NONLINEAR AND NONSTANDARD
OPTION PRICING
A thesis submitted to the University of Manchester
for the degree of Doctor of Philosophy
in the Faculty of Engineering and Physical Sciences
2008
Kristoﬀer John Glover
School of Mathematics
Contents
Abstract 11
Declaration 12
Copyright Statement 13
Acknowledgements 14
Dedication 15
1 Introduction 16
1.1 Evidence of increased interest in liquidity . . . . . . . . . . . . . . . . 17
1.2 A brief history . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
1.3 Derivative pricing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
1.3.1 European options . . . . . . . . . . . . . . . . . . . . . . . . . 21
1.3.2 Arbitrage pricing . . . . . . . . . . . . . . . . . . . . . . . . . 22
1.3.3 The FeynmanKac representation theorem . . . . . . . . . . . 24
1.3.4 From FeynmanKac to BlackScholes . . . . . . . . . . . . . . 25
1.3.5 American options . . . . . . . . . . . . . . . . . . . . . . . . . 27
1.3.6 Optimal stopping problems . . . . . . . . . . . . . . . . . . . 29
1.3.7 Freeboundary problems . . . . . . . . . . . . . . . . . . . . . 30
1.4 Supply and demand economics . . . . . . . . . . . . . . . . . . . . . . 32
1.5 Liquidity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
1.5.1 Deﬁning liquidity . . . . . . . . . . . . . . . . . . . . . . . . . 35
1.5.2 Measuring liquidity . . . . . . . . . . . . . . . . . . . . . . . . 36
1.6 Price formation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
2
1.7 Option pricing in illiquid markets: a literature review . . . . . . . . . 40
1.8 Introduction to perturbation methods . . . . . . . . . . . . . . . . . . 45
1.9 Layout of the thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
2 The Modelling Framework 48
2.1 Technical asides . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
2.1.1 Markovian processes . . . . . . . . . . . . . . . . . . . . . . . 53
2.1.2 Applicability of Itˆ o’s formula . . . . . . . . . . . . . . . . . . . 54
2.2 Alternative models . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
2.2.1 Transactioncost models . . . . . . . . . . . . . . . . . . . . . 56
2.2.2 Reactionfunction (equilibrium) models . . . . . . . . . . . . . 57
2.2.3 Reducedform SDE models . . . . . . . . . . . . . . . . . . . . 58
2.3 A uniﬁed framework . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
2.3.1 Cetin et al. (2004) . . . . . . . . . . . . . . . . . . . . . . . . 59
2.3.2 Platen and Schweizer (1998) . . . . . . . . . . . . . . . . . . . 59
2.3.3 Mancino and Ogawa (2003) . . . . . . . . . . . . . . . . . . . 60
2.3.4 Lyukov (2004) . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
2.3.5 Sircar and Papanicolaou (1998) . . . . . . . . . . . . . . . . . 61
3 Firstorder Feedback Model 64
3.1 Analysis close to expiry: European options . . . . . . . . . . . . . . . 67
3.2 Analysis close to expiry: American put options . . . . . . . . . . . . . 72
3.3 The vanishing of the denominator . . . . . . . . . . . . . . . . . . . . 77
4 Fullfeedback Model 83
4.1 Putcall parity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
4.2 A solution by inspection . . . . . . . . . . . . . . . . . . . . . . . . . 87
4.3 Similarity solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
4.4 Perturbation expansions . . . . . . . . . . . . . . . . . . . . . . . . . 90
4.5 Numerical solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
4.6 Analysis close to expiry . . . . . . . . . . . . . . . . . . . . . . . . . . 93
4.7 Numerical results  full problem . . . . . . . . . . . . . . . . . . . . . 97
4.7.1 A second solution regime . . . . . . . . . . . . . . . . . . . . . 99
3
5 Smoothed Payoﬀs  Another Breakdown 102
5.1 Local analysis about the singularities . . . . . . . . . . . . . . . . . . 106
5.1.1 Asymptotic matching . . . . . . . . . . . . . . . . . . . . . . . 108
5.1.2 Properties of the inner solution . . . . . . . . . . . . . . . . . 110
5.1.3 Introduction to phaseplane analysis . . . . . . . . . . . . . . 111
5.1.4 Deriving an autonomous system . . . . . . . . . . . . . . . . . 114
5.1.5 Behaviour of the ﬁxed points . . . . . . . . . . . . . . . . . . 116
5.1.6 Structure of the phase portrait . . . . . . . . . . . . . . . . . 120
5.1.7 Other ﬁxed points . . . . . . . . . . . . . . . . . . . . . . . . . 122
6 Perpetual Options 127
6.1 Analytic solutions and perturbation methods . . . . . . . . . . . . . . 131
7 Other Models 137
7.1 Frey (1998, 2000) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
7.2 Frey and Patie (2002) . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
7.3 Sircar and Papanicolaou (1998) . . . . . . . . . . . . . . . . . . . . . 140
7.4 Bakstein and Howison (2003) . . . . . . . . . . . . . . . . . . . . . . 141
7.4.1 Nonsmooth solutions to the Bakstein and Howison (2003) model146
7.4.2 New nonsmooth solutions to the BlackScholes equation . . . 147
7.5 Liu and Yong (2005) . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
7.5.1 Vanishing of the denominator . . . . . . . . . . . . . . . . . . 150
7.6 Jonsson and Keppo (2002) . . . . . . . . . . . . . . . . . . . . . . . . 152
7.6.1 Connections with the other modelling frameworks . . . . . . . 154
8 Explaining the Stock Pinning Phenomenon 155
8.1 Linear price impact . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156
8.2 Nonlinear price impact . . . . . . . . . . . . . . . . . . . . . . . . . . 159
8.3 A new nonlinear price impact model . . . . . . . . . . . . . . . . . . 161
9 The British Option 164
9.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164
9.2 The noarbitrage price . . . . . . . . . . . . . . . . . . . . . . . . . . 167
4
9.2.1 The gain function . . . . . . . . . . . . . . . . . . . . . . . . . 170
9.3 Numerical treatment . . . . . . . . . . . . . . . . . . . . . . . . . . . 171
9.4 Free boundary analysis far from expiry . . . . . . . . . . . . . . . . . 175
9.5 Analysis close to expiry . . . . . . . . . . . . . . . . . . . . . . . . . . 180
9.6 Financial analysis of the British put option . . . . . . . . . . . . . . . 186
9.7 The British call option . . . . . . . . . . . . . . . . . . . . . . . . . . 193
9.7.1 Analysis far from expiry . . . . . . . . . . . . . . . . . . . . . 196
9.7.2 Analysis close to expiry . . . . . . . . . . . . . . . . . . . . . 198
9.8 Integral representations of the free boundary . . . . . . . . . . . . . . 198
9.8.1 The American put option . . . . . . . . . . . . . . . . . . . . 198
9.8.2 The British put option . . . . . . . . . . . . . . . . . . . . . . 201
10 Conclusions 204
A Maximum Principles 223
A.1 Nonlinear equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 226
A.2 Uniqueness of PDEs . . . . . . . . . . . . . . . . . . . . . . . . . . . 227
A.2.1 The linear BlackScholes equation . . . . . . . . . . . . . . . . 228
A.2.2 The nonlinear (illiquid) BlackScholes equation . . . . . . . . . 229
A.3 Monotonicity in λ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 230
B Nondimensionalisation of the British Put 232
C The Probability Density Function 233
Word count 69834
5
List of Figures
3.1 Value of European call options with ﬁrstorder feedback (T = 1, r =
0.04, σ = 0.2, K = 1) for λ = 0, 1, 2, 5, 10; the variation with λ
appears to be monotonic. . . . . . . . . . . . . . . . . . . . . . . . . . 66
3.2 Value of European put options with ﬁrstorder feedback (T = 1, r =
0.04, σ = 0.2, K = 1) for λ = 0, 1, 2, 5, 10; the variation with λ
appears to be monotonic. . . . . . . . . . . . . . . . . . . . . . . . . . 66
3.3 Asymptotic Matching. . . . . . . . . . . . . . . . . . . . . . . . . . . 70
3.4 Inner solution minus the payoﬀ for put and call options, r = 0.04,
σ = 0.2, K = 1 and for λ = 0.1, 0.15, 0.2, . . ., 0.4. . . . . . . . . . . . 72
3.5 Value of American put options, T = 1, r = 0.04, σ = 0.2, K = 1 and
for λ = 0, 1, 2, 5, 10; the variation with λ appears to be monotonic. . 73
3.6 Firstorder feedback put (with early exercise), location of free bound
ary (as τ →0) with λ, K = 1, r = 0.04, σ = 0.2. . . . . . . . . . . . . 76
3.7 Location of the vanishing of the denominator of (2.9) with λ = 0.1,
K = 1, r = 0.04 and σ = 0.2. . . . . . . . . . . . . . . . . . . . . . . 78
3.8 The ﬁrst derivative (∆) of the BlackScholes equation (3.2) (dotted
line) and the ﬁrst order feedback PDE (2.9) (solid line) for τ = 0.01,
0.015, . . ., 0.05. Compare the location of the vanishing denominator 3.7. 79
3.9 The second derivative (Γ) of the BlackScholes equation (3.2) (dotted
line) and the ﬁrst order feedback PDE (2.9) (solid line) for τ = 0.01,
0.015, . . ., 0.05. Compare the location of the vanishing denominator 3.7. 79
6
3.10 Firstorder feedback put option value for two diﬀerent values of λ at
various times to expiry; τ = 0.0125, 0.0375, 0.075. For r = 0.04,
σ = 0.2, K = 1 and λ = 0.09 (solid line) and λ = 0.1 (dotted line).
Compare with ﬁgure 3.7. . . . . . . . . . . . . . . . . . . . . . . . . . 81
4.1 The leading order correction term V
1
(S, τ) to the BlackScholes (i.e.
λ = 0) European put option for various time to expiry. K = 1,
r = 0.04, σ = 0.2, T = 1 and τ = 0.1, 0.2, . . . , 1. . . . . . . . . . . . . 92
4.2 Deltas for fullfeedback (European) put, K = 1, r = 0.04, σ = 0.2,
λ = 0.1 and T = 1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
4.3 Local (τ →0) solution of a fullfeedback put, K = 1, λ = 0.1, r = 0.04
and σ = 1, 0.95, . . ., 0.15. . . . . . . . . . . . . . . . . . . . . . . . . 96
4.4 Full feedback put, K = 1, r = 0.04, σ = 0.2 and λ = 0.1; modiﬁed
numerical scheme. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
4.5 Full feedback put, K = 1, r = 0.04, σ = 0.2 and λ = 0.1; modiﬁed
numerical scheme. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
4.6 Full feedback call, K = 1, r = 0.04, σ = 0.2 and λ = 0.1; modiﬁed
numerical scheme. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
4.7 Full feedback put, smoothed payoﬀ, K = 1, r = 0.04, σ = 0.1, λ = 0.1
and τ = 0.01. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
5.1 Phase portrait of the autonomous system (5.24). Note the ﬁxed point
at u =
_
243
80
_1
3
, v = 0 and the ﬁeld direction lines. The dotted line rep
resents an analytic envelope for the phase portrait close to the singular
line v =
5u
3
, cf. equation (5.31). . . . . . . . . . . . . . . . . . . . . . 120
6.1 Full feedback American put, K = 1, r = 0.04, σ = 0.2, λ = 0.25,
ρ = 0.15 (smoothed payoﬀ), τ = 0, 1, . . . , 10. Note that we are in the
regime λ < 2ρ and so we should expect no singular behaviour. . . . . 128
6.2 Perpetual fullfeedback American put, K = 1, r = 0.04, σ = 0.2,
λ = 0, 0.1, 0.2, . . . , 1.1; freeboundary location as indicated. . . . . . . 130
7
6.3 The ﬁrst order correction to the BlackScholes perpetual American put
option (solid line) compared to the diﬀerence of the fully numerical
option value with the BlackScholes (dotted line). K = 1, r = 0.04,
σ = 0.2 and λ = 0.1. . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
6.4 The ﬁrst order correction to the BlackScholes perpetual American put
option (solid line) compared to the diﬀerence of the fully numerical
option value with the BlackScholes (dotted line) for various values of
λ. K = 1, r = 0.04, σ = 0.2 and λ = 0.1, 0.5, 1. . . . . . . . . . . . . 136
7.1 Location of the vanishing of the denominator of the Frey (1998, 2000)
(solid line) and Sch¨ onbucher and Wilmott (2000) (dotted line) model
with λ = 0.1, K = 1, r = 0.04 and σ = 0.2. . . . . . . . . . . . . . . 139
7.2 Local (τ → 0) call solution of the Sircar and Papanicolaou (1998)
model K = 1, r = 0.04, σ = 0.2, and
ˆ
λ = 0, 0.05, . . . , 0.2. . . . . . . . 141
7.3 Local (τ → 0) put solution of the Sircar and Papanicolaou (1998)
model K = 1, r = 0.04, σ = 0.2, and
ˆ
λ = 0 0.05, . . ., 0.3. . . . . . . . 142
7.4 Solution to equation (7.9) for a put option with λ = 0.01, 0.5, 1, . . .,
5, σ = 0.2, r = 0.04, K = 1, and α = 1.5. . . . . . . . . . . . . . . . . 144
7.5 Local (τ →0) put solution of the Bakstein and Howison (2003) model
K = 1, r = 0.04, σ = 0.2, α = 1.5, and
ˆ
λ = −5, 4.75, . . ., 5. . . . . . 147
7.6 Local (τ →0) call solution of the Bakstein and Howison (2003) model
K = 1, r = 0.04, σ = 0.2, α = 1.5, and
ˆ
λ = −5, 4.75, . . ., 5. . . . . . 148
7.7 Nonsmooth solution of the BlackScholes equation. K = 1, r = 0.04,
σ = 0.2. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
7.8 Location of the vanishing of the denominator for the Liu and Yong
(2005) model for various value of β. K = 1, r = 0.04, σ = 0.2,
λ = 0.1, and β = 1 10
5
, 2 10
5
, . . ., 1 10
6
. . . . . . . . . . . . . . 151
7.9 Local (τ → 0) call solution of the Jonsson and Keppo (2002) model
K = 1, σ = 0.2, and a = −1, 0.9, . . ., 1. . . . . . . . . . . . . . . . . 153
7.10 Local (τ → 0) put solution of the Jonsson and Keppo (2002) model
K = 1, σ = 0.2, and a = −1, 0.9, . . ., 1. . . . . . . . . . . . . . . . . 153
8
8.1 The pinning probability (8.5) for values of nE = 0.5, 1, . . ., 5. T −t =
0.1, K = 1, and σ = 0.2. . . . . . . . . . . . . . . . . . . . . . . . . . 158
8.2 Comparing the pinning probability associated with (8.6) (solid line)
with the model of Avellaneda and Lipkin (2003) (dotted line) for nE =
0.1, T −t = 0.1, K = 1, σ = 0.2, and r +
1
2
σ
2
= 0. . . . . . . . . . . . 159
8.3 Solution to (8.7) for p = 0.8, 0.9, . . . , 1.2, T − t = 0.1, K = 1, σ = 0.2
and r +
1
2
σ
2
= 0. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
8.4 Solution to equation (8.8) (solid line) compared to (8.5) (dotted line)
for T = 0.1, K = 1, σ = 0.2, and r +
1
2
σ
2
= 0. . . . . . . . . . . . . . 163
9.1 The British put option free boundary for varying values of the contract
drift. T = 1, K = 1, σ = 0.4, r = 0.1, D = 0, and µ
c
= 0.11, 0.115,
0.12, . . ., 0.16. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173
9.2 The British put option free boundary for varying volatilities. T = 1,
K = 1, µ
c
= 0.125, r = 0.1, D = 0, and σ = 0.05, 0.1, . . ., 0.5. . . . . 173
9.3 The zero of the Hfunction, i.e. S
h
(t), for varying values of the contract
drift. µ
c
= 0.102, 0.104, . . . , 1. T = 50, K = 1, r = 0.1, D = 0, and
σ = 0.4. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177
9.4 The asymptotic approximation for the British put option free bound
ary close to expiry, i.e. (9.27) (dotted line) compared with fully nu
merical value (solid line). T = 0.01, σ = 0.4, r = 0.1, µ
c
= 0.125, and
D = 0. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 186
9.5 Location of the free boundary for the British (solid line) and American
(dotted line) put option under investigation in ﬁgures 9.6, 9.7 and 9.8.
T = 1, K = 1, σ = 0.4, r = 0.1, µ
c
= 0.125, and D = 0. . . . . . . . . 189
9.6 The diﬀerence in the percentage return of the British put option and
the American put option at every possible stopping location. The solid
lines denote contours at increments of 10% from 10% to 60%. The
dotted line represents the zero contour. S
0
= 1, T = 1, K = 1, σ = 0.4,
r = 0.1, D = 0, and µ
c
= 0.125. . . . . . . . . . . . . . . . . . . . . . 190
9
9.7 The diﬀerence in the percentage return of the British put option and
the European put option. Again the solid lines denote contours at
increments of 10% from 0% to 70%. The dotted line represents the
zero contour. S
0
= 1, T = 1, K = 1, σ = 0.4, r = 0.1, D = 0, and
µ
c
= 0.125. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 190
9.8 The diﬀerence in the percentage return of the American put option and
the European put option. The solid lines denote contours at increments
of 10% from 70% to 30%. The dotted line represents the zero contour.
S
0
= 1, T = 1, K = 1, σ = 0.4, r = 0.1, and D = 0. Note the change
of orientation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191
9.9 Schematic representation of the regions in which atthemoney Eu
ropean, American and British put option would provide the greatest
return on an investment. The dotted lines represent the free bound
aries of the American and British put option for reference. T = 1,
K = 1, σ = 0.4, r = 0.1 and D = 0. . . . . . . . . . . . . . . . . . . . 192
9.10 Figures representing the region in which American put options would
provide a greater expected return that its British option counterpart,
for increasing moneyness. T = 1, K = 1, σ = 0.4, r = 0.1 and D = 0. 194
9.11 The British call option free boundary for varying values of the contract
drift. T = 1, K = 1, σ = 0.4, r = 0.1, D = 0, and µ
c
= 0.05, 0.055,
0.06, . . ., 0.09. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195
9.12 The British call option free boundary for varying volatilities. T = 1,
K = 1, µ
c
= 0.08, r = 0.1, D = 0, and σ = 0.05, 0.1, . . ., 0.5. . . . . . 195
9.13 The asymptotic approximation for the British call option free bound
ary close to expiry, i.e. (9.32) (dotted line) compared with fully nu
merical value (solid line). T = 0.01, K = 1, σ = 0.4, r = 0.1, µ
c
= 0.08
and D = 0. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 199
10
The University of Manchester
Kristoﬀer John Glover
Doctor of Philosophy
The Analysis of PDEs Arising in Nonlinear and Nonstandard Option
Pricing
October 23, 2008
This thesis examines two distinct classes of problem in which nonlinearities arise in
option pricing theory. In the ﬁrst class, we consider the eﬀects of the inclusion of ﬁ
nite liquidity into the BlackScholesMerton option pricing model, which for the most
part result in highly nonlinear partial diﬀerential equations (PDEs). In particular,
we investigate a model studied by Sch¨ onbucher and Wilmott (2000) and furthermore,
show how many of the proposed existing models in the literature can be placed into
a uniﬁed analytical framework. Detailed analysis reveals that the form of the nonlin
earities introduced can lead to serious solution diﬃculties for standard (put and call)
payoﬀ conditions. One is associated with the inﬁnite gamma and in such regimes
it is necessary to admit solutions with discontinuous deltas, and perhaps even more
disturbingly, negative option values. A second failure (applicable to smoothed payoﬀ
functions) is caused by a singularity in the coeﬃcient of the diﬀusion term in the
optionpricing equation. It is concluded in this case is that the model irretrievably
breaks down and there is insuﬃcient ‘ﬁnancial modelling’ in the pricing equation.
The repercussions for American options are also considered.
In the second class of problem, we investigate the properties of the recently intro
duced British option (Peskir and Samee, 2008a,b), a new nonstandard class of early
exercise option, which can help to mediate the eﬀects of a ﬁnitely liquid market,
since the contract does not require the holder to enter the market and hence incur
liquidation costs. Here we choose to focus on the interesting nonlinear behaviour of
the earlyexercise boundary, speciﬁcally for times close to, and far from, expiry.
In both classes, detailed asymptotic analysis, coupled with advanced numerical tech
niques (informed by the asymptotics) are exploited to extract the relevant dynamics.
11
Declaration
No portion of the work referred to in this thesis has been
submitted in support of an application for another degree
or qualiﬁcation of this or any other university or other
institute of learning.
12
Copyright Statement
i. The author of this thesis (including any appendices and/or schedules to this
thesis) owns any copyright in it (the “Copyright”) and s/he has given The
University of Manchester the right to use such Copyright for any administrative,
promotional, educational and/or teaching purposes.
ii. Copies of this thesis, either in full or in extracts, may be made only in accor
dance with the regulations of the John Rylands University Library of Manch
ester. Details of these regulations may be obtained from the Librarian. This
page must form part of any such copies made.
iii. The ownership of any patents, designs, trade marks and any and all other
intellectual property rights except for the Copyright (the “Intellectual Property
Rights”) and any reproductions of copyright works, for example graphs and
tables (“Reproductions”), which may be described in this thesis, may not be
owned by the author and may be owned by third parties. Such Intellectual
Property Rights and Reproductions cannot and must not be made available
for use without the prior written permission of the owner(s) of the relevant
Intellectual Property Rights and/or Reproductions.
iv. Further information on the conditions under which disclosure, publication and
exploitation of this thesis, the Copyright and any Intellectual Property Rights
and/or Reproductions described in it may take place is available from the Head
of the School of Mathematics.
13
Acknowledgements
I am extremely grateful to my supervisors Professor Peter W. Duck and David P.
Newton for their expert guidance and continued support throughout the course of
this Ph.D. In particular, I thank Peter for his boundless knowledge, enthusiasm and
eﬃciency, and David for his caring supervision and his conﬁdence in my abilities. In
addition, EPSRC funding is gratefully acknowledged.
I thank my parents for their love and unwavering support for which these mere
expressions of gratitude do not suﬃce. Thank you to my colleagues and friends
for their invaluable advice and numerous enlightening discussions. In particular, to
Goran Peskir for his time and enthusiasm for the subject, and to Erik Ekstr¨ om for
his insight and friendship.
To my close friends, both old and new, and in particular to Jonathan Causey, Helen
Burnip, Philip Haines, John Heap, Sebastian Law and Vicky Thompson, I thank you
for creating the good times and for being there through the bad. I hope, despite
the distances between us, our friendships can continue to ﬂourish. Finally, I thank
Hannah for everything, I hope we both ﬁnd what we’re looking for.
14
Dedication
To Gran, in loving memory.
15
Chapter 1
Introduction
Nowadays people know the price of everything and the value of nothing.
 Oscar Wilde (18541900)
The Picture of Dorian Gray (1891)
Mathematical ﬁnance is not a branch of the physical sciences. There are no laws of
nature just waiting to be discovered; one is not trying to model Mother Nature and
her laws, but the nature of man and his markets. However, this does not preclude
us from trying to quantify the ﬁnancial markets and to utilise the powerful tools of
mathematics in order to better understand such markets.
Since the deﬁnitive papers of Black and Scholes (1973) and Merton (1973), much of
the work undertaken in mathematical ﬁnance has been aimed at relaxing a number
of the modelling assumptions. One of the more subtle was that the market in the
underlying asset
1
was inﬁnitely (or perfectly) elastic, such that trading had no impact
on the price of the underlying. If we relax this assumption, then we see some rather
interesting and possibly counterintuitive behaviours. As we shall show later, this is
partly due to the fact that any model incorporating such a feature will inevitably
lead to nonlinear behaviour (feedback). In particular, we shall be concerned for the
most part with nonlinear partial diﬀerential equations (PDEs) arising from the study
1
Termed underlying in the sequel.
16
CHAPTER 1. INTRODUCTION 17
of ﬁnitely elastic markets. Work that has led to this class of PDEs in ﬁnance to
date includes Whalley and Wilmott (1993) in relation to transaction costs, which
was one of the ﬁrst nonlinear PDEs to arise in the ﬁeld of mathematical ﬁnance.
In addition, there is the so called BlackScholesBarrenblatt equation introduced by
Avellaneda et al. (1995) in the study of uncertain volatility models. These models
involve optimisation over all possible values of volatility, and as a result are also
highly nonlinear.
The aim of modelling the behaviour of the underlying is to capture the dynamics
of the observed market prices as faithfully as possible. One approach to incorporate
these dynamics is to ﬁnd a stochastic process that ﬁts most closely the distribution of
returns of the underlying. This is an exogenous strategy, and as such provides little
insight into which of the many factors aﬀecting the price dynamics are actually the
most important. In addition, the exogenous processes required tend to be diﬃcult to
handle mathematically, for example L´evy processes. An alternative approach (and
that to be followed in this thesis) is to retain one the simplest stochastic process,
geometric Brownian motion, but to provide an endogenous mechanism by which
the dynamics diﬀer from this standard geometric Brownian motion. This provides
much greater insight into how prices are actually formed in the market, and has the
advantage of being consistent with the bulk of the literature over the past thirtyﬁve
years.
In this chapter we introduce the basic ideas and concepts and review the results of
the classical BlackScholesMerton option pricing theory used in later chapters. It
is by no means a complete treatment of the relevant theories, just enough for the
unfamiliar reader to understand the contributions of the following chapters.
1.1 Evidence of increased interest in liquidity
Recent worries about the health of the modern ﬁnancial system have deterred people
from getting involved in the derivatives markets. This has resulted in trading volumes
CHAPTER 1. INTRODUCTION 18
decreasing and hence increased liquidity problems. David Oakley of the Financial
Times
2
warns that
...the sharp slowdown in these [derivative] markets is a serious warning
sign of the growing problems in the ﬁnancial world as they are usually
highly liquid, turning over vast amounts of trade every day.
Further, Rachel Lomax, the Bank of England’s Deputy Governor goes on to describe
the recent ﬁnancial turmoil in the wake of the American subprime mortgage prob
lems
3
as
...the largest ever peacetime liquidity crisis.
The current liquidity crisis can be traced back to the collapse of the US subprime
mortgage market. In August 2007 the Financial Times is quoted as saying that
4
...as market turmoil rises ﬁnancial problems are no longer simply conﬁned
to a risky corner of the US mortgage market. This stems from another
key theme now haunting the markets: namely that liquidity is evaporating
from numerous corners of the ﬁnancial world, as both investors in hedge
funds and the banks that lend to them try to cut and run from recent
losses.
Clearly, in times of crisis, liquidity becomes an ever important issue, motivating
further investigation into the eﬀects of reduced liquidity on all aspects of the ﬁnancial
markets. In a recent blog entry regarding the subprime induced liquidity crisis Paul
Wilmott states that
5
...this should spur on the implementation of mathematical models for
feedback... which may in turn help banks and regulators to ensure that
2
See Derivative liquidity crisis ‘to continue’, David Oakley, FT.com, November 23 2007.
3
Quoted in Bank deputy downbeat on economy, Chris Giles, FT.com, February 27 2008.
4
See Liquidity alarm bells sound, Paul J Davies, Gillian Tett, Joanna Chung and StacyMarie
Ishmael, FT.com, August 1 2007.
5
Quoted in Science in Finance IV: The feedback eﬀect Paul Wilmott, blog entry at http://www.
wilmott.com/blogs/paul/, January 29 2008.
CHAPTER 1. INTRODUCTION 19
the press that derivatives are currently getting is not as bad as it could
be.
1.2 A brief history
In 1828 Robert Brown (17731858), a Scottish botanist, observed the apparently ran
dom motion of pollen particles suspended in water and subsequently during the 19th
century it became clear that the pollen particles were being bombarded by a multi
tude of molecules of the surrounding water, whose aggregate eﬀect was (apparently)
random. In addition, wherever we look we see a random world and therefore Brow
nian motion (named in honour of Robert Brown) is an invaluable tool for describing
this randomness. In fact, the ubiquitous nature of Brownian motion can be seen as
the dynamic counterpart of the ubiquitous nature of the normal distribution, which
rests ultimately on the Central Limit Theorem.
6
The origins of much of ﬁnancial mathematics trace back to a dissertation (entitled
Th´eorie de la sp´eculation
7
) published in 1900 by Louis Bachelier (18701946). In
it he proposed to model the movement of stock prices with a diﬀusion process or
Brownian motion. Note that this was ﬁve years before Einstein’s seminal paper
outlining the theory of Brownian motion, and it was not until the 1920s that the
rigorous mathematical underpinnings of the theory of Brownian motion was provided
by Norbert Wiener (18941964).
Meanwhile, as quantum mechanics emerged in the 1920s it began to become clear
that the quantum picture is both inescapable at the subatomic level and intrinsically
probabilistic. The work of Richard P. Feynman (19181988) in the late 1940s on quan
tum mechanics using path integrals, introduced the Wiener measure into the heart
of quantum theory. Feynman’s work was made mathematically rigorous by Mark
Kac (19141984) and the socalled FeynmanKac formula, which gives a stochastic
6
See for example Jacod and Protter (2003).
7
For a translated version with commentary and a foreword by Paul Samuelson see Davis and
Etheridge (2006).
CHAPTER 1. INTRODUCTION 20
representation for the solution to certain classes of PDEs, was introduced (see section
1.3.3).
In 1944 Kiyoshi Itˆ o (1915) went on to develop stochastic calculus, the machinery
needed in order to use Brownian motion to model stock prices successfully, and which
would later become an essential tool of modern ﬁnance. However, it was not until 1965
that economist Paul Samuelson (1915) resurrected Bachelier’s work and advocated
Itˆ o’s geometric Brownian motion model as a suitable model for stock price movements.
After this it was not long until Black, Scholes and Merton wrote down their famous
equation for the price of a European call and put option in 1969, work for which the
surviving members (Scholes and Merton) received the Nobel Prize for economics in
1997.
A more comprehensive overview of the early years of mathematical ﬁnance can be
found in Jarrow and Protter (2004).
1.3 Derivative pricing
When we discretise a problem it becomes easier to deﬁne or understand but much
harder to solve without the use of continuous time calculus; this thesis deals solely
with continuous time models. In continuoustime modelling there are two main ap
proaches to calculating the price of a given derivative security, the socalled martin
gale approach and the PDE approach. In the former, a stochastic process for the
underlying is speciﬁed and an equivalent probability measure is found that turns the
discounted underlying into a martingale. The price of the derivative is then deﬁned
as the conditional expectation of its discounted payoﬀ under this new (riskneutral)
measure. Alternatively, in the PDE approach, a stochastic process for the underlying
is likewise speciﬁed and then Itˆ o’s formula for a function of the underlying stochastic
process is used to derive a PDE involving the coeﬃcients of the underlying process.
The two approaches are deeply linked via the famous FeynmanKac formula (outlined
in section 1.3.3) and it should be noted that both approaches can be used for complete
CHAPTER 1. INTRODUCTION 21
and incomplete markets. In the latter case, arriving at a unique price for a derivative
requires additional assumptions. If one is using the martingale approach, then this
arbitrariness is reﬂected in the choice of equivalent martingale measure, whereas using
the PDE approach the choice of martingale measure is analogous to specifying the
socalled market price of risk of the nontraded variable. Since the models introduced
in this thesis result in complete markets,
8
i.e. all sources of risk are traded, both the
martingale approach and the PDE approach should arrive at the same price.
The fair price of a derivative security (and all other ﬁnancial instruments) is de
termined by the expected discounted value of some future payoﬀ, which is itself
dependent on the future value of the underlying asset. Of course, the future value
of the underlying is not known a priori, and price processes are often modelled by
stochastic processes. Therefore, an understanding of the behaviour of such stochastic
processes is a valuable prerequisite for the study of derivative pricing; this section
attempts to provide such an understanding. The derivative securities studied in this
thesis, without exception, are options contracts. A brief overview of the types of
contracts referred to in the main body of the thesis will be considered next.
1.3.1 European options
European options are the simplest type of options contract and within this class the
most common are call options and put options. The holder of a call option written
on a certain underlying asset (usually a stock) has the right, but not the obligation,
to buy the underlying at some predetermined date, denoted T, and at some pre
determined price, denoted K. If the underlying at time t = T, S
T
, is worth more
then K then the (rational) holder would exercise the option and make a proﬁt S
T
−K.
Alternatively, if S
T
is less than K, then the holder would not exercise, resulting in
the option expiring worthless. Thus, the value of the call option at expiry (T) is
given by
V
C
(S
T
, T) = (S
T
−K)
+
:= max¦S
T
−K, 0¦. (1.1)
8
Under suitable restrictions, see chapter 2.
CHAPTER 1. INTRODUCTION 22
Similarly, the holder of a put option has the right to sell the underlying for the
exercise price K, resulting in the value of the put option:
V
P
(S
T
, T) = (K −S
T
)
+
:= max¦K −S
T
, 0¦. (1.2)
The functions (1.1) and (1.2) are called payoﬀ proﬁles and will be referred to as such
throughout this thesis. There are, of course, many diﬀerent options contracts with
more general payoﬀ proﬁles, h(S
T
) say. For an option to be described as European,
its contract must specify that exercise is only possible at a single maturity time, T.
Note that these contracts dependent only on the price of the underlying at expiry,
S
T
, and not on the path of the price prior to maturity; this results in tractability in
many situations. Options that allow exercise at times prior to expiry are said to have
an earlyexercise feature. More speciﬁcally if the option allows exercise at any time
prior to expiry such an option is referred to as an American option. These options
are very popular in practise, and will play an important role in much of this thesis.
Indeed we shall return to them shortly in section 1.3.5.
1.3.2 Arbitrage pricing
An arbitrage opportunity corresponds to a riskfree proﬁt. More formally, it is the
opportunity to construct a trading strategy (i.e. buying and selling ﬁnancial instru
ments) in such a way that the initial investment (at t = 0) is zero and the wealth at
time T is nonnegative with a nonzero probability of a strictly positive wealth. In
an eﬃcient market there should be no such arbitrage opportunities and indeed the
seminal work by Black and Scholes (1973) and Merton (1973) used the noarbitrage
principle to arrive at a unique price for the fair value of an option contract. To state
their results, we have a market consisting of a bank account which grows according
to the (deterministic) dynamics
dB = rBdt,
and one risky asset, with stochastic price dynamics
dS
t
= µS
t
dt +σS
t
dW
P
t
, (1.3)
CHAPTER 1. INTRODUCTION 23
where r is the positive (constant) interest rate, µ the drift and σ the volatility of
the underlying price process. W
P
t
denotes a standard Brownian motion under the
probability measure P. The fair value or price of a European option contract V (S, t)
with payoﬀ proﬁle h(S
T
) can be shown to be given by
V (S, t) = E
Q
S,t
_
e
−r(T−t)
h(S
T
)
¸
, (1.4)
in words, the expected discounted future payoﬀ. The indices indicate that the pro
cess for S
t
is started at S at time t and also that the expectation is calculated under
the socalled riskneutral probability measure, Q, as opposed to the real world mea
sure, P, deﬁned by the process (1.3).
9
The riskneutral measure is deﬁned as the
unique measure equivalent to P under which the discounted price process is a mar
tingale. Consequently, the stock price process (1.3) can then be described in terms
of a standard QBrownian motion W
Q
t
as
dS
t
= rS
t
dt +σS
t
dW
Q
t
. (1.5)
Note that the dynamics of S
t
under the riskneutral measure Q are the same as
the dynamics under the realworld measure P, except that the drift of S
t
under Q
is equal to the interest rate r instead of µ. Consequently the drift parameter µ
does not appear anywhere in the pricing formula for European options; this fact
undoubtedly contributed to the widespread application of the BlackScholesMerton
pricing methodology in the years subsequent to its publication, since in practise
the drift parameter is notoriously diﬃcult to measure from past time series of the
underlying process.
10
The model analysed above is an example of a complete market model. The simplest
deﬁnition of a complete market is one in which every derivative security can be repli
cated by a selfﬁnancing trading strategy in the stock and bond. In this model, any
security whose payoﬀ h(S
T
) is known at time T (where h(S
T
) is any T
T
measurable
9
This subtlety was the main innovation of option pricing research in the 1970s. Prior to this,
expectations had been taken under the real world measure P.
10
In fact, Liptser and Shiryaev (2001) show that the expected waiting time to obtain an estimate
of the drift (via the naive approximation S
t
/t) that is within of the true drift is proportional to
−2
. For example if = 0.01 it would take ∼ 10, 000 years to obtain such an estimate.
CHAPTER 1. INTRODUCTION 24
random variable with E[h
2
(S
T
)] < ∞) can be replicated by some unique selfﬁnancing
trading strategy. Finally, we note that in a complete market, a characterisation of
the arbitragefree principle is that there exists a unique equivalent martingale mea
sure Q, under which the discounted prices of traded securities are martingales. For
more on this characterisation see the original works of Harrison and Kreps (1979)
and Harrison and Pliska (1981, 1983).
Expected values of solutions to stochastic diﬀerential equations (SDEs), such as the
pricing equation (1.4), are linked to the solution of (linear) parabolic partial diﬀeren
tial equations (PDEs) via the famous FeynmanKac representation theorem. Thus,
the price of a European option can be studied using both stochastic methods and
parabolic PDE methods; this thesis focuses primarily on the latter. In the following
section we describe the FeynmanKac representation theorem.
1.3.3 The FeynmanKac representation theorem
Suppose we are given the PDE for the unknown function u(S, t)
∂u
∂t
+
1
2
σ
2
(S, t)
∂
2
u
∂S
2
+µ(S, t)
∂u
∂S
= 0, (1.6)
subject to the ﬁnal condition
u(S, T) = h(S),
where µ(S, t), σ(S, t) and h(S) are known functions and T a parameter. This equation
is sometimes called the Kolmogorov backward equation. The FeynmanKac formula
tells us that the solution can be written as an expectation,
u(S, t) = E
P
S,t
[h(S
T
)]
where S
t
is a stochastic process given by the equation
dS
t
= µ(S
t
, t)dt +σ(S
t
, t)dW
P
t
. (1.7)
The indices on the expectation indicates that the process S
t
is started at S at time t
and in addition the superscript indicates that the expectation is taken under the prob
ability measure P, corresponding to the stochastic process (1.7), with a PBrownian
CHAPTER 1. INTRODUCTION 25
motion W
P
t
. This useful representation allows us to solve deterministic PDEs via
stochastic methods and, conversely, expectations of functions of stochastic processes
via deterministic PDEs.
Proof. The proof of the FeynmanKac representation is fairly straightforward and so
we shall outline the basic idea here. Consider an unknown function u(S, t). Applying
Itˆ o’s formula we have
du =
_
∂u
∂t
+µ(S, t)
∂u
∂S
+
1
2
σ
2
(S, t)
∂
2
u
∂S
2
_
dt +σ(S, t)
∂u
∂S
dW
P
t
.
Now, by assumption the O(dt) terms above are zero if u(S, t) is assumed to be the
solution of the PDE (1.6). Integrating the above equation we obtain
_
T
t
du = u(S
T
, T) −u(S
t
, t) =
_
T
t
σ(S, t)
∂u
∂S
dW
P
t
.
Next, taking expectations and reorganising a little we arrive at
u(S, t) = E
P
S,t
[u(S
T
, T)] −E
P
S,t
__
T
t
σ(S, t)
∂u
∂S
dW
P
t
_
.
Finally, it can be shown that the expectation of an Itˆ o integral with respect to a
Brownian motion is zero (see, for example, prop. 4.4 of Bj¨ ork, 2004) resulting in the
required result
u(S, t) = E
P
S,t
[u(S
T
, T)] = E
P
S,t
[h(S
T
)] .
1.3.4 From FeynmanKac to BlackScholes
Having satisﬁed ourselves of the validity of the FeynmanKac representation theorem,
we can now use it to represent the expectation given in (1.4), representing the price
of a European option, as the solution to a secondorder linear parabolic PDE. The
ﬁrst point to note is that (1.4) involves discounting and so it is useful to make the
transformation
V (S, t) = e
−r(T−t)
u(S, t)
CHAPTER 1. INTRODUCTION 26
in equation (1.6) to obtain the PDE
∂V
∂t
+
1
2
σ
2
(S, t)
∂
2
V
∂S
2
+µ(S, t)
∂V
∂S
−rV = 0,
which we have shown can be represented as the conditional expectation
V (S, t) = E
P
S,t
_
e
−r(T−t)
h(S
T
)
¸
.
However, note that the expectation in (1.4) is taken under the riskneutral measure
Q and so the corresponding PDE representation of (1.4) is given by
∂V
∂t
+
1
2
σ
2
S
2
∂
2
V
∂S
2
+rS
∂V
∂S
−rV = 0, (1.8)
with the following condition
V (S, T) = h(S), (1.9a)
V (0, t) = h(0)e
−r(T−t)
, (1.9b)
V (S, t) →h(S)e
−r(T−t)
as S →∞, (1.9c)
where we have used the riskneutral process (1.5). Note that in what follows this
shall be referred to as in the BlackScholes equation (which should also be credited
to Merton). Moreover, if we assume a stochastic process of the much more general
form (1.7), then the corresponding (generalised) BlackScholes equation obtained via
the FeynmanKac formula is given by
11
L
BS
(V ) =
∂V
∂t
+
1
2
σ
2
(S, t)
∂
2
V
∂S
2
+rS
∂V
∂S
−rV = 0, (1.10)
with the same boundary conditions as previously, i.e. (1.9).
However, it can be shown that standard FeynmanKac type results only hold under
(quite restrictive) analytic conditions on the coeﬃcients of the SDE and PDE, as
sumptions that are often not satisﬁed by many models used in practise. Remarkably,
this problem is often glossed over or simply not mentioned in the literature. What
follows is a brief overview of the some of these analytic conditions. In some sense the
11
Again note the independence of the realworld drift µ(S, t).
CHAPTER 1. INTRODUCTION 27
behaviour of the models presented in this thesis can be attributed to the failure of
the coeﬃcients of the relevant equations to satisfy the conditions outline below.
In order for the conditional expectation (1.4) to be the unique classical solution to the
BlackScholes equation (1.10) with the conditions (1.9) then the diﬀusion coeﬃcient
σ(S, t) must be suﬃciently regular. More precisely, it must be locally Lipschitz, i.e.
[σ(S
1
, t) −σ(S
2
, t)[ ≤ C[S
1
−S
2
[
for some C > 0, and also satisfy a linear growth condition in S, i.e.
[σ(S, t)[ ≤ D(1 +[S[)
for some constant D > 0.
12
Another condition is that the operator L
BS
must be
uniformly elliptic, meaning (in this onedimensional situation) that the coeﬃcient
σ(S, t) must be strictly positive at every point in the solution domain (S, t) ∈ Ω
[0, T], where Ω is the domain of the process S
t
, for example Ω = ¦S > 0¦ for geometric
Brownian motion. In other words, we have the restriction that
σ
2
(S, t) > 0 ∀(S, t),
i.e. the diﬀusion coeﬃcient σ
2
(S, t) cannot degenerate be zero. Note that even in the
simplest cases, such as geometric Brownian motion where σ(S, t) = σS, the volatility
term degenerates in certain regions of state space. Speciﬁcally lim
S→0
σ(S, t) = 0.
We can avoid this diﬃculty here (and also in many other more general situations)
by making the change of variable x = log S giving σ(x, t) = σ which is no longer
degenerate.
1.3.5 American options
Unlike European options discussed in section 1.3.1, American options have the extra
feature that they can be exercised at any time prior to expiry, T. The time γ at which
12
The stochastic process derived in chapter 2 can be seen to exhibit singular behaviour and, as
such, these conditions are no longer satisﬁed. Hence, we are no longer in a regime where standard
results from SDE and PDE theory can be applied. In addition, here the nonLipschitz nature of the
coeﬃcients means that the solutions to the corresponding SDE need no longer remain continuous;
jumps may be seen at the location where the diﬀusion coeﬃcient becomes singular.
CHAPTER 1. INTRODUCTION 28
the option is exercised is called the exercise time and because the market cannot be
anticipated, the holder of the option needs to decide whether to exercise at each
point in time based only on the information up to time t ≤ T (i.e. the information
contained in the ﬁltration T
t
).
The terms European and American were ﬁrst coined in Samuelson (1965) and the
story behind their naming is noteworthy. According to a private communication
with Robert C. Merton, Samuelson visited many practitioners on Wall Street prior to
writing his paper. One of his industry contacts explained to him that there were two
types of options available, one more complex (that could be exercised early) and one
much simpler (that could only be exercised at expiry). The practitioner commented
that only the more sophisticated European mind (as opposed to the American mind)
could understand the former. In response, when Samuelson (an American) wrote the
paper, he used the European and American preﬁxes but reversed the ordering.
If the payoﬀ proﬁle is given by h, and the holder of the American option decides to
exercise early then she receives the amount h(S
γ
) at time γ. Using the theory of
optimal stopping (cf. Peskir and Shiryaev, 2006), the unique noarbitrage price of an
American option can be shown to be given intuitively by
V (S, t) = sup
t≤γ≤T
E
Q
S,t
_
e
−r(γ−t)
h(S
γ
)
¸
, (1.11)
i.e. the supremum of the expected value of the discounted payoﬀ over all random times
γ that are stopping times with respect to the ﬁltration generated by the Brownian
motion used to specify the dynamics of the underlying process for S
t
. This is a rather
intuitive deﬁnition of the American option price.
Immediately from the deﬁnition (1.11) we have the inequality
V (S, t) ≥ h(S) (1.12)
since the stopping time γ = t is included in the supremum. This is a natural condition
since if V (S, t) < h(S) then there would be an obvious instant arbitrage at time t.
CHAPTER 1. INTRODUCTION 29
In addition, choosing γ = T gives the further inequality
V (S, t) ≥ V
E
(S, t),
where V
E
(S, t) is the corresponding European option price. Again, this is intuitive,
since an American option gives its holder more rights than the corresponding Euro
pean option with the same payoﬀ function and expiration date.
Another point to note is that when pricing American options we cannot, without
loss of generality, set the interest rate to zero, which can be done for their European
counterparts. Pricing American derivatives is mathematically more involved than
the European case and closedform expressions for American option prices are rarely
obtained. However, it can be shown by noarbitrage arguments that, for nonnegative
interest rates and no dividends, the price of an American call option is the same as
its corresponding European option (see, for example, prop. 7.14 of Bj¨ ork, 2004). In
other words, the supremum in expression (1.11) is attained for the stopping time
γ = T when considering the payoﬀ function of a call option. Thus, the price of an
American call reduces to the price of a European call, which does have an explicit
formula, ﬁrst derived by Black and Scholes (1973).
It can also be shown that the price of an American put option is, in general, strictly
higher than the price of the corresponding European put option. Indeed it can be
seen (directly from its wellknown analytic expression) that the European put option
price crosses below the payoﬀ function (1.2) for suﬃciently small S, violating the
condition (1.12). Hence the value of the American contract cannot coincide with
that of its European counterpart. We therefore use a put option as our canonical
example of an American option throughout the remainder of this thesis.
1.3.6 Optimal stopping problems
The observant reader may have noticed already that there is a strong link between
pricing American options and optimal stopping problems. When faced with an optimal
CHAPTER 1. INTRODUCTION 30
stopping problem, there are two facets of the solution that we are most interested
in. The ﬁrst is to determine the price of the option V (called the value function in
optimal stopping terminology) and the second to determine the optimal strategy for
the option holder, in other words to determine the stopping time that realises the
supremum in (1.11). Determining the value function will be discussed shortly, but
ﬁrst we state a key result from the theory of optimal stopping. If the function h is
continuous, in addition to some other technical conditions,
13
then the supremum is
attained for the stopping time
γ
∗
:= inf¦u ≥ t : V (S
u
, u) = h(S
u
)¦,
i.e. the ﬁrst time that the price of the American option drops down to the value of
its payoﬀ. Alternatively, and more practically, the optimal stopping time γ
∗
can be
formulated as the ﬁrst exit time from the continuation region deﬁned by
( := ¦(S, t) : V (S, t) > h(S)¦,
i.e. as
γ
∗
:= inf¦u ≥ t : (S
u
, u) / ∈ (¦.
The continuation region is so named due to the fact that in this region it is not
optimal to exercise the option. Clearly, if the value V (S, t) at some time t is strictly
larger than the payoﬀ proﬁle h(S), then it is not optimal to exercise the option.
1.3.7 Freeboundary problems
Analogous to the FeynmanKac representation theorem for European options (out
lined in section 1.3.3), the price of American options can be shown to satisfy partial
diﬀerential inequalities. For a nonnegative payoﬀ function h, the price of an American
13
See, for example, Peskir and Shiryaev (2006)
CHAPTER 1. INTRODUCTION 31
option as deﬁned in (1.11) is given by the solution to the following linear complemen
tarity problem:
V (S, t) ≥ h(S, t), (1.13a)
L
BS
(V ) =
∂V
∂t
+
1
2
σ
2
S
2
∂
2
V
∂S
2
+rS
∂V
∂S
−rV ≤ 0, (1.13b)
L
BS
(V ).
_
h(S, t) −V (S, t)
_
= 0, (1.13c)
to be solved in the entire domain ¦(S, t) : S > 0, 0 ≤ t ≤ T¦ with the ﬁnal condition
V (S, T) = h(S).
Further to this, it can be shown that the BlackScholes equation holds at all points in
the continuation region and that at the boundary of the continuation region, we must
apply the smooth pasting or smooth ﬁt principle.
14
This principle states that the value
function V (S, t) must be at least C
1,1
diﬀerentiable,
15
not only in the continuation
regions, but also over the boundary of the continuation regions, denoted by ∂(. It
also transpires that for a standard American put option there is an increasing function
S
f
(t), the free boundary, separating the continuation region from the stopping region,
compare Jacka (1991). As such the linear complementarity problem (1.13) can be
formulated as the freeboundary problem
16
∂V
∂t
+
1
2
σ
2
S
2
∂
2
V
∂S
2
+rS
∂V
∂S
−rV = 0, (1.14a)
V (S
f
, t) = K −S
f
, (1.14b)
V
S
(S
f
, t) = −1, (1.14c)
V (S, T) = (K −S)
+
, (1.14d)
V (S, t) →0 as S →∞, (1.14e)
to be solved in the domain ¦(S, t) : 0 ≤ t ≤ T, S > S
f
(t)¦, in other words the
boundary of the domain is to be solved as part of the problem. This implies that for
S > S
f
(t) the value V (S, t) must satisfy V (S, t) > (K − S)
+
, and for S ≤ S
f
(t) the
14
In fact the principle of smooth ﬁt in probability, the principle of no arbitrage in ﬁnance and the
conservation of energy law in the physical sciences can be seen as diﬀerent formulations of the same
principle. This is alluded to in Peskir (2005b).
15
At least for points at which the payoﬀ proﬁle h(S, t) is C
1,1
diﬀerentiable.
16
See for example Karatzas and Shreve (1998)
CHAPTER 1. INTRODUCTION 32
value satisﬁes V (S, t) = (K−S)
+
. Furthermore, the existence and uniqueness of the
freeboundary problem (1.14) can be proved. In addition, for put options without
dividends, Chen et al. (2008) have recently proved the convexity of the resulting free
boundary.
Explicit solutions to parabolic freeboundary problems are rare, however it can be
shown (cf. Jacka, 1991) that the American put option free boundary S
f
(t) is a mono
tonically increasing function and that it approaches K as t approaches T. The asymp
totic behaviour of S
f
(t) for times close to expiry can also be determined and indeed
this shall be expounded upon in further detail in chapter 9.
1.4 Supply and demand economics
Many of the models presented in this thesis make assumptions about the structure
of the markets and the intentions of the participants of these idealised markets. This
motivates a brief discussion of how prices are actually formed in these markets, in
short a discussion of supply and demand, the backbone of a market economy.
Starting with the basics, a market is a place where buyers (providing demand) and
sellers (providing supply) meet. In a free market, prices are determined solely by
the interaction of demand and supply; nothing more, nothing less. In addition, all
being equal, there will be more demand for an asset at a lower price than at a
higher price and, hence, we should expect an inverse relationship between price and
quantity demanded. Conversely, an increase in price will usually lead to an increase
in the number of people wishing to sell at that price, hence we should expect a
positive relationship between price and supply. In the economics literature, these
relationships are often called the law of demand and the law of supply. In a market,
the price at which supply matches demand is often called the equilibrium price or
market clearing price, so called because it is at this price that all the surpluses are
cleared from the market and the forces of demand and supply are not acting to change
this equilibrium. If disequilibrium exists, then the forces of demand and supply will
CHAPTER 1. INTRODUCTION 33
automatically adjust the market to equilibrium. With excess demand, prices will be
forced upwards due to the shortage that exists, and with excess supply, prices will be
forced downwards, due to the surplus that exists.
An important concept crucial to the models discussed in this thesis is that of elasticity.
At its heart this concept is a purely mathematical one which aims to measure the
responsiveness of one variable to a change in another variable. More speciﬁcally given
any functional relationship y = f(x) the point elasticity, , is deﬁned as
=
dy/y
dx/x
=
dy
dx
x
y
=
d(log y)
d(log x)
,
i.e. the ratio of percentage changes. Similarly, given a function of more than one
variable y = f(x
1
, x
2
, . . . , x
n
) the partial point elasticities are given by
i
=
∂y
∂x
i
x
i
y
=
∂(log y)
∂(log x
i
)
.
Applied to the economics of supply and demand the price elasticity of demand (PED)
is deﬁned as
PED =
dq/q
dp/p
=
dq
dp
p
q
,
where q is the quantity demanded of an asset and p is the price per unit of that asset.
The PED measures the responsiveness of the quantity demanded to the change in
price. PED > 1 implies that the good is price elastic, PED < 1 implies that the
good is price inelastic and when PED = 1 we have unit elasticity. The limiting cases
PED = 0 and PED = ∞ imply that the asset is perfectly price inelastic and elastic
respectively. The price elasticity of supply (PES) is deﬁned similarly.
An important point to note at this stage is that elasticity and liquidity are not the
same, though there is a tendency to confuse the two. Elasticity deﬁnes a relationship
between price and the quantity demanded (as deﬁned above), whereas liquidity is
concerned with the availability to trade the underlying asset at a given price. How
ever (unlike elasticity) liquidity is not a welldeﬁned concept, hence there is much
ambiguity in the connection between the two concepts. The next section explores the
concepts of liquidity in much more detail.
CHAPTER 1. INTRODUCTION 34
1.5 Liquidity
Risk can be classiﬁed into the following categories
17
• Market risk,
• Credit risk,
• Model risk,
• Operational risk,
• Liquidity risk.
The standard models implicitly assume that the only risk experienced by a trader
is that due to the uncertain nature of the market. More relevant to this thesis,
these standard models assume that the trader will not experience any liquidity risk,
implicitly assuming a level of liquidity that is without limits. Liquidity risk arises
in situations where a party interested in trading an asset cannot do so because she
cannot ﬁnd a willing counterparty to that trade. Liquidity risk becomes particularly
important to parties who are about to hold or currently hold an asset, since it aﬀects
their ability to trade. In fact one of the most important attributes of ﬁnancial markets
is to provide immediate liquidity to investors. Of course, some markets are more liquid
than others, and the liquidity of a given market varies over time and in addition can
dramatically dry up in times of crisis.
Recent crises in the ﬁnancial markets have triggered studies on the subject of market
liquidity. For example, the stock market crises in October 1987 and 1989, the Asian
crisis in 1997 and the problems at LongTerm Capital Management Fund (LTCM)
led the Committee on the Global Financial System to conduct several studies dis
cussing the importance of liquid ﬁnancial markets, including Bank for International
Settlements (1999) and Bank for International Settlements (2001).
17
See Protter (2006).
CHAPTER 1. INTRODUCTION 35
1.5.1 Deﬁning liquidity
Market liquidity is often associated with the ability to quickly buy or sell a particular
item without causing a signiﬁcant movement in the price. However, the concept
of liquidity is multifaceted and illdeﬁned. Many researchers have attempted to do
so but the best that can be done is to classify its many dimensions. Kyle (1985)
describes market liquidity in terms of three attributes, namely the tightness, depth
and resilience of the market. Liu (2006) identiﬁes four dimensions to liquidity, namely,
trading quantity, trading speed, trading cost, and price impact. Alternatively, Sarr
and Lybek (2002) state that liquid markets exhibit ﬁve characteristics: tightness, i.e.
having low transaction costs, such as a small bidask spread as well as other implicit
costs; immediacy, i.e. the speed with which orders can be executed, reﬂecting the
eﬃciency of the trading, clearing and settlement systems; depth, i.e. the existence of
abundant orders both above and below the price at which an asset currently trades;
breadth, i.e. orders are both numerous and large in volume with minimal impact on
prices; and ﬁnally resiliency, i.e. new orders ﬂow quickly to correct order imbalances.
Clearly, liquidity is a tricky concept to deﬁne (let alone measure), and due to this
multidimensional nature comparing individual assets liquidities is also problematic,
since one asset could be more liquid along one dimension of liquidity while the other
is more liquid in a diﬀerent dimension. One particular interpretation of liquidity in
the literature ﬁts nicely with the philosophy of this thesis; Howison (2005) states
that market liquidity can manifest itself in three possible forms. First, there is a
diﬀerence between the prices for buying and selling the asset, the socalled bidask
spread. Second, the price paid for trading the asset depends on the quantity traded,
due to limited availability of a stock at the quoted price. In fact, even for a highly
liquid market, trading beyond the quoted depth of the market usually results in a
higher purchase price (or a lower selling price) for part, if not all, of the trade; this
is often termed the liquidation cost. Third, and most relevant to this thesis, is that
the action of a large trade may itself impact the price, independent of all the other
factors aﬀecting the price dynamics; this is termed price impact.
CHAPTER 1. INTRODUCTION 36
1.5.2 Measuring liquidity
Because there are many dimensions of liquidity, there is no single method for mea
suring it. Measures which are often used in the empirical literature on liquidity and
asset pricing include the bidask spreads, various measures of the price impact of
order ﬂow, and various measures of order ﬂow. Measures of the price impact of or
der ﬂow include price changes regressed on signed volume, or absolute price changes
regressed on absolute volume, or daily changes regressed on daily volume. Measures
of volume include numbers of trades and daily volume measured in dollars. Of all
these measures, the price impact of order ﬂow is perhaps the most widely used, the
advantage of this measure being that it is based on the actual observed price changes
associated with trades. However, despite the advantages of using the price impact of
order ﬂow as a measure of liquidity, tricky econometric issues, such as measurement
error, selection bias and simultaneity bias are involved when using this measure.
Sarr and Lybek (2002) classify the existing liquidity measures into four categories.
18
The ﬁrst is transaction cost measures that capture the costs of trading ﬁnancial
assets and trading frictions in secondary markets. One particularly intuitive measure
of transaction costs is the percentage bidask spread, deﬁned as
BAS = 2
_
P
A
−P
B
P
A
+P
B
_
,
where the ask price P
A
and bid price P
B
can be calculated from the quotes on the
market or using a weighted average of actual executed trades over a period of time,
the latter being a better estimate of the actual transaction costs since trades may not
take place at the actual quoted prices, in this case the spread is called the realised
spread. In the second category are volumebased measures that attempt to distinguish
liquid markets by the volume of transactions compared to the price variability, this
is primarily used to measure the breadth and depth of the market. Trading volume
is traditionally used to measure the existence of numerous market participants and
transactions and is deﬁned as
Vol =
n
i=1
P
i
Q
i
(1.15)
18
See Sarr and Lybek (2002) for a good review of many examples of each class of liquidity measure.
CHAPTER 1. INTRODUCTION 37
where Vol is the dollar volume traded, P
i
and Q
i
are prices and quantities of the ith
trade during a speciﬁed period. This can be given more meaning by relating it to
the outstanding volume of the asset. The resulting turnover rate gives an indication
of the number of times the outstanding volume of the asset changes hands. The
turnover can thus be deﬁned as
TO =
Vol
NP
where Vol is the trading volume deﬁned in (1.15), N is the outstanding stock of the
asset and P is the average price of the n trades in (1.15). There are many other
volumebased measures. The third category of liquidity measures are equilibrium
pricebased measures that try to capture orderly movements towards equilibrium
prices; in the main these attempt to measure resiliency of the market. The fourth
and ﬁnal category, and the most relevant to the focus of this thesis, are marketimpact
measures that attempt to diﬀerentiate between price movements due to the degree
of liquidity from other factors, such as general market conditions or arrival of new
information; these attempt to measure both elements of resiliency and speed of price
discovery.
However, clearly no single measure can manage to fully capture the multifaceted na
ture of liquidity, and as such there is no universally accepted measure of liquidity.
Most of the existing literature attempting to measure liquidity has focused on the
diﬀerent dimensions of liquidity individually. In fact this problem of no universal liq
uidity measure has resulted in many unanswered questions in market microstructure
theory, which focuses on determining the processes by which information is incorpo
rated into prices. One such question is whether liquidity is priced in asset returns.
For example Amihud and Mendelson (1986) (who simply use the bidask spread)
and Datar et al. (1998) (who instead use the turnover rate) argue that liquidity is
priced, whereas others, such as Chalmers and Kadlec (1998), Chen and Kan (1995)
and Eleswarapu and Reinganum (1993) suggest that it is not.
More recently Liu (2006) introduced a new measure of liquidity (called the standard
ised turnoveradjusted number of zero trading volumes over the prior 12 months) that
CHAPTER 1. INTRODUCTION 38
aims to capture multiple dimensions of liquidity. Using this measure Liu (2006) out
lines a twofactor, liquidity risk adjusted capital asset pricing model (CAPM) that
well explains the crosssection of stock returns, (possibly) answering the question
whether liquidity is priced. In addition, the new twofactor CAPM model is able to
account for the booktomarket eﬀect, which the Fama and French (1996) threefactor
model fails to explain.
1.6 Price formation
We have alluded to the fact that the price of ﬁnancial instruments may be considered
as entirely dependent on supply and demand. However knowledge about how these
prices are actually formed in the market are of great interest, since we wish to see ex
actly whereabouts in the price formation process liquidity issues become important.
From a market microstructure perspective, price movements are caused primarily
through the arrival of information. The dynamics by which this information is incor
porated into the current price is addressed in the market microstructure literature,
where many models of price formation have been proposed; for an overview of this
topic see O’Hara (1995). Such models are not referred to speciﬁcally in this thesis
and so it suﬃces to describe brieﬂy the role of some of the more important market
participants.
One of the most important members of any ﬁnancial market are the socalled market
makers. These are individuals or ﬁrms that will take both long and short positions
in a given security in order to facilitate trading, and thus add to the liquidity and
depth of the market. The marketmaker accepts a certain level of risk in holding the
ﬁnancial instrument or commodity but hopes to be compensated by making a proﬁt
on the bidask spread.
In the United States, many markets have oﬃcial market makers for each given se
curity, known as specialists. Their main function being to provide the other side of
trades when there are shortterm buyandsellside imbalances in customers orders.
CHAPTER 1. INTRODUCTION 39
In return, the specialist is granted various informational and trade execution advan
tages. On the London Stock Exchange (LSE) there are oﬃcial market makers for
many securities (except for the largest and most heavily traded companies, which
instead use an automated system called SETS). On the LSE one can always buy and
sell stock; each stock always has at least two market makers and they are obliged to
deal. This is in contrast with much smaller order driven markets in which it can be
extremely diﬃcult to determine at what price one would be able to buy or sell any
of the many illiquid stocks.
In traditional exchange ﬂoor markets the burden of providing liquidity is given to
market makers or specialists. Nowadays, however, most ﬁnancial markets have be
come fully electronic and operate on what is called a matched bargain or order driven
basis. In these markets, when a buyer’s bid price meets a seller’s oﬀer price the stock
exchange’s matching system will decide that a deal has been executed. In an order
driven market there are numerous types of orders that can be placed, each catering to
the diﬀerent needs of diﬀerent market participants. The two main type of orders are
the market order, which is an order to buy or sell immediately at the best available
price, and as such gives no guarantee on the price but is guaranteed to be executed
immediately. Alternatively we have limit orders which are not to be executed unless
the speciﬁed price is met (or bettered) by current bids or asks. Here, we are not
guaranteed execution but we are guaranteed price. It should, however, be noted that
limit orders often incur higher commission fees. Further, in these orderdriven mar
kets liquidity now becomes selforganised, in the sense that any agent can choose, at
any instant of time, either to provide or to consume liquidity; providing liquidity by
posting limit orders or consuming liquidity by issuing a market order.
The introduction of electronic markets has seen a sharp increase in another type of
market participant, the program trader. A program trader is one who uses a computer
to automate his trades. This may be to exploit arbitrage opportunities such as index
arbitrage (the misalignment of the price of an index and the sum of its constituent
stocks) or to perform portfolio insurance, the automated execution of a deterministic
CHAPTER 1. INTRODUCTION 40
hedging strategy. Program traders are thought to have been a contributing factor
of the October 19, 1987 market crash
19
and to be responsible for an increased stock
market volatility, since they quickly dump large orders on the market at critical times.
These large orders can contribute to the existing momentum of the market, thereby
increasing market volatility. This shall be seen in a more mathematical framework
in chapter 3.
1.7 Option pricing in illiquid markets: a literature
review
Authors such as Kreps (1979) and Bick (1987, 1990) have placed the classical Black
ScholesMerton formulation into the framework of a consistent model for market
equilibrium with interacting agents having very speciﬁc investment characteristics
(see section 1.6). Moreover Bick (1987, 1990) showed how geometric Brownian mo
tion, one of the fundamental assumptions of the BlackScholesMerton model, can be
derived in a general equilibrium model with pricetaking agents.
Furthermore F¨ ollmer and Schweizer (1993) were the ﬁrst to use a microeconomic
approach to construct diﬀusion models for asset price movements. They deﬁne in
formation traders who believe in a fundamental value of the asset, and noise traders
whose demands are from hedging requirements. They derived equilibrium diﬀusion
models for the asset price based on interaction between the two. Many of the models
discussed in this thesis such as Platen and Schweizer (1998), Sircar and Papanicolaou
(1998) and Sch¨ onbucher and Wilmott (2000) were inspired by the temporary equi
librium approach of F¨ ollmer and Schweizer (1993). Starting from a microeconomic
equilibrium and deriving a diﬀusion model for stock prices which endogenously in
corporates the demand due to hedgers and in particular delta hedgers.
19
Jacklin et al. (1992) argue that one of the causes was actually information about the extent
of portfolio insurancemotivated trading suddenly becoming known to the rest of the market. This
prompted the realisation that assets had been overvalued because the information content of trades
induced by hedging concerns had been misinterpreted. Consequently, general price levels fell sharply.
CHAPTER 1. INTRODUCTION 41
The literature on liquidity falls broadly into two approaches. The ﬁrst involves the
price impact due to a large trade. In such models the large trader can move the price
by his actions. Jarrow (1992, 1994) provided a discretetime model which allows
the large trader to impact the market via some reaction function. He showed that
the price of a derivative in this framework must be equal to the hedge cost, but
this cost, and hence the price, is dependent on the large trader’s position in the
underlying and the derivative asset; leading to nonlinearity. However in markets that
allow large traders to impact the price of the asset there is the possibility of price
manipulation and so called market corners and market squeezes. A market corner
is a successful eﬀort of a trader to manipulate the price of a futures contract by
gaining eﬀective control over trading in the futures and the supply of the deliverable
goods. In a market squeeze, the trader achieves control by disruption in the supply of
the cash commodity. Although price manipulation violates the Commodity Exchange
Act, there have been many examples of such activities, especially in (less regulated)
developing markets. An example of a market corner is the Hunt silver manipulation
of 19791980, a detailed and readable account of which can be found in Williams
(1995). An example of a market squeeze is the (alleged) soybean manipulation of 1989
for which more details can be found in Pirrong (2004). However in the theoretical
framework proposed by Jarrow (1992, 1994) it was showen that to prevent any such
manipulation the price impact mechanism must not exhibit any delay. In addition a
suﬃcient condition to exclude proﬁtable market manipulation (in discretetime) was
given, i.e. that the price mechanism must be independent of the history of the trades,
and only dependent on the current position of the trades. Bank and Baum (2004)
later extended Jarrow’s results to continuous time.
Moreover, in the presence of price impact, it is not clear that an option is still perfectly
replicable; hence it is no longer straightforward how to derive option prices from
the prices of the underlying. Frey and Stremme (1997) studied the perturbation of
volatility induced by a delta hedging strategy for a European option whose price is
given by a classical BlackScholes formula with constant volatility. They concluded
CHAPTER 1. INTRODUCTION 42
that if a hedging strategy is used which does not take into account the feedback eﬀect
(which we term ﬁrstorder feedback), then it is not possible to replicate perfectly
an option, and hence there is still risk associated with hedging in illiquid markets.
They did show, however, that increasing heterogeneity of the distribution of hedged
contracts reduces both the level and price sensitivity of this unhedged risk. Frey
(1998, 2000) then showed that if feedback is taken into account in a more general
hedging strategy (which we term full feedback), then it is possible to replicate an
option perfectly (provided certain conditions on market liquidity and the nonlinearity
of the payoﬀ condition are satisﬁed). In the discretetime framework of Jarrow (1994),
the question as to whether options could be perfectly replicated in a ﬁnitely elastic
market reduces to solving (recursively) a ﬁnite number of equations. In the continuous
time framework of Frey (1998), this can be characterised more succinctly as the
solution of a nonlinear PDE, for which Frey (1998) gave existence and uniqueness
results. These results, however, place a heavy restriction on the amount of market
illiquidity that the model allows and rely on the terminal payoﬀ being suﬃciently
smooth, both of which can be seen as undesirable restrictions.
20
Frey and Patie
(2002) extended the work of Frey (2000) with an asset dependent liquidity parameter
which attempts to incorporate so called liquidity drops, whereby market liquidity
drops if the stock price drops, the aim being to reproduce, more eﬀectively, the
volatility smile.
Other continuous time models similar to Frey (1998) include Sch¨ onbucher and Wilmott
(2000), who used a market microstructure equilibrium model to derive a modiﬁed
stochastic process under the inﬂuence of price impact. The PDEs derived by these
latter authors correspond to those derived in chapter 2 of the present study. Sircar
and Papanicolaou (1998) derived a slightly diﬀerent nonlinear PDE that depends on
the exogenous income process of the reference traders and the relative size of the
program traders. Platen and Schweizer (1998) proposed a model using an approach
that attempted to explain the volatility smile and its skewness endogenously and
20
For further discussion on these restrictions see chapter 4.
CHAPTER 1. INTRODUCTION 43
Mancino and Ogawa (2003) proposed a very similar model in the same vein. Lyukov
(2004) then extended the model of Platen and Schweizer (1998) with more realistic
assumptions about market equilibrium conditions (taking into account the presence
of a market maker) and also obtained a very similar nonlinear PDE to that derived
in chapter 2. Another ‘tweak’ of these models was made by Liu and Yong (2005) who
attempted to regularise the PDE close to expiry. The majority of these models will
be considered in more detail in chapter 7.
The second approach to liquidity seen in the literature involves the price impact due
to the immediacy provisions of market makers. In these models, supply and demand
are equalised by the market maker in the shortterm market. The approach is relevant
if an agent wishes to trade a large amount in a short time. These models have been
considered by Rogers and Singh (2006) and Cetin and Rogers (2007), amongst others,
who propose a series of independent auctions. The main diﬀerence with the ﬁrst class
of models is that these are now local in time models, without longterm eﬀects, i.e.
the actions of the traders do not inﬂuence the underlying stochastic process. These
models eliminate the feedback eﬀects discussed above and, as such, they are concerned
more with the liquidation cost than permanent price impact. Bakstein and Howison
(2003) adopted a similar approach to Rogers and Singh (2006) but the former study
leads to feedback eﬀects, which the latter study was trying to avoid. Another model
in this category is the work of Cetin et al. (2004), who modelled the liquidation cost
as dependent on the quadratic variation of the trading strategy which again leads to
a nonlinear PDE when considered in continuous time.
Other notable work in the area of liquidity and price impact includes Agliardi and
Andergassen (2001) who extended the work of Sch¨ onbucher and Wilmott (2000) and
Frey (1998) to include sink transaction costs as an additional source of illiquidity. Mo
tivated by empirical evidence, Esser and Moench (2003) extended the work of Frey
(2000) to incorporate stochastic liquidity into the price impact framework. More
recently, HenryLabord´ere (2004) incorporated the feedback model of Sch¨ onbucher
and Wilmott (2000) into the portfolio optimisation of Markowitz (1959) to ﬁnd that
CHAPTER 1. INTRODUCTION 44
portfolio optimisation has the eﬀect of reducing market volatility and thus the price
of options in that market. Brennan and Schwartz (1989) also analysed the transfor
mation of market volatility under the impact of portfolio insurance and under the
assumption of CRRA utility. They showed an increase in BlackScholes implied mar
ket volatility between 1% and 7% for values for the fraction of the market subject
to portfolio insurance varying between 1% and 20%. Cvitanic and Ma (1996) and
Cuoco and Cvitanic (1998) studied a diﬀusion model for price dynamics when the
drift and volatility coeﬃcient are functions of the large traders’ trading strategy. Fi
nally Jonsson and Keppo (2002) derived a somewhat diﬀerent nonlinear PDE, using
a model with an exogenously deﬁned exponential price eﬀect function.
According to Rogers and Singh (2006) and Cetin and Rogers (2007), the price impact
models of a large trader have numerous shortcomings. The ﬁrst is the socalled free
round trip phenomenon, which can result in the possibility of market manipulation
by a large agent. Recall that Jarrow (1992, 1994) gave suﬃcient conditions to ex
clude proﬁtable market manipulation, but only in the discretetime framework, for a
detailed discussion of how things can go wrong in continuous time see Sch¨ onbucher
and Wilmott (2000). The second, and by far the most important problem with these
models is that if we allow the action of one large agent to aﬀect the price, then we
must allow the actions of all large agents to aﬀect the price and furthermore the eﬀect
of hedging a multitude of options on the underlying. Clearly this would result in an
impossibly cumbersome problem.
There has also been some recent work using alternative pricing paradigms in place
of portfolio replication, such as that by Bank (2006) who attempted to price options
on illiquid underlyings using the utility indiﬀerence of the market maker. This ap
proach leads to a stochastic optimisation problem and the HamiltonJacobiBellman
equation, which is also inherently nonlinear.
One of the few attempts to analyse the aforementioned models from an empirical
standpoint was carried out recently by Sanfelici (2007) who considered the model of
CHAPTER 1. INTRODUCTION 45
Mancino and Ogawa (2003) (amongst others) and attempted to calibrate the model to
market data, comparing the results with other popular models. Sanfelici concluded
that the nonlinear models above can contribute to the explanation of the implied
volatility smile but not as well as the other possible explanations, such as jump
diﬀusion or stochastic volatility, due in the most part to the models’ limited capability
to reproduce skewed probability distributions. However, the study did ﬁnd that the
model of Mancino and Ogawa (2003) was more stable through time and consistent
with market data.
Any other relevant literature will be discussed in the body of the thesis where it is
appropriate.
1.8 Introduction to perturbation methods
Perturbation methods, also known as asymptotic methods, are a collection of mainly
analytical techniques that can be used to solve (or simplify) mathematical problems
involving a small or large parameter. They are used extensively throughout this
thesis and so a brief introduction, including their rich history, is outlined below.
Two phases in the history of asymptotics may be identiﬁed. The ﬁrst, often called
classical asymptotics dates back to the work of Poincar´e (1886) on the farﬁeld be
haviour of linear ordinary diﬀerential equations. These techniques are primarily con
cerned with coordinate expansions, i.e. asymptotic expansions in which the indepen
dent variable, x say, plays the role of the large or small parameter. Taylor expansions
are the most well known of this class of expansions. Coordinate expansions can also
be used to investigate the behaviour of diﬀerential equations near any singular point
x
0
as x −x
0
→0 or for large values of the independent variable, i.e. as x →∞, both
of which we exploit in this thesis. More recent developments in asymptotic theory
have been associated with so called parametric expansions. These ideas have been
developed alongside the theory of ﬂuid dynamics (especially, but not exclusively),
a theory in which the governing equations are highly nonlinear and as such have
CHAPTER 1. INTRODUCTION 46
tractability only in the simplest of situations. Exploiting the smallness of certain
parameters and seeking a power series solution in the smallness parameter can often
reduce the original system of equations to a much simpler asymptotic set of equations,
whose solutions (both analytical or numerical) will often be much easier to ﬁnd. The
relative simplicity of the asymptotic solution does, however, come at the expense
of the approximate nature of the results. Nevertheless for many practical purposes
the results obtained can be made suﬃciently accurate, and indeed can often provide
invaluable insight into the qualitative behaviour of the problem under investigation.
During the ﬁrst half of the twentieth century it was demonstrated that a number
of problems involving small or large parameters developed a pathological behaviour,
which we now refer to as singular perturbations. This behaviour stemmed from the
fact that in certain regions of the solution domain we have a so called asymptotic
breakdown, where the power series expansion solution sought no longer becomes an
asymptotic series, under certain circumstances. An early attempt to tackle this prob
lem was due to Prandtl (1904) in his paper on boundarylayer theory. Prandtl’s idea
was to subdivide the solution domain into separate regions where diﬀerent asymp
totic forms apply. Over the next half century these ideas were developed by many
(Friedrichs (1954), Kaplun (1967), Kaplun and Lagersrom (1957), Cole (1968) and
van Dyke (1964) to name but a few) and so the method of matched asymptotic ex
pansions emerged. For a recent overview of how these techniques have been exploited
in ﬁnance to date see Howison (2005).
1.9 Layout of the thesis
The remainder of this thesis is organised as follows. Chapter 2 provides an heuristic
derivation of one of the more intuitive models attempting to incorporate liquidity into
option pricing theory and in addition shows how the majority of models introduced
in the current literature can be formulated in this framework. Chapter 3 investigates
the socalled ﬁrstorder feedback model (an exceptional case of a linear PDE) and
CHAPTER 1. INTRODUCTION 47
furthermore highlights interesting diﬀerences of both the option value and American
option earlyexercise boundary from the classical BlackScholesMerton case. Chap
ter 4 investigates the fully nonlinear full feedback model using both analytical and
numerical techniques. Chapter 5 takes a closer look at this model in a regime in which
it appears no classical solutions exist; here socalled phase plane analysis is found to
be useful in determining such behaviour. In chapter 6 the perpetual options of the full
feedback model are considered. Chapter 7 takes a look at the existing models in the
literature from a viewpoint of the results found in chapters 26. Chapter 8 considers
brieﬂy the socalled stock pinning eﬀect which appears to be well explained by the
models outlined in this thesis. Chapter 9 concerns itself with the related topic of the
British option, which can provide an investor with protection against the liquidity
issues discussed in the rest of the thesis; here we concern ourselves mainly with the
behaviour of the free boundary which exhibits some interesting qualitative diﬀerences
with the standard American option free boundary. Finally chapter 10 provides some
concluding remarks and ideas for future research.
Chapter 2
The Modelling Framework
In the end, a theory is accepted not because it is conﬁrmed by conven
tional empirical tests, but because researchers persuade one another that
the theory is correct and relevant.
 Fischer Black (19381995) in 1986
In this chapter we present an heuristic derivation of one particular class of model for
incorporating liquidity into option pricing theory. We also attempt to highlight the
links between the existing models and furthermore we transpose these models into
a single intuitive analytical framework. This has not previously been done in the
literature.
In order to provide a derivation of the primary governing PDEs considered in this
thesis, we present the following arguments, which are similar to those employed in
Lipton (2001) and Liu and Yong (2005). We shall assume the underlying process
to be a geometric Brownian motion (but this can be generalised to any stochastic
process)
dS = µSdt +σSdW
t
, (2.1)
where S is the price of the underlying, µ and σ are the (constant) drift and volatility
respectively and W
t
is a standardised Brownian motion. It is possible to add a forcing
48
CHAPTER 2. THE MODELLING FRAMEWORK 49
term, f(S, t), to the process, which is dependent on the stock price and time, i.e.
dS = µSdt +σSdW
t
+λ(S, t)df, (2.2)
where λ(S, t) is an arbitrary function. Note that at this stage no assumptions need
be made regarding the form of the functions λ(S, t) and f(S, t), and particular ﬁnan
cial interpretations can conveniently be postponed until certain manipulations are
complete.
Since f(S, t) is a function of S and t only, it is possible to incorporate the additional
contribution to the price dynamics into the drift and volatility coeﬃcients µ and σ.
We commence by using Itˆ o’s formula on the function f(S, t), to obtain
df =
∂f
∂t
dt +
∂f
∂S
dS +
1
2
∂
2
f
∂S
2
(dS)
2
+. . . ,
which substituting into (2.2), gives to leading order
1
_
1 −λ
∂f
∂S
_
dS =
_
µS +λ
∂f
∂t
_
dt +
λ
2
∂
2
f
∂S
2
(dS)
2
+σSdW
t
. (2.3)
In order to proceed further we require an expression for (dS)
2
, which can be obtained
by simply squaring equation (2.3) to yield, as dt →0
(dS)
2
=
σ
2
S
2
dt
_
1 −λ
∂f
∂S
_
2
+o(dt), (2.4)
where we have used the condition that (dW
t
)
2
→ dt as dt → 0. Substituting (2.4)
into (2.3), and with a little rearranging, we arrive at the following stochastic process,
analogous to (2.1):
dS = ˆ µ(S, t)dt + ˆ σ(S, t)dW
t
, (2.5)
where
ˆ µ(S, t) =
1
1 −λ
∂f
∂S
_
µS +λ
_
∂f
∂t
+
1
2
ˆ σ
2
∂
2
f
∂S
2
__
, (2.6a)
ˆ σ(S, t) =
σS
1 −λ
∂f
∂S
. (2.6b)
1
Note the term on the lefthandside and how, even at this early stage in the derivation, we begin
to observe possible singular behaviour, . We shall return to this issue numerous times throughout
the remainder of the thesis.
CHAPTER 2. THE MODELLING FRAMEWORK 50
We can interpret the function f(S, t) as a forcing mechanism on an underlying stochas
tic process which results in the process (2.1) being modiﬁed to the process (2.5). The
ﬁnancial interpretation of this forcing term will be considered in full detail shortly.
It is tempting to expect that the stochastic process (2.5) will exhibit some boundary
behaviour at the location of the singularity in the drift, i.e. when
∂f
∂S
= 1/λ, since
at this location the drift will become inﬁnite and consequently the process might
be contained within a ﬁxed domain. However the volatility of the process also be
comes singular at the same location and so it may be possible to move beyond the
‘boundary’. The interested reader is referred to Sch¨ onbucher and Wilmott (2000)
who provide an analysis of the behaviour of the modiﬁed stochastic process at such
singular locations.
We now turn our attention to option pricing under the modiﬁed stochastic process.
To do this we will use the wellknown Generalised BlackScholes equation (for a more
detailed derivation see for example Duﬃe, 1996), which leads to the following pricing
PDE for the modiﬁed stochastic process incorporating the aforementioned forcing
term
∂V
∂t
+
1
2
σ
2
S
2
_
1 −λ
∂f
∂S
_
2
∂
2
V
∂S
2
+rS
∂V
∂S
−rV = 0. (2.7)
Note that, consistent with standard BlackScholes arguments, the drift of the modiﬁed
process ˆ µ(S, t) does not appear in the option pricing PDE.
Thus far we have been deliberately vague about the ﬁnancial interpretation of the
forcing term in (2.1). In the context of markets with ﬁnite elasticity, we can deﬁne
f(S, t) to be the number of extra shares that should be held due to some deterministic
hedging/trading strategy and hence df(S, t) will specify the number of shares needed
to be bought or sold at time t and price S due to such a strategy. Also λ(S, t) can
be interpreted as some function dependent on how we choose to model the form of
price impact and liquidity; we shall return to this issue in section 2.3.
Markets are not complete to traders who do not have the opportunity to trade contin
uously. Only large institutions can trade close to continuously and so, in a complete
CHAPTER 2. THE MODELLING FRAMEWORK 51
market, options provide no extra trading opportunities to them. For small traders
however, options open up new trading possibilities, resulting in many large institu
tions selling the options to the small traders and then hedging the risk by replicating
the option. This results in a net long position (of the large institution) for stocks in
the market. In such a market there is a high demand for these replicating strategies
and it is not, therefore, unreasonable to assume that a trading strategy that could
impact the price signiﬁcantly is that of delta hedging. In this case the trading strat
egy f, in equation (2.7) should be set to an option delta, based on some form of
option V
∗
, i.e.
2
f = ∆
∗
=
∂V
∗
∂S
. (2.8)
This leads to an interesting question about which strategy the hedgers are assumed
to follow. A naive strategy would be if V
∗
were the BlackScholes value V
BS
and thus
distinct from the solution V of equation (2.7). This leads to the linear PDE
∂V
∂t
+
1
2
σ
2
S
2
_
1 −λ
∂
2
V
BS
∂S
2
_
2
∂
2
V
∂S
2
+rS
∂V
∂S
−rV = 0. (2.9)
This case we call ﬁrstorder feedback, which is analysed in detail in chapter 3. Another
(more interesting and challenging) case is when the hedger is assumed to be aware of
the feedback eﬀect and so would change the hedging strategy accordingly. We shall
call this case full feedback, which corresponds to the case when V
∗
≡ V , i.e. the
trading strategy adopted has to be found as part of the problem. This leads to the
fully nonlinear PDE,
∂V
∂t
+
1
2
σ
2
S
2
_
1 −λ
∂
2
V
∂S
2
_
2
∂
2
V
∂S
2
+rS
∂V
∂S
−rV = 0, (2.10)
which is dealt with in chapter 4. Note that although λ can be scaled out of (2.10)
using the simple substitution V (S, t) =
1
λ
w(S, t), this would then introduce λ into the
standard payoﬀ conditions, and we therefore chose not to do this. Note too that for
simplicity, similar to Liu and Yong (2005), we have assumed that European contingent
2
Note that this is for a net long position in the market, if there was a net short position then
we would set f = −∆
∗
. This case is considered in the model of stock pinning by Avellaneda and
Lipkin (2003) and will be discussed further in chapter 8.
CHAPTER 2. THE MODELLING FRAMEWORK 52
claims are settled by physical delivery of the underlying asset at maturity. In this
case we do not need to introduce liquidation costs at maturity into the replicating
portfolio. This is primarily due to the fact that the exact liquidation value is diﬃcult
to determine, since it depends on the liquidation strategy chosen by the investor;
optimal liquidation strategies are discussed, for example, in Almgren and Chriss
(2001). How we choose to close out the contracts is not just an academic exercise
since, as noted by Frey (1998), in relatively illiquid markets, such market traders who
know that some other market participant has to dissolve a large hedged portfolio
in the near future, can try to proﬁt from this information by front running the
anticipated trades.
Equation (2.10) has appeared in the literature several times with diﬀering forms of
the function λ(S, t) according to the modelling assumptions. The simplest case occurs
in Sch¨ onbucher and Wilmott (2000), who have λ(S, t) constant and a dimensionless
measure of the liquidity of the market. Frey (1998) has the similar form λ(S, t) =
ˆ
λS
where
ˆ
λ ∈ R is again some measure of the liquidity of the market. Another form
for λ(S, t) can be found in Liu and Yong (2005) in which λ(S, t) =
ˆ
λ(1 − e
−β(T−t)
)
where
ˆ
λ is a constant price impact coeﬃcient, T −t is time to expiry and β a decay
coeﬃcient. As far as we can ascertain there is little ﬁnancial justiﬁcation for this, and
it appears to be introduced for numerical expediency to avoid diﬃculties associated
with the growing option gamma as expiration approaches. Indeed, the precise manner
in which this is achieved is discussed in section 7.5, and this highlights that this factor
fundamentally changes the optionprice dynamics. In what follows it is assumed for
the most part that λ(S, t) is a constant, analogous with the work of Sch¨ onbucher and
Wilmott (2000), although a good deal of the analysis close to expiry presented here
is quite widely applicable to other models.
Finally, the question as to whether or not the model described above leads to a
complete market is of interest. In this context the answer to this question reduces to
showing whether or not an option can be perfectly hedged in such a market, hence
that there exists a unique solution to equations (2.9) and (2.10) with the appropriate
CHAPTER 2. THE MODELLING FRAMEWORK 53
payoﬀ proﬁle. In the ﬁrstorder feedback case the linear nature of the equation makes
it easy to show this. For the full feedback problem it is not immediately clear. Recall,
however, that Frey (1998) showed such existence and uniqueness results, although
under some fairly restrictive assumptions; for more information see section 1.7.
2.1 Technical asides
In this section we describe some of the more technical details regarding the application
of standard analytic methods to the problem as outline above.
2.1.1 Markovian processes
In the full feedback case we eﬀectively break the link between the solution of the
PDE and the solution of the SDE. It is well understood that the solution of a linear
parabolic second order PDE can be expressed as an expectation of a Markov pro
cess. Introducing the nonlinearity in the above manner results in a breakdown of the
FeynmanKac representation, and so the solution to the PDE no longer corresponds
to
V (S, t) = E
Q
S,t
_
e
−r(T−t)
(K −S
T
)
+
¸
,
where the expectation is taken under the riskneutral measure, Q, of the modiﬁed
stochastic process (2.5). This is not to say that one cannot use the FeynmanKac
representation theorem (see section 1.3.3) to formulate the solution to a nonlinear
PDE as an expectation. It is possible to do so if the nonlinear PDE can be linearised
by an appropriate transformation. For example, the nonlinear equation
3
u
t
+au
xx
+bu
2
x
= 0,
u(x, T) = h(x),
(2.11)
3
Note that this equation can be obtained from an appropriate transformation of Burgers’ equa
tion, see for example Rosenerans (1972).
CHAPTER 2. THE MODELLING FRAMEWORK 54
can be linearised using a ColeHopf transformation w(x, t) = exp
_
b
a
u(x, t)
_
. The
resulting linear system
w
t
+aw
xx
= 0,
w(x, T) = exp
_
b
a
h(x)
_
,
can thus be written as an expectation using the standard FeynmanKac representation
theorem for Markovian processes, i.e.
w(x, t) = E
P
x,t
_
exp
_
b
a
h(x
T
)
__
,
where x
t
follows the dynamics dx
t
=
√
2adW
P
t
. The linearising transformation can
then be applied to the above expectation to recover the original (nonlinear) value
function u(x, t), and so a link is reestablished between the nonlinear PDE and a
Markovian stochastic process. This is a speciﬁc example of the fact that nonlinear
equations can be expressed as a functional of a Markov process. In the above example
the functional is given by the ColeHopf transform. A functional is any function on
the sample path of a process, for example the integral process or the maximum
process, compared to a function which is just dependent on the value of the process
at the current time (i.e. Markovian).
Note, however, that equation (2.11) is a quasilinear PDE as opposed to equation
(2.10) which is fully nonlinear. This full nonlinearity makes it extremely unlikely
that such a linearising transform can be found and indeed the author could ﬁnd
no such transform. For more on the ColeHopf transform solution to the nonlinear
Burgers’ equation see Rosenerans (1972).
2.1.2 Applicability of Itˆ o’s formula
In the application of Itˆ o’s formula it is assumed that the function f(S, t) is C
2,1
diﬀerentiable, if it is not then we have to include local time contributions. For a
detailed deﬁnition of local time see section IV.5 of Protter (1990) or Peskir (2003).
Consider for example a function that is C
2,1
in spacetime in two separate regions
CHAPTER 2. THE MODELLING FRAMEWORK 55
separated by a kink along a curve in spacetime. If we are to apply Itˆ o’s formula
across the curve, then we have to take into account the local time spent on that
curve (see Peskir, 2005a).
To illustrate this we brieﬂy describe (a slight modiﬁcation of) the socalled Itˆ oTanaka
formula, where we wish to apply Itˆ o’s formula to the function f(S, t) = [S−K[ where
K can be thought of as the strike price.
4
The ﬁrst point to note is that this function
does not have a welldeﬁned second derivative at S = K and so df is illdeﬁned at this
point. We can, however, overcome this by approximating the nonsmooth function
by a smooth function such as
f(S, t) =
_
¸
¸
¸
_
¸
¸
¸
_
K −S, S ∈ (0, K −);
(S−K)
2
2
+

2
, S ∈ (K −, K +);
S −K, S ∈ (K +, ∞),
(2.12)
where > 0 is a small parameter. Note that here we have a welldeﬁned second
derivative in the entire domain. Evaluating df of (2.12) gives
df(S, t) =
_
¸
¸
¸
_
¸
¸
¸
_
−dS, S ∈ (0, K −);
(S−K)

dS +
1
2
(dS)
2
, S ∈ (K −, K +);
dS, S ∈ (K +, ∞).
Taking the limit ↓ 0 leads to
5
df = sgn(S −K)dS + lim
↓0
_
1
2
¦(dS)
2
[S ∈ (K −, K +)¦
_
. (2.13)
Next we can calculate
(dS)
2
=
σ
2
S
2
dt
_
1 −λ sgn(S −K)
_
2
,
and thus we obtain
df = sgn(S −K)dS +σ
2
K
2
lim
↓0
_
1
2
¦dt[S
t
∈ (K −, K +)¦
_
. (2.14)
4
This would correspond to a trading strategy in which we would always hold stock and the size
of our holding would be equal to the distance the current stock price was from the strike price K.
5
The sgn function is deﬁned as
sgn(x) =
−1, x < 0;
0, x = 0;
1, x > 0.
CHAPTER 2. THE MODELLING FRAMEWORK 56
We can deﬁne the local time spent on the curve S
t
= K as L
t
given by
L
t
= lim
↓0
_
1
2
¦u ∈ [0, t][S
u
∈ (K −, K +)¦
_
,
and hence equation (2.13) will become
df = sgn(S −K)dS +σ
2
K
2
dL
t
.
Note that for the BlackScholes model the value function of a standard put option can
be shown to be C
2,1
diﬀerentiable everywhere in the interior of the solution domain
with nonsmoothness only on the domain’s boundary (due to the payoﬀ proﬁle).
Since Itˆ o’s formula is not being applied across the kink in the payoﬀ proﬁle then Itˆ o’s
formula, without local time, is all that is required. Hence our application of Itˆ o’s
formula is valid for the ﬁrstorder feedback case. However smoothness of the solution
to the full feedback PDE has not been determined a priori and if it transpires that
it is not C
2,1
diﬀerentiable in the interior of the domain then we will have to apply
a local time correction to our application of Itˆ o’s formula in a similar manner to the
above.
2.2 Alternative models
Bordag and Frey (2007) identify three distinct frameworks that attempt to model
illiquid markets: transactioncost models; reactionfunction or equilibrium models;
and reducedform SDE models. The derivation of the governing PDEs provided in
this chapter aims to illustrate the links between these frameworks, in particular the
latter two. A brief overview of each framework is provided below.
2.2.1 Transactioncost models
The main model is this class is due to Cetin et al. (2004). In this framework the
transaction price S
t
is given by the formula
S
t
(α) = e
ρα
ˆ
S
t
CHAPTER 2. THE MODELLING FRAMEWORK 57
where
ˆ
S
t
is some fundamental stock price (usually given by geometric Brownian
motion d
ˆ
S
t
= µ
ˆ
S
t
dt +σ
ˆ
S
t
dW
t
), ρ > 0 is a liquidity parameter and α is the number of
shares being traded. Obviously the trader will pay more than the fundamental price
for the stock when buying (i.e. when α > 0) and receive less when selling (α < 0).
Cetin et al. (2004) show that this transaction cost is proportional to the quadratic
variation of the stock trading strategy.
2.2.2 Reactionfunction (equilibrium) models
The models in this class include Jarrow (1994), Frey and Stremme (1997), Platen and
Schweizer (1998), Frey and Coauthors (19962001), Sircar and Papanicolaou (1998),
Sch¨ onbucher and Wilmott (2000), Lyukov (2004) and Bank and Baum (2004). In such
models there are two types of traders in the market, namely ordinary investors and a
large investor. The overall supply of the stock is normalised to one. The normalised
stock demand of the ordinary investors at time t is modelled as a function D(S
t
, U
t
, t),
where S
t
is the price of the stock. The normalised stock demand of the large investor
is written in the form λΨ
t
, where λ ≥ 0 is a parameter that measures the size of the
trader’s position relative to the total supply of the stock. The equilibrium price S
t
is
then determined by the market clearing condition, (where supply meets demand)
D(S
t
, U
t
, t) +λΨ
t
= 1.
Furthermore assuming that this strategy Ψ
t
is Markovian, i.e. that Ψ
t
= f(S
t
, t) for
a smooth function f, then the market clearing condition becomes
D(S
t
, U
t
, t) +λf(S
t
, t) = 1.
It can be shown that under suitable assumptions on the function D, the above admits
a unique solution and hence it can be inverted to solve for S
t
. Diﬀerent models propose
diﬀerent functional forms for D which lead to slightly diﬀerent prices. We shall
return to this framework after we have considered the somewhat more intuitive ﬁnal
framework. For a very general analysis of the dynamics of selfﬁnancing strategies in
reactionfunction models, see Bank and Baum (2004).
CHAPTER 2. THE MODELLING FRAMEWORK 58
2.2.3 Reducedform SDE models
The models in this class are due to Frey (2000), Frey and Patie (2002), Jandaˇcka and
ˇ
Sevˇecoviˇc (2005) and Liu and Yong (2005). Here investors are assumed to be large
traders in the sense that their actions aﬀect the equilibrium price and the liquidity
adjusted price process in the presence of a large trader is given directly by
dS
t
= µ(S
t
, t)dt +σ(S
t
, t)dW
t
+λ(S
t
, t)dΨ
t
, (2.15)
where as before Ψ
t
is a semimartingale representing the trading strategy. Again we
can make the assumption that the strategy is Markovian, i.e. that Ψ
t
= f(S
t
, t) for
a smooth function f. This framework provides a very intuitive means of obtaining
the stock price process in the presence of a large trader, since the ﬁnal term in the
SDE can be considered as incorporating price impact. Note that the model outlined
in the ﬁrst half of this chapter falls under the reducedform SDE class. In the next
section we aim to illustrate how all three frameworks may be ‘transposed’ into a
reducedform SDE (2.15).
2.3 A uniﬁed framework
In this section we attempt to unify the equilibrium and reducedform SDE models.
We show that any of the equilibrium models existing in the literature can be recast in
terms of a reducedform SDE and so (at least for the purposes of analysis) the reduced
form SDE models can be seen as the more general framework, whose properties we
aim to analyse in this thesis. We shall consider, in turn, each of the equilibrium
models and show that they can be rewritten as an SDE of the form (2.15) and hence
all the modelling can be encapsulated into the function λ(S, t). In addition, the
transaction cost model described in section 2.2.1 can also be rewritten as a reduced
form SDE and we shall discuss this next.
CHAPTER 2. THE MODELLING FRAMEWORK 59
2.3.1 Cetin et al. (2004)
Rewriting the model of Cetin et al. (2004) as
S
t
= e
λf(S,t)
ˆ
S
t
we can take diﬀerentials to get
dS
t
= d
_
e
λf
¸
ˆ
S
t
+e
λf
d
ˆ
S
t
= λdfe
λf
ˆ
S
t
+e
λf
_
µ
ˆ
S
t
dt +σ
ˆ
S
t
dW
t
_
+O(dt)
= e
λf
ˆ
S
t
(λdf +µdt +σdW
t
) +O(dt)
= S
t
(λdf +µdt +σdW
t
) +O(dt)
⇒dS
t
= µS
t
dt +σS
t
dW
t
+λS
t
df(S
t
, t) +O(dt)
where we have not included any quadratic variation terms in the calculation as this
would aﬀect only the drift of the process and for the purposes of option pricing we are
only really interested in the volatility and price impact terms. The function f(S, t)
in the above model is identiﬁed as the number of shares traded rather than held and
so it is not obvious that we should thus identify f(S, t) to be the trading strategy
as before. However, in the interest of brevity we will not investigate this model any
further.
2.3.2 Platen and Schweizer (1998)
The model of Platen and Schweizer (1998) deﬁnes the market clearing condition
D(S
t
, U
t
, t) = constant
where the demand is modelled as
D(S
t
, U
t
, t) = U
t
+γ (log S
t
−log S
0
) +f(S
t
, t),
where S
0
and γ constant. Hence
6
dD(S
t
, U
t
, t) = dU
t
+γ
dS
t
S
t
+df(S
t
, t) +O(dt) = 0,
6
Again, note that we are neglecting the quadratic variation terms since these would only inﬂuence
the drift and clutter the algebra.
CHAPTER 2. THE MODELLING FRAMEWORK 60
where S
t
denotes the equilibrium price. Also the authors assume that dU
t
= mdt +
νdW
t
for m, ν ∈ R and so we have that
dS
t
S
t
= −γ
−1
_
mdt +νdW
t
+df(S
t
, t)
_
+O(dt),
hence we can identify
σ(S
t
, t) = −νγ
−1
,
λ(S
t
, t) = −γ
−1
.
There is discussion about the sign of the parameter γ in the paper of Platen and
Schweizer (1998), but in this formulation it is clear that it must be negative in order
to produce a price process that has a positive drift and volatility.
2.3.3 Mancino and Ogawa (2003)
The model of Mancino and Ogawa (2003) extends the work of Platen and Schweizer
(1998) and so they also deﬁne the market clearing condition as
D(S
t
, U
t
, t) = constant,
where the demand is modelled as
D(S
t
, U
t
, t) = U
t
+γ (log S
t
−log S
0
) +H
_
f(S
t
, t)
_
,
for any function H. Again dU
t
= mdt +νdW
t
so we have that
dS
t
S
t
= −γ
−1
_
mdt +νdW
t
+
dH
df
df(S
t
, t)
_
+O(dt),
hence we can identify
σ(S
t
, t) = −νγ
−1
,
λ(S
t
, t) = −γ
−1
dH
df
.
CHAPTER 2. THE MODELLING FRAMEWORK 61
2.3.4 Lyukov (2004)
The model of Lyukov (2004) deﬁnes the market clearing condition to be
dU
t
+df(S
t
, t) = dM
t
,
where the right hand side is the supply from the market maker. He also deﬁnes
liquidity L to be
L =
dM
t
/N
dS
t
/S
t
and
dU
t
= Nmdt +NνdW
t
,
where N is the total number of shares and m, ν ∈ R. Putting these together we have
dS
t
S
t
=
m
L
dt +
ν
L
dW
t
+
1
LN
df(S
t
, t),
hence we can identify
µ(S
t
, t) = mL
−1
,
σ(S
t
, t) = νL
−1
,
λ(S
t
, t) = (LN)
−1
.
2.3.5 Sircar and Papanicolaou (1998)
The model of Sircar and Papanicolaou (1998) deﬁne the (normalised) market clearing
condition to be (cf. section 2.2.2)
D(S
t
, U
t
, t) +ρf(S
t
, t) = 1, (2.16)
where f(S
t
, t) is the normalised aggregate demand per security traded and ρ is the
ratio of options being hedging to total supply. The authors also note that if we
restrict the demand function to the form
D(S
t
, U
t
, t) = D(S
t
, U
t
),
CHAPTER 2. THE MODELLING FRAMEWORK 62
i.e. there is no explicit time dependence, then in order for the model to reduce to the
BlackScholes model in the limit of ρ →0 the demand must take the form
D(S
t
, U
t
) = φ(U
γ
t
/S
t
),
where
dU
t
U
t
= µ
1
dt +σ
1
dW
t
, (2.17)
and γ is the ratio
γ =
σ
0
σ
1
,
where σ
0
is the volatility of the BlackScholes process given by
dS
BS
t
S
BS
t
= µ
0
dt +σ
0
dW
t
,
hence γ is the ratio of the volatility of the reference process U
t
to the volatility of the
BlackScholes process S
BS
t
. Further the authors assume that
φ(z) = βz,
i.e. linear, hence
D(S
t
, U
t
, t) =
βU
γ
t
S
t
. (2.18)
Therefore the market clearing condition becomes
7
dD(S
t
, U
t
, t) +ρdf = 0,
⇒d
_
βU
γ
t
S
t
_
+ρdf = 0,
⇒
βU
γ
t
S
t
_
γ
dU
t
U
t
−
dS
t
S
t
_
+ρdf +O(dt) = 0,
⇒
dS
t
S
t
= γ
dU
t
U
t
+
ρS
t
βU
γ
t
df +O(dt).
Now it is clear from (2.16) and (2.18) that
βU
γ
t
S
t
= D(S
t
, U
t
, t) = 1 −ρf,
and so we have
dS
t
S
t
= γ
dU
t
U
t
+
ρdf
1 −ρf
+O(dt).
7
Ignored the quadratic variation terms.
CHAPTER 2. THE MODELLING FRAMEWORK 63
Finally substitution for U
t
from (2.17) gives
dS
t
S
t
=
σ
0
µ
1
σ
1
dt +σ
0
dW
t
+
ρ
1 −ρf
df +O(dt),
hence we can identify
σ(S
t
, t) = σ
0
,
λ(S
t
, t) =
ρ
1 −ρf
.
With this survey complete we now proceed in the next chapter to investigate the case
of ﬁrstorder feedback (with λ constant for simplicity) and show how the illiquidity,
even in this case, has a signiﬁcant eﬀect on the option replication price, especially as
we approach expiry.
Chapter 3
Firstorder Feedback Model
As a starting point to investigating how liquidity can aﬀect the option value, we
assume that a hedger holds the number of stocks dictated by the analytical Black
Scholes delta, rather than the delta from the modiﬁed option price. This leads to the
linear PDE (assuming λ constant)
∂V
∂t
+
1
2
σ
2
S
2
_
1 −λ
∂
2
V
BS
∂S
2
_
2
∂
2
V
∂S
2
+rS
∂V
∂S
−rV = 0. (2.9)
which is somewhat easier to solve than the fullfeedback problem PDE (2.10), but
still has important and interesting diﬀerences from the classical BlackScholes PDE.
This chapter investigates the analytical properties of equation (2.9), in particularly
in the region close to expiry, highlighting the diﬀerences with the classical Black
Scholes model. We also consider American options in this framework which has not
previously been attempted.
This idea of ﬁrstorder feedback leading to a modiﬁed, but still linear PDE also
appears in Sch¨ onbucher and Wilmott (2000), but under a diﬀerent guise. They call
the solution to the PDE (2.9) the price taker’s price. In an illiquid market inﬂuenced
by a large trader (or by an equivalent large group of small traders) following the
BlackScholes hedging strategy, a small trader can trade any number of shares, on a
small scale, without aﬀecting the price. Hence equation (2.9) models the replicating
64
CHAPTER 3. FIRSTORDER FEEDBACK MODEL 65
cost of an option for such small traders, who are aware of the large traders inﬂuence
on the market; for these traders only, the market appears liquid.
Figure 3.1 shows numerical results from the solution of equation (2.9) (obtained using
a CrankNicolson procedure
1
) for European call options (all with time to maturity,
T = 1 year, riskfree rate, r = 0.04, volatility, σ = 0.2, and exercise price, K = 1)
for λ = 0, 1, 2, 5, 10. Here, the standard call expiry payoﬀ at t = T has been
implemented, i.e. condition (1.1) on page 21. The result with λ = 0 is, of course,
the classic BlackScholes result. As λ is increased, the option value is apparently
eroded monotonically towards the amount by which the contract is currently in the
money or, if out of the money, zero (some of the analysis below will conﬁrm this).
Corresponding results for put options (using the same parameters as for ﬁgure 3.1),
with the standard put payoﬀ condition (1.2) are presented in ﬁgure 3.2, and these
too strongly point to a monotonic asymptote on to the payoﬀ function (for ﬁxed T)
as the liquidity parameter λ increases. Although the illiquid results appear to be
rather qualitatively similar to the liquid (λ = 0) results, a more detailed analysis
(applicable for times close to expiry) follows, and this highlights some subtle, but
important diﬀerences.
It should be noted at this stage that the size of errors and related implementation
details will be omitted from the presentation of numerical results. The techniques
used are more often than not standard and have well understood error estimates. In
all calculations the error levels were more than acceptable, often with errors within
the line width of the graphs presented.
1
Note that ﬁnite diﬀerence methods will be the preferred method of solution throughout this
thesis. Finite diﬀerence methods have the advantage over the alternative ﬁnite element methods
because we almost exclusively work in rectangular domains (asset price and time); in such situations
ﬁnite diﬀerence methods are much easier to implement. For more on ﬁnite diﬀerence methods see
Smith (1978).
CHAPTER 3. FIRSTORDER FEEDBACK MODEL 66
0
0.1
0.2
0.3
0.4
0.5
0.6
0 0.2 0.4 0.6 0.8 1 1.2 1.4
PSfrag replacements
S
C
a
l
l
v
a
l
u
e
λ = 0
λ = 10
Figure 3.1: Value of European call options with ﬁrstorder feedback (T = 1, r = 0.04,
σ = 0.2, K = 1) for λ = 0, 1, 2, 5, 10; the variation with λ appears to be monotonic.
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0 0.2 0.4 0.6 0.8 1 1.2 1.4
PSfrag replacements
S
P
u
t
v
a
l
u
e
λ = 0
λ = 10
Figure 3.2: Value of European put options with ﬁrstorder feedback (T = 1, r = 0.04,
σ = 0.2, K = 1) for λ = 0, 1, 2, 5, 10; the variation with λ appears to be monotonic.
CHAPTER 3. FIRSTORDER FEEDBACK MODEL 67
3.1 Analysis close to expiry: European options
In this section we consider the behaviour of the option value close to expiry. This is
generally the most critical and intricate period for option pricing models and oﬀers
us some insight into the valuation dynamics, shedding more light on the value of
options as the parameter λ is increased. The standard substitution, setting τ = T −t
(representing time to expiry), transforms (2.9) into
V
τ
−
σ
2
S
2
V
SS
2 (1 −λV
BS
SS
)
2
−rSV
S
+rV = 0, (3.1)
where subscripts now denote derivatives and V
BS
is the solution to the corresponding
BlackScholes equation
V
BS
τ
−
1
2
σ
2
S
2
V
BS
SS
−rSV
BS
S
+rV
BS
= 0. (3.2)
We next investigate the small τ behaviour of (3.1) (i.e. for times close to expiry), for
which we also need to know the behaviour of the BlackScholes equation (3.2) in this
limit. To obtain this, we wish to zoom in on the solution domain close to strike and
expiry. This zoomed domain we refer to as the inner region and the solution in this
domain, the inner solution. Naturally the domain and solution outside of this region
are given the preﬁx outer. Mathematically this zooming is done via the following well
known transformation, hence as τ →0 the solution takes the form
V
BS
= τ
1
2
f(η) +O(τ), (3.3)
where
η =
S −K
τ
1
2
, (3.4)
which is often called the inner variable and whose form can be veriﬁed a posteriori;
such a solution is sometimes called a selfsimilar solution.
2
It can be shown (see for
example Wilmott et al., 1995) that the solution f when η = O(1), is given by the
solution to the ordinary diﬀerential equation (ODE)
σ
2
K
2
f
ηη
+ηf
η
−f = 0. (3.5)
2
For a survey of these methods applied to the equations arising in continuum mechanics see
Barenblatt (1996).
CHAPTER 3. FIRSTORDER FEEDBACK MODEL 68
For a put the appropriate boundary conditions become
f →0 as η →∞, f →−η as η →−∞
and for a call
f →η as η →∞, f →0 as η →−∞.
Using these boundary conditions it is straightforward and well known to show (cf.
Wilmott et al., 1995) that equation (3.5) leads to the solution for a put
f
P
(η) = η
_
Φ
_
η
Λ
_
−1
_
+
Λ
√
2π
e
−
1
2
(
η
Λ
)
2
(3.6)
and likewise for a call
f
C
(η) = ηΦ
_
η
Λ
_
+
Λ
√
2π
e
−
1
2
(
η
Λ
)
2
,
where Λ = σK and Φ() is the standard normal cumulative distribution function
deﬁned as
Φ(x) =
1
√
2π
_
x
−∞
e
−
1
2
y
2
dy. (3.7)
Considering the second derivative gives
f
P
ηη
= f
C
ηη
=
1
Λ
√
2π
e
−
1
2
(
η
Λ
)
2
,
and, consequently, rewriting in terms of the original variables, the local BlackScholes
gamma is given by
V
BS
SS
=
1
σK
√
2πτ
e
−
(S−K)
2
2τσ
2
K
2
. (3.8)
We can now proceed to incorporate this into the ﬁrstorder feedback illiquid problem
(3.1). To investigate the small τ behaviour of this equation, we can perform a local
similarity analysis similar to that just performed for the liquid BlackScholes equation.
To illustrate this powerful technique the analysis will be explained in detail. First we
seek a solution of the form
V = τ
α
g(ξ), (3.9)
where
ξ =
S −K
τ
β
, (3.10)
CHAPTER 3. FIRSTORDER FEEDBACK MODEL 69
and is assumed to be O(1). α and β are constants to be determined. The ﬁrst
point to note is that if the BlackScholes local solution described above is used, i.e.
α = β =
1
2
, then the second derivative term in (3.1) becomes much smaller than the
other terms in the equation in the limit τ →0. Therefore it would appear that close
to expiry the dynamics of the ﬁrstorder feedback model are signiﬁcantly diﬀerent
from the standard BlackScholes model. Substituting (3.9) and (3.10) into (3.1) and
rewriting the approximation of V
BS
SS
close to expiry (3.8) in terms of the new inner
variable ξ gives
τ
α−1
(αg −βξg
ξ
) −
σ
2
K
2
τ
α−2β
g
ξξ
2
_
1 −
λ
σK
√
2πτ
e
−
ξ
2
τ
2β−1
2σ
2
K
2
_
2
+rKτ
α−β
g
ξ
+rτ
α
g = 0. (3.11)
It is clear that the denominator of the second term is O(τ
−1
) in the limit τ → 0
(provided that 2β −1 > 0, which can veriﬁed a posteriori ). Consequently the second
derivative term is O(τ
α−2β+1
) as τ →0. To obtain an appropriate scaling we balance
this with the time derivative term which is O(τ
α−1
), leading to the conclusion that
α −1 = α −2β + 1 ⇒β = 1.
To ﬁx α we exploit the fact that the inner solution τ
α
g(ξ) must match with the outer
solution, the solution that is valid outside of our inner region close to strike and
expiry. In this (outer) region the second derivative of the solution is eﬀectively zero
since the payoﬀ proﬁle has no curvature here. Hence the outer solution is given by
the solution to equation (2.9) without the diﬀusion term, i.e.
V
τ
−rSV
S
+rV = 0. (3.12)
This can be solved using the method of characteristics, in conjunction with the ap
propriate initial condition, to obtain
V
P
(S, τ) =
_
Ke
−rτ
−S
_
+
, (3.13a)
V
C
(S, τ) =
_
S −Ke
−rτ
_
+
, (3.13b)
for puts and calls respectively. Thus we require τ
α
g(ξ) = S −Ke
−rτ
for a call option
in the limit ξ →∞. We shall discuss this matching procedure in more detail shortly,
CHAPTER 3. FIRSTORDER FEEDBACK MODEL 70
however the preceding arguments suggest that for the matching to work we require
S − K = O(τ
α
). In addition, by construction ξ = O(1) in the inner region, hence
(3.10) suggests we must also have S −K = O(τ
β
). We conclude that
3
α = β = 1.
PSfrag replacements
S
τ
K
ξ →∞
ξ →−∞
V
P
(S, τ) = 0
V
C
(S, τ) = S −Ke
−rτ
V
P
(S, τ) = Ke
−rτ
−S
V
C
(S, τ) = 0
V
P
(S, τ) = τg
P
(ξ)
V
C
(S, τ) = τg
C
(ξ)
O(τ)
Figure 3.3: Asymptotic Matching.
Note that the scaling here implies a region O(τ) in asset space S, close to the strike
price, that is somewhat smaller than the classical BlackScholes model, (3.3), which is
O(τ
1
2
) as τ →0 (see (3.4)); this is clearly an important diﬀerence from the standard
BlackScholes model behaviour close to expiry. Substituting for α and β into (3.11)
gives
g −ξg
ξ
−
σ
2
K
2
τ
−1
g
ξξ
2
_
1 −
λ
σK
√
2πτ
e
−
ξ
2
τ
2σ
2
K
2
_
2
−rKg
ξ
+rτg = 0,
and taking the O(1) terms leads to
πσ
4
K
4
λ
2
g
ξξ
+ (ξ +rK) g
ξ
−g = 0. (3.14)
Note that on the relatively short ξ = O(1) scale, we are eﬀectively replacing (3.8)
with
V
BS
SS
=
1
σK
√
2πτ
. (3.15)
3
Note that this satisﬁes our a priori assumption that 2β −1 > 0.
CHAPTER 3. FIRSTORDER FEEDBACK MODEL 71
The appropriate boundary conditions to equation (3.14) for standard puts and calls
are obtained from an asymptotic matching procedure (cf. van Dyke, 1964). For
asymptotic matching we require the inner solution, τg(ξ) in the limit [ξ[ → ∞ to
match with the outer solution (3.13). A schematic of this is given in ﬁgure 3.3, and
formally the matching condition is deﬁned as
lim
S→K,τ→0
V
OUTER
(S, τ) = lim
ξ→∞
τg(ξ). (3.16)
Practically we rewrite the outer solution in terms of the inner variable ξ and then
equate the result to the inner solution τg(ξ) to give the appropriate boundary con
ditions. This procedure leads to the boundary conditions for a put:
g →0 as ξ →∞, g →−(ξ +rK) as ξ →−∞,
and for a call:
g →ξ +rK as ξ →∞, g →0 as ξ →−∞.
Equation (3.14) can be solved analytically; for a put, the solution is
g
P
(ξ) = (ξ +rK)
_
Φ
_
ξ +rK
κ
_
−1
_
+
κ
√
2π
e
−
1
2
(
ξ+rK
κ
)
2
, (3.17)
where κ =
√
πσ
2
K
2
λ
, and for a call
g
C
(ξ) = (ξ +rK)Φ
_
ξ +rK
κ
_
+
κ
√
2π
e
−
1
2
(
ξ+rK
κ
)
2
, (3.18)
where Φ() is the cumulative normal distribution function deﬁned by (3.7). Note
that increasing illiquidity (λ → ∞) implies κ → 0 and this in turn indicates that
(3.17) and (3.18) become increasingly focused about ξ = −rK, i.e. S = K(1 − rτ),
taking on the payoﬀ form away from this point, consistent with our observations
above regarding ﬁgures 3.1 and 3.2. Furthermore diﬀerentiating (3.17) and (3.18)
with respect to λ directly gives
∂g
P
∂λ
=
∂g
C
∂λ
= −
σ
2
K
2
√
2λ
2
e
−
1
2
(
ξ+rK
κ
)
2
< 0,
conﬁrming that, on the inner scale at least, the option value is monotonic decreasing
in the liquidity parameter λ.
CHAPTER 3. FIRSTORDER FEEDBACK MODEL 72
Figure 3.4 shows the diﬀerence between the inner solution and the payoﬀ function for
both put and call options for various values of the liquidity parameter λ. In addition
to conﬁrming the monotonic behaviour in λ, it can be seen that call values always lie
above the payoﬀ curve (i.e. g
C
−[ξ]
+
> 0 for all ξ). As with the classical BlackScholes
result for European calls, in the present case it is never optimal to early exercise calls.
In the case of the puts, it is possible for g
P
− [−ξ]
+
< 0 (i.e. V − payoﬀ < 0) for
certain ranges of ξ, which opens up the potential for the optimal early exercise (on
the ξ scale), and so a consideration of this possibility is considered next.
0.05
0
0.05
0.1
0.15
0.2
0.25
0.3
2 1.5 1 0.5 0 0.5 1 1.5 2
P
S
f
r
a
g
r
e
p
l
a
c
e
m
e
n
t
s
g
P
(
ξ
)
−
[
−
ξ
]
+
g
C
(
ξ
)
−
[
ξ
]
+
ξ
λ ↓ 0
(a) Put
0.05
0
0.05
0.1
0.15
0.2
0.25
0.3
2 1.5 1 0.5 0 0.5 1 1.5 2
P
S
f
r
a
g
r
e
p
l
a
c
e
m
e
n
t
s
g
P
(
ξ
)
−
[
−
ξ
]
+
g
C
(
ξ
)
−
[
ξ
]
+
ξ
λ ↓ 0
(b) Call
Figure 3.4: Inner solution minus the payoﬀ for put and call options, r = 0.04, σ = 0.2,
K = 1 and for λ = 0.1, 0.15, 0.2, . . ., 0.4.
3.2 Analysis close to expiry: American put op
tions
The remarks above naturally beg the question as to the value of a put option on a
ﬁnitely liquid underlying if early exercise is permitted. In the context of ﬁrstorder
feedback, the most consistent model has the delta in (2.8) on page 51 computed
using the liquid (λ = 0) American put value V
BS
AM
, which does permit early exercise;
note that the free boundary (optimal exercise price) of the illiquid put option, V ,
need not necessarily be the same as the free boundary of the liquid option V
BS
AM
.
Figure 3.5 shows results for the American put with the same ﬁnancial parameters
as for the earlier European options, obtained via a standard Projected Successive
CHAPTER 3. FIRSTORDER FEEDBACK MODEL 73
Over Relaxation (PSOR) iterative procedure. The convergence criteria chosen for
the algorithm was that maximum error between subsequent iterations was less than
1 10
−6
.
At each node the BlackScholes American value was computed using a PSOR algo
rithm and then this value was used in the PSOR algorithm for the ﬁrstorder feedback
PDE (2.9) subject to (1.2). Again we see the ‘collapse’ of the option value on to the
payoﬀ as the liquidity parameter λ is increased (which implies the location of the
free boundary always moves towards the exercise price as λ increases). However,
although the results appear to be qualitatively similar to the λ = 0 case, there are
subtle diﬀerences, as we shall now show.
0
0.05
0.1
0.15
0.2
0.25
0.8 0.9 1 1.1 1.2 1.3 1.4 1.5
PSfrag replacements
S
P
u
t
v
a
l
u
e
λ = 0
λ = 10
Figure 3.5: Value of American put options, T = 1, r = 0.04, σ = 0.2, K = 1 and for
λ = 0, 1, 2, 5, 10; the variation with λ appears to be monotonic.
Analysis of the liquid (λ = 0) American put option close to expiry leads to a somewhat
complicated structure, as detailed by Kuske and Keller (1998). Here, the η scale
deﬁned in (3.4) can be shown to fail to capture the free (exercise) boundary. Instead,
the free boundary is located at a somewhat larger distance (O(
√
−τ log τ)) from
the exercise price (with a signiﬁcant price variation in a region O(
_
−τ/ log τ) of this
exercise boundary). It was shown by Widdicks (2002) that as τ →0, on the η = O(1)
scale, the solution of the liquid (λ = 0) American option takes the same form as that
CHAPTER 3. FIRSTORDER FEEDBACK MODEL 74
of its European counterpart, i.e. (3.6). Therefore, it is entirely self consistent to
use this form and, indeed, the European gamma (3.8), for the American case when
η = O(1) or smaller, which is relatively distant from the free boundary.
However, recall that for the case when λ ,= 0, with ξ = O(1) we have clear indications
of the possibility of early exercise (on the ξ scale) for the illiquid put. In other words
the ξ scaling for the ﬁrstorder feedback equation encompasses the free boundary
(unlike the η scaling for the liquid (λ = 0) case). Therefore, the American problem
in this case reduces to the solution of (3.14), subject to the conditions
4
g →0 as ξ →∞, (3.19a)
g = −ξ and g
ξ
= −1 on ξ = ξ
f
, (3.19b)
where we have used the usual smooth pasting conditions (continuity of the option
value and its derivative) and ξ
f
denotes the location of the free boundary (on the
ξ scale). It is straightforward to solve the system (3.14), (3.19) fully numerically,
however it is also possible to reduce the above problem to a transcendental equation
for ξ
f
, a procedure which has not previously been done in the literature and shall be
outlined below.
It is convenient to make the shift z = ξ + rK which transforms the system (3.14),
(3.19) to
κ
2
g
zz
+zg
z
−g = 0, (3.20a)
g →0 as z →∞, (3.20b)
g = −z +rK, g
z
= −1 on z = z
f
, (3.20c)
where κ is as before and z
f
the free boundary on the zscale. To proceed we seek a
solution of the form,
g(z) = zˆ g(z)
4
Note that here the outer solution is given by V
P
(S, τ) = (K − S)
+
rather than V
P
(S, τ) =
(Ke
−rτ
− S)
+
since we are dealing with the American option and so the outer solution will be
trivially the payoﬀ proﬁle.
CHAPTER 3. FIRSTORDER FEEDBACK MODEL 75
which upon substitution leads to
ˆ g
zz
ˆ g
z
= −
2
z
−
z
κ
2
.
Integrating once gives
ˆ g
z
=
A
z
2
e
−
1
2
(
z
κ
)
2
,
where A is a constant of integration. Integrating once more gives
[ˆ g]
∞
z
=
_
∞
z
A
y
2
e
−
1
2
(
y
κ
)
2
dy,
⇒ ˆ g(∞) − ˆ g(z) =
_
−
A
y
e
−
1
2
(
y
κ
)
2
_
∞
z
−
A
κ
2
_
∞
z
e
−
1
2
(
y
κ
)
2
dy,
⇒ ˆ g(z) = −
A
z
e
−
1
2
(
z
κ
)
2
+
A
κ
2
_
∞
z
e
−
1
2
(
y
κ
)
2
dy,
where we have applied the boundary condition (3.20b). Returning to the original
function g(z) gives
g(z) = zˆ g(z) = −Ae
−
1
2
(
z
κ
)
2
+
Az
κ
2
_
∞
z
e
−
1
2
(
y
κ
)
2
dy. (3.21)
It is now required to determine the value of A using the remaining boundary condi
tions on the free boundary. To do this we diﬀerentiate equation (3.21) to obtain
g
z
(z) =
A
κ
2
_
∞
z
e
−
1
2
(
y
κ
)
2
dy.
We can now apply the two boundary conditions at z
f
, i.e.
g(z
f
) = −z
f
+rK = −Ae
−
1
2
(
z
f
κ
)
2
+
Az
f
κ
2
_
∞
z
f
e
−
1
2
(
y
κ
)
2
dy, (3.22)
g
z
(z
f
) = −1 =
A
κ
2
_
∞
z
f
e
−
1
2
(
y
κ
)
2
dy. (3.23)
From (3.23) we have
A =
−κ
2
_
∞
z
f
e
−
1
2
(
y
κ
)
2
dy
,
hence substituting into (3.22) gives
−z
f
+rK =
κ
2
e
−
1
2
(
z
f
κ
)
2
_
∞
z
f
e
−
1
2
(
y
κ
)
2
dy
−z
f
,
⇒rK
_
∞
z
f
e
−
1
2
(
y
κ
)
2
dy = κ
2
e
−
1
2
(
z
f
κ
)
2
,
CHAPTER 3. FIRSTORDER FEEDBACK MODEL 76
which can be written in the form of the standard cumulative normal distribution
function as
Φ
_
z
f
κ
_
= 1 −
κ
rK
√
2π
e
−
1
2
(
z
f
κ
)
2
.
Transforming back to the original ξ variable yields
Φ
_
ξ
f
+rK
κ
_
= 1 −
κ
rK
√
2π
e
−
1
2
ξ
f
+rK
κ
2
,
or further
Φ
_
λ(ξ
f
+rK)
√
πσ
2
K
2
_
= 1 −
σ
2
K
√
2rλ
e
−
λ
2
2π
ξ
f
+rK
σ
2
K
2
2
. (3.24)
6
4
2
0
2
4
6
8
0 1 2 3 4 5
PSfrag replacements
l
o
g
(
−
ξ
f
)
λ
Figure 3.6: Firstorder feedback put (with early exercise), location of free boundary
(as τ →0) with λ, K = 1, r = 0.04, σ = 0.2.
Figure 3.6 shows the variation of the local free boundary ξ
f
(more particularly
log(−ξ
f
)) with λ for the ﬁnancial parameters considered earlier, i.e. r = 0.04, σ = 0.2,
K = 1. The key point to note is that solutions of the system do exist, i.e. the short
S −K = O(τ), ξ = O(1) scale captures the location of the free boundary with ﬁrst
order feedback, whilst as noted above, the liquid (λ = 0) case evolves on a relatively
longer scale of S − K = O(
√
−τ log τ); consistent with this as λ → 0, ξ
f
→ −∞.
Further asymptotic analysis can describe this behaviour, but is omitted in the in
terests of brevity. Note that transforming back to the original (S, τ) variables using
equation (3.10) we can see that
S
f
(τ) = K +ξ
f
τ +. . .
CHAPTER 3. FIRSTORDER FEEDBACK MODEL 77
where ξ
f
is determined by equation (3.24). Hence the free boundary approaches the
strike price at expiry linearly for small times to expiry.
3.3 The vanishing of the denominator
One further interesting property of the ﬁrstorder feedback model is that we have
the possibility of the denominator of the volatility term vanishing, hence a possible
breakdown in the diﬀusion term of the ﬁrstorder feedback PDE.
5
Considering equa
tion (2.9) one might naively think that a singularity occurs when 1 − λV
BS
SS
= 0.
Using the analytic solution for the BlackScholes equation we ﬁnd that
V
BS
SS
=
e
−
1
2
d
1
(S,τ)
2
σS
√
2πτ
for both puts and calls where
d
1
(S, τ) =
log
_
S
K
_
+
_
r +
1
2
σ
2
_
τ
σ
√
τ
.
Hence the location where the denominator of equation (2.9) vanishes is given by the
solution to the equation
1 −
λe
−
1
2
d
1
(S
∗
,τ)
2
σS
∗
√
2πτ
= 0. (3.25)
This equation can be solved explicitly by setting x = ln(S/K) and doing so we arrive
at
S
∗
(τ) = K exp
_
−
_
r +
3
2
σ
2
_
τ ±
√
2σ
2
τ
_
(r +σ
2
)τ −log
_
λ
σK
√
2πτ
__1
2
_
.
Figure 3.7 shows the results for the same set of ﬁnancial parameters as used through
out this chapter.
It can be seen that equation (3.25) has two distinct solutions for τ ∈ (0, τ
0
) for some
ﬁnite timetoexpiry τ
0
, determined by the solution of the transcendental equation
(r +σ
2
)τ
0
= log
_
λ
σK
√
2πτ
0
_
.
5
This shall be even more important in our consideration of the full feedback problem discussed
in chapter 4.
CHAPTER 3. FIRSTORDER FEEDBACK MODEL 78
0.97
0.98
0.99
1
1.01
1.02
0 0.005 0.01 0.015 0.02 0.025 0.03 0.035 0.04 0.045
P
S
f
r
a
g
r
e
p
l
a
c
e
m
e
n
t
s
S
∗
τ
τ
0
Figure 3.7: Location of the vanishing of the denominator of (2.9) with λ = 0.1,
K = 1, r = 0.04 and σ = 0.2.
For τ > τ
0
the equation has no solutions, i.e. the denominator does not vanish in this
region. Note that when considering the Frey (2000) model an explicit expression for
τ
0
can be obtained (see section 7.1). Also note that the location of the singularity at
τ = 0 is K as one might expect.
Having determined the location in which the denominator of equation (2.9) vanishes
it is now interest to investigate the behaviour of the solution at these points. It
is a well known property of ODEs
6
that singularities of the solution occur only at
singularities of the equation when it is written in the standard form
V
(n)
= F
_
V
(n−1)
, V
(n−2)
, . . . , V, S
_
,
i.e. for a second order ODE in the form
V
SS
= F (V
S
, V, S) .
Hence possible singularities of the solution occur at singularities of the function F().
Even though equation (2.9) is a PDE, and thus the above result cannot be applied
directly, it is informative to rewrite (2.9) in standard form to give
V
SS
=
2
σ
2
S
2
(V
τ
−rSV
S
+rV )
_
1 −λV
BS
SS
_
2
,
6
See, for example, Kruskal et al. (1997)
CHAPTER 3. FIRSTORDER FEEDBACK MODEL 79
from which we can see that the right hand side has no singularities except at [V
BS
SS
[ →
∞, which only occurs at (S, τ) = (K, 0). Hence it is expected that no singularities
will appear in the solution for τ > 0, instead it is clear that at the locations of the
vanishing of the denominator the second derivative will be forced to zero.
1
0.8
0.6
0.4
0.2
0
0.85 0.9 0.95 1 1.05 1.1 1.15
P
S
f
r
a
g
r
e
p
l
a
c
e
m
e
n
t
s
S
∆
=
∂
V
∂
S
τ = 0.01
τ = 0.05
Figure 3.8: The ﬁrst derivative (∆) of the BlackScholes equation (3.2) (dotted line)
and the ﬁrst order feedback PDE (2.9) (solid line) for τ = 0.01, 0.015, . . ., 0.05.
Compare the location of the vanishing denominator 3.7.
0
5
10
15
20
25
30
35
40
45
0.85 0.9 0.95 1 1.05 1.1 1.15
P
S
f
r
a
g
r
e
p
l
a
c
e
m
e
n
t
s
S
Γ
=
∂
2
V
∂
S
2
τ = 0.01
τ = 0.05
S
∗
Figure 3.9: The second derivative (Γ) of the BlackScholes equation (3.2) (dotted
line) and the ﬁrst order feedback PDE (2.9) (solid line) for τ = 0.01, 0.015, . . ., 0.05.
Compare the location of the vanishing denominator 3.7.
Figures 3.8 and 3.9 show the ﬁrst and (more importantly) the second derivative
CHAPTER 3. FIRSTORDER FEEDBACK MODEL 80
of the solution of (2.9) in the vicinity of S
∗
respectively. It can be seen that the
solution at S
∗
appears to be at least C
2,1
diﬀerentiable and further that V
SS
remains
positive, just ‘skimming’ zero as S
∗
is approached, thus indicating that the solution
remains convex as τ increases; despite the vanishing of the denominator. Indeed
convexity preservation is a wellknown property of option prices under the assumption
of geometric Brownian motion, see for example Bergman et al. (1996) or El Karoui
et al. (1998). Convexity preservation means that given convexity at t the option price
remains convex for all times prior to t. This is true only in certain models and is
not necessarily true for certain models when in higher dimensions. Ekstr¨ om et al.
(2005) show that geometric Brownian motion is the only convexity preserving model
in higher dimensions. More importantly they show that (in one dimension) a local
volatility model, i.e. a model for the underlying with an arbitrary volatility function
σ(S, τ) as we have here, will be convexity preserving. The numerical investigations
of the ﬁrstorder feedback equation described above indicate that the solution proﬁle
does indeed remain convex, i.e V
SS
≥ 0 for all (S, τ), in agreement with the stated
results.
One ﬁnal subtlety of the behaviour of the ﬁrstorder feedback equation comes when we
take a closer look at the hypothesised monotonic decreasing behaviour of the option
value with the liquidity parameter λ. We have shown that for small times to expiry
the solution is indeed monotonically decreasing in the parameter λ, however this in
not the whole story for τ = O(1). Figure 3.10 shows the solution to equation (3.1)
for two slightly diﬀerent values of λ (the dotted line representing the higher value)
and also at three diﬀerent times to maturity. It can be seen that for the closest time
to maturity a monotonic decreasing relationship is shown, however as we increase the
time to expiry, this relationship appears to be increasing for suﬃciently large times
away from maturity.
We can begin to see what may be happening if we take a closer look at the modiﬁed
volatility term of the ﬁrstorder feedback case, i.e.
ˆ σ
2
(S, τ) =
σ
2
S
2
(1 −λV
BS
SS
)
2
. (3.26)
CHAPTER 3. FIRSTORDER FEEDBACK MODEL 81
0
0.02
0.04
0.06
0.08
0.1
0.12
0.14
0.16
0.85 0.9 0.95 1 1.05 1.1 1.15
P
S
f
r
a
g
r
e
p
l
a
c
e
m
e
n
t
s
S
V
τ = 0.0125
τ = 0.0375
τ = 0.075
Figure 3.10: Firstorder feedback put option value for two diﬀerent values of λ at
various times to expiry; τ = 0.0125, 0.0375, 0.075. For r = 0.04, σ = 0.2, K = 1 and
λ = 0.09 (solid line) and λ = 0.1 (dotted line). Compare with ﬁgure 3.7.
Recall that the results of Ekstr¨ om et al. (2005) suggest that the solution to the PDE
(3.1) (for convex payoﬀ proﬁles) remains convex for all (S, τ). As such we have the
relationship that an increase in volatility will result in an increase in option price,
7
hence to determine the option price dependence on λ we need only determine the
dependence of the modiﬁed volatility on λ. Diﬀerentiating (3.26) with respect to λ
yields
∂ˆ σ
2
∂λ
=
σ
2
S
2
V
BS
SS
2 (1 −λV
BS
SS
)
3
(3.27)
and it is clear that the sign of the above may change, dependent on the size of V
BS
SS
.
In fact, (3.27) indicates that the solution dependency on the liquidity parameter λ is
monotonic increasing in regions in which V
BS
SS
< 1/λ and monotonic decreasing for
V
BS
SS
> 1/λ. However in the limit as λ → ∞ the region V
BS
SS
> 1/λ expands to ﬁll
the whole domain and so this will become the dominant behaviour for a ﬁxed time
to expiry. The above agrees well with the results shown in ﬁgure 3.10, since for the
parameters used, ﬁgure 3.7 indicates that the denominator of equation (3.1) does not
vanish for τ 0.04 and so we are in the region in which V
BS
SS
< 1/λ and we should
thus expect monotonic increasing behaviour in λ, which it appears we have.
7
See for example Bergman et al. (1996), El Karoui et al. (1998) or Janson and Tysk (2003)
CHAPTER 3. FIRSTORDER FEEDBACK MODEL 82
With this necessary background complete, in the next chapter we may proceed to
investigate the more challenging fullfeedback model.
Chapter 4
Fullfeedback Model
We now turn our attention to the full feedback case, namely the equation
V
τ
−
σ
2
S
2
V
SS
2 (1 −λ(S, τ)V
SS
)
2
−rSV
S
+rV = 0, (4.1)
where the trading strategy assumed to aﬀect the price is not simply the BlackScholes
delta hedging strategy as discussed in the previous chapter, but rather is based on
the actual delta of the modiﬁed price, and as a consequence the price impact is fully
considered in the trading strategy. This corresponds to a situation where all market
participants performing such hedging strategies are aware of the eﬀect that their
strategies have on the price. In this case the trading strategy has to be determined
as part of the problem, resulting in nonlinearity.
The full feedback equation has been studied extensively in the literature with various
functional forms of the liquidity parameter λ(S, t). Existence and uniqueness was
given by Frey (1998) which shows that options in such a market can be perfectly
replicated, hence the market is complete. These results, however, are not applica
ble to standard put or call options (nor any nonsmooth payoﬀ proﬁle) and so here
we investigate the solutions to the PDE in such regimes. The aim being to illus
trate exactly how things go wrong, with the view to informing modellers on how to
incorporate nonsmooth payoﬀs into the intuitive framework outlined in chapter 2.
83
CHAPTER 4. FULLFEEDBACK MODEL 84
The fullfeedback model described here is what Sch¨ onbucher and Wilmott (2000) call
the paper value replication for the large trader, socalled because the value satisfying
the PDE is just the paper value of the option. Liquidating the portfolio would change
the price, and due to the negative slope of the demand curve the realised value would
be less than the paper value. However, numerous diﬃculties arise when liquidation
strategies are incorporated into such dynamic hedging strategies. It is for this reason
that the majority of models in the literature (and the present study) consider only
the paper value or make the assumption that the option is closed out using physical
delivery to bypass any diﬃculties with the liquidation value.
Nonlinear diﬀusion equations are a frequent occurrence in the physical sciences and
the work done in these disciplines can provide much insight into the nonlinear be
haviour of the models arising in mathematical ﬁnance. As such the aim of this chapter
is investigate thoroughly the properties of the nonlinear PDE (4.1), with the standard
put and call payoﬀ proﬁles (1.2) and (1.1), using various analytical and numerical
techniques. We also give a brief overview of some of the techniques used to solve
general nonlinear PDEs and appraise their appropriateness for equation (4.1).
Note that equation (4.1) has the form
V
τ
−
1
2
F(S, τ, V
SS
)S
2
V
SS
−rSV
S
+rV = 0,
and as such is fully nonlinear. Equations of this form were ﬁrst studied by Barenblatt
and coworkers in the context of hydrodynamics and more recently have arisen in
models occurring in quantitative ﬁnance, for example in many transaction cost models
such as Barles and Soner (1998).
The nonlinearity of equation (4.1) has many important consequences. If we have a
nonlinear PDE then one of the most striking diﬀerences with linear equations is that
the sum of two or more solutions is no longer necessarily a solution itself. As such
there is a signiﬁcant diﬀerence in the value of the option to someone holding a long
position as opposed to a short position, hence we must take care to specify the option
CHAPTER 4. FULLFEEDBACK MODEL 85
position. It is quite remarkable therefore that the inherent nonlinearity of the pricing
equations leads, quite naturally, to the concept of the bidask spread.
The eﬀect of transaction costs and liquidity costs are always a sink of money for
hedgers. As an example of this, let us consider any nonlinear transaction cost model.
A portfolio consisting of the same option held both long and short would be priced
at zero since this portfolio has a zero payoﬀ and is thus worthless. However when
pricing each option separately, transaction costs will be incurred on both options and
such costs will aggregate causing a disparity between the value of the portfolio and
the sum of its constituent parts; the diﬀerence being that of the total transaction
costs incurred.
When faced with a partial diﬀerential equation, there are many questions that one
must ask. The most fundamental being does there exist a solution, and if so is this
solution unique? The latter question becomes even more important when dealing
with nonlinear equations, well known to exhibit possible multiple solutions. For
linear (diﬀusion) equations, existence and uniqueness results are well known, and in
fact the BlackScholes equation can be shown to be a suﬃciently wellposed problem.
Stating some results from general PDE theory, a problem is wellposed if:
1
1. The problem has a solution,
2. This solution is unique,
3. The solution depends continuously on the data given in the problem.
In addition if we require the solution of a PDE of order k to be at least k times
continuously diﬀerentiable, then we call a solution with this much smoothness a
classical solution of the PDE. Many PDEs do not have classical solutions but are
nonetheless wellposed if we allow for properly deﬁned generalised or weak solutions.
As it turns out this idea of a weak solution is exactly what is needed when tackling
the solution to equation (4.1) with nonsmooth payoﬀ proﬁles, see section 4.6.
1
See for example Evans (1998)
CHAPTER 4. FULLFEEDBACK MODEL 86
As previously mentioned, Frey (1998) showed that equation (4.1) was wellposed for
λ(S, τ) =
ˆ
λS and for suﬃciently smooth payoﬀ proﬁles. This was done by diﬀeren
tiating the fully nonlinear PDE in V to obtain a quasilinear PDE in the hedging
strategy ∆(S, τ) for which existence and uniqueness was shown using arguments sim
ilar to Ladyzenskaja et al. (1968).
2
In the sequel we will be considering the more
general scenario of nonsmooth payoﬀ proﬁles, where existence an uniqueness results
have not been established.
4.1 Putcall parity
Firstly we note that putcall parity can be shown to still hold, even in this highly
nonlinear situation. This can be shown by simply substituting the put call parity
relationship
V
P
= V
C
−S +Ke
−rτ
into the nonlinear equation for the put value V
P
(S, τ), i.e.
V
P
τ
−
σ
2
S
2
V
P
SS
2 (1 −λV
P
SS
)
2
−rSV
P
S
+rV
P
= 0,
and the payoﬀ proﬁle
V
P
(S, 0) = (K −S)
+
.
Doing so we obtain
V
C
τ
−
σ
2
S
2
V
C
SS
2 (1 −λV
C
SS
)
2
−rSV
C
S
+rV
C
= 0,
with
V
C
(S, 0) = (S −K)
+
,
hence we can recover the call option value from the put option value. Note that the
nonlinear equation (4.1) diﬀers from the BlackScholes equation by a function of the
second derivative of the option value only, i.e. σ = σ(S, τ, V
SS
), hence it is clear that
the parity relationship will still hold, since the second derivative of the put and call
option coincide.
2
An alternative and novel approach to providing existence and uniqueness results, exploiting the
maximum principle for parabolic equations, can be found in section A.2.2.
CHAPTER 4. FULLFEEDBACK MODEL 87
4.2 A solution by inspection
It should be noted that it is possible to ﬁnd exact analytic solutions to the PDE (4.1).
Such solutions can be obtained by exploiting symmetries of the governing equation
(and boundary conditions). These similarity solutions, based on the theory of Lie
groups, can be useful in investigating nonlinear problems, however the application of
such methods tends to be limited by the fact that many nonlinear PDEs do not have
such symmetries. In addition, the solutions obtained in this way are generally only
valid for very restrictive boundary conditions. As an example of this, if we assume
that λ(S, τ) in equation (4.1) is a function of τ only, i.e. λ = λ(τ), then we can seek
a solution of the following form
V (S, τ) = S
2
h(τ).
Substitution into (4.1) gives
h
τ
−
σ
2
h
(1 −2λh)
2
−rh = 0,
which is now an ODE and so open to standard solution techniques. Making the
further assumption that λ is constant we can arrive at the implicit solution for h:
ln(h) +
σ
2
2r
ln
_
(1 −2λh)
2
+
σ
2
r
_
+
σ
√
r
arctan
_√
r
σ
(1 −2λh)
_
= (r +σ
2
)τ +A,
where A is a constant determined by the payoﬀ proﬁle. It should be emphasised here
that we have been restricted to a constant λ(S, t) in order to obtain this solution
(without any real ﬁnancial justiﬁcation), and furthermore is only valid for payoﬀ
proﬁles depending quadratically on S. Note too that regularity issues arise if h =
1
2λ
,
which are not unrelated to the analysis of chapter 5.
4.3 Similarity solutions
In order to gain more analytical insight into the behaviour of the highly nonlinear
PDE (4.1) we can employ the powerful technique of similarity solutions. It should be
CHAPTER 4. FULLFEEDBACK MODEL 88
noted, however, that the similarity solution technique is rarely successful in solving a
complete boundary value or Cauchy problems, because it requires special symmetries
in the equation and the initial/boundary conditions. On the other hand, it can be
extremely useful when considering a local analysis in space or in time, for example the
initial behaviour of the American option freeboundary problem and the value of an
atthemoney option shortly before exercise, which are diﬃcult to resolve numerically.
The idea behind similarity solutions is to exploit the fact that the equations and
the initial (ﬁnal) and boundary conditions are invariant under a certain scaling, for
example S → νS, τ → ν
2
τ for any real number ν. Such a scaling is called a
oneparameter group of transformations. Under this transformations an equation is
invariant and so
S
√
τ
is the only combination of S and τ that is independent of ν and
hence the solution must be a function of
S
√
τ
only.
The global similarities (of Lie type) to equation (4.1) can be found in Bordag (2007).
In that paper, under the assumption that λ = λ(S), it is proved that λ ∼ S
k+1
where k ∈ R is the only case with a nontrivial symmetry group and a complete set
of invariant solutions is also found. Hence for equation (4.1) where λ(S, τ) =
ˆ
λS
k+1
,
there exists a global similarity transform of the form
V (S, τ) = S
1−k
u(z) where z = log S +aτ, a ,= 0. (4.2)
Note that k = −1 corresponds to the model of Sch¨ onbucher and Wilmott (2000), i.e.
with constant liquidity parameter λ. In addition the case k = 0 corresponds to the
model developed by Frey and his coworkers (see section 7.1). Substitution of (4.2)
into equation (4.1) gives the following (highly nonlinear) ODE
(a −r)u
z
+ rku −
σ
2
_
u
zz
+ (1 −2k)u
z
−k(1 −k)u
_
2
_
1 −
ˆ
λ
_
u
zz
+ (1 −2k)u
z
−k(1 −k)u
_
_
2
= 0. (4.3)
If k = 1 then equation (4.3) becomes the much simpler
(a −r)u
z
+ru −
σ
2
_
u
zz
−u
z
_
2
_
1 −
ˆ
λ(u
zz
−u
z
)
_
2
= 0, (4.4)
CHAPTER 4. FULLFEEDBACK MODEL 89
and when k = 0 (corresponding to the Frey model) we have the (even simpler)
equation
(a −r)u
z
−
σ
2
_
u
zz
+u
z
_
2
_
1 −
ˆ
λ(u
zz
+u
z
)
_
2
= 0. (4.5)
Under the assumption of nonzero interest rates we can quite easily obtain a solution
to (4.5) by setting a = r. This leads to the ODE
u
zz
+u
z
= 0
which has the most general solution
u(z) = A−Be
−z
where A and B are constants to be determined by the boundary and payoﬀ conditions.
After transforming back to the ﬁnancial variables this gives
V (S, τ) = AS −Be
−rτ
which is indeed a valid solution of equation (4.1), however it does not satisfy any
practical boundary and payoﬀ conditions. In addition Bordag and her coworkers
showed families of explicit solutions to equations (4.4) and (4.5) under the assumption
of zero interest rates (r = 0); for the k = 1 case see Bordag (2007) and for k = 0
see Bordag and Chmakova (2007) and Bordag and Frey (2007). However, all the
solutions found in this way correspond to diﬀerentiable payoﬀ proﬁles, which diﬀer
from the majority of the payoﬀ proﬁles considered in the present thesis (which are
more ﬁnancially relevant). However, these global solutions may be useful to test the
accuracy of numerical techniques applied to such highly nonlinear systems.
For the present model being studied, i.e. k = −1, equation (4.3) reduces to
(a −r)u
z
−ru −
σ
2
_
u
zz
+ 3u
z
+ 2u
_
2
_
1 −
ˆ
λ
_
u
zz
+ 3u
z
+ 2u
_
_
2
= 0,
which appears to have no analytic solution and so we are forced to turn to numerical
techniques. We can start to see already that singular behaviour is inherent in this
CHAPTER 4. FULLFEEDBACK MODEL 90
highly nonlinear system by rearranging the above equation into a quadratic equation
in u
zz
, i.e.
_
ˆ
λ
2
ψ
_
u
2
zz
+
_
σ
2
2
−2
ˆ
λψ + 2
ˆ
λ
2
φψ
_
u
zz
+
_
σ
2
φ
2
+ψ −2
ˆ
λφψ +
ˆ
λ
2
φ
2
_
= 0
where φ = 3u
z
+ 2u and ψ = (r − a)u
z
− ru. If we solve this using the quadratic
formula we obtain
u
zz
=
1
ˆ
λ
−φ −
σ
2
4
ˆ
λ
2
ψ
_
_
1 ±
_
1 −
8
ˆ
λψ
σ
2
_1
2
_
_
,
from which it is clear that diﬃculties will occur if the square root were to become
negative. We shall return to this in section 4.6.
4.4 Perturbation expansions
Since it appears that no analytical solution can be found with the required boundary
conditions, we are forced to turn to numerical solutions. However it is possible to
ﬁnd an approximate solution for small values of the parameter λ. This can be done
by exploiting the techniques of asymptotic expansions. First we rewrite equation
(4.1) in the more convenient form
(1 −λV
SS
)
2
(V
τ
−rSV
S
+rV ) −
1
2
σ
2
S
2
V
SS
= 0. (4.6)
Now we expand V (S, τ) as follows
V (S, τ) = V
0
(S, τ) +λV
1
(S, τ) +λ
2
V
2
(S, τ) +. . .
where V
n
(S, τ) are functions to be found. Substituting this expansion into (4.6) and
collecting together terms of the same order in λ gives
3
O(λ
0
) : V
0τ
−
1
2
σ
2
S
2
V
0SS
−rSV
0S
+rV
0
= 0,
O(λ
1
) : V
1τ
−
1
2
σ
2
S
2
V
1SS
−rSV
1S
+rV
1
= 2V
0SS
(V
0τ
−rSV
0S
+rV
0
) ,
O(λ
2
) : V
2τ
−
1
2
σ
2
S
2
V
2SS
−rSV
2S
+rV
2
= 2V
0SS
(V
1τ
−rSV
1S
+rV
1
)
+
_
2V
1SS
−V
2
0SS
_
(V
0τ
−rSV
0S
+rV
0
) .
3
Note that a similar perturbation analysis is outlined in the appendix of Sch¨ onbucher and
Wilmott (2000).
CHAPTER 4. FULLFEEDBACK MODEL 91
This reveals some structure in the successive approximations, the lefthandside is
merely the BlackScholes operator L
BS
acting on the n
th
approximation and the right
hand side is a function of the previous approximations (which have been found), i.e.
L
BS
V
0
= 0,
L
BS
V
1
= f
1
(V
0
),
L
BS
V
2
= f
2
(V
0
, V
1
),
.
.
.
L
BS
V
n
= f
n
(V
0
, V
1
, . . . , V
n−1
).
This recursive process is continued until the desired level of accuracy is required (al
though in practise solving past the second correction term V
2
becomes too analytically
cumbersome or numerically expensive). However, it should be noted that in order to
permit a regular asymptotic expansion of the kind outlined above we must assume
suﬃcient regularity in the function V (S, τ), speciﬁcally we need the derivatives of
V (S, τ) to be bounded.
4
This cannot be guaranteed a priori and the unbounded
second derivative of the payoﬀ proﬁle suggests that we may not be able to apply such
a regular expansion in the region around any singular points. For more on singular
perturbations see Johnson (2004).
Figure 4.1 shows the ﬁrstorder correction term V
1
(S, τ) to the price of a European
put option. These results indicate that (similar to ﬁrstorder feedback in the regime
[V
SS
[ < 1/λ) the inclusion of market illiquidity increases the put option price. More
over, it can be shown (see appendix A) that if we restrict ourselves to the regime
in which [V
SS
[ < 1/λ in the entire solution domain (corresponding to suﬃciently
smooth payoﬀ proﬁles) then the solution to equation (4.1) is monotonic increasing in
the liquidity parameter λ; this is in agreement with the result shown in ﬁgure 4.1.
4
See for example Johnson (2004).
CHAPTER 4. FULLFEEDBACK MODEL 92
0
0.05
0.1
0.15
0.2
0.25
0.3
0.6 0.8 1 1.2 1.4
P
S
f
r
a
g
r
e
p
l
a
c
e
m
e
n
t
s
V
1
(
S
,
τ
)
S
τ increasing
Figure 4.1: The leading order correction term V
1
(S, τ) to the BlackScholes (i.e.
λ = 0) European put option for various time to expiry. K = 1, r = 0.04, σ = 0.2,
T = 1 and τ = 0.1, 0.2, . . . , 1.
4.5 Numerical solutions
Consider next a numerical treatment of (4.1) with λ constant for simplicity, subject
to the put payoﬀ condition (1.2). Figure 4.2 shows results obtained using a simi
lar CrankNicolson scheme to that successfully employed on the ﬁrstorder feedback
model, but of course, incorporating iteration in order to treat properly the inherent
nonlinearity in the problem. The results (for the delta) are clearly erroneous, even
though they were obtained with a relatively ﬁne grid (timestep δτ of 10
−3
, gridsize
δS of 5 10
−4
); in addition, the output was found to be highly dependent on the
choice of grid. Note that this erroneous behaviour is not simply due to the well
documented ‘ringing’ behaviour associated with the CrankNicolson ﬁnitediﬀerence
scheme (see Duﬀy, 2004).
This sort of diﬃculty is understandably sidestepped in published works (for speciﬁc
details see chapter 7) but a study of its causes will surely be helpful for the next
phase of modelling in the ﬁeld. In fact there are two problematic issues with regard
to these diﬃculties, which are not unconnected. The ﬁrst is linked to the inevitable
inﬁnite behaviour of the gamma with standard payoﬀ conditions, which even a cur
sory inspection of (4.1) suggests will be problematic; this is considered below. The
CHAPTER 4. FULLFEEDBACK MODEL 93
second diﬃculty (again revealed by a cursory inspection of (4.1)) is the likelihood of
diﬃculties if there is a zero in the denominator of the volatility term. A discussion of
this issue, which is associated with smoothed payoﬀ functions, will be deferred until
chapter 5.
1.5
1
0.5
0
0.5
1
1.5
0.96 0.98 1 1.02 1.04
PSfrag replacements
∆
=
∂
V
∂
S
S
τ = 1
τ = .1
Figure 4.2: Deltas for fullfeedback (European) put, K = 1, r = 0.04, σ = 0.2,
λ = 0.1 and T = 1.
4.6 Analysis close to expiry
As noted earlier, a thorough asymptotic analysis of the option valuation close to
maturity (τ →0) can yield signiﬁcant insight into the dynamics of the problem, and
consequently this limit is studied next. For this we seek a local solution for the put
value of the form (which can be justiﬁed a posteriori )
V (S, τ) = −τ
1
2
ηH(−η) +τφ(η) +. . . (4.7)
where η is deﬁned in (3.4). H() denotes the Heaviside function, which is necessary to
‘mimic’ the behaviour of the payoﬀ, close to expiry. Consequently, we have a diﬀerent
form for the valuation equation in two regions, one in S > K (above the strike) and
the other in S < K (below the strike). Thus, although the option value is assumed to
be continuous, clearly we are allowing for a discontinuous delta close to expiry at the
CHAPTER 4. FULLFEEDBACK MODEL 94
exercise price. Indeed, we sought solutions with continuous deltas, without success,
and it is our assertion that such solutions do not exist for this problem. It should
be noted that the above indicates that the crucial regime is within a distance O(τ
1
2
)
of the exercise price as τ → 0 (a result determined through asymptotic analysis),
similar to the λ = 0 liquid options (as discussed in the previous section), which is
rather broader than the scale appropriate for the ﬁrstorder feedback options (where
S −K = O(τ)).
In the region S > K (i.e. η > 0) the following equation describes φ:
φ −
η
2
φ
η
−
σ
2
K
2
φ
ηη
2 (1 −λφ
ηη
)
2
= 0, (4.8)
with φ →0 as η →∞. In the region S < K (i.e. η < 0) the appropriate equation is
φ −
η
2
φ
η
−
σ
2
K
2
φ
ηη
2 (1 −λφ
ηη
)
2
+rK = 0, (4.9)
with φ → −rK as η → −∞. At η = 0, smooth pasting (φ, φ
η
continuous) is
appropriate. Sample results for a put option are shown in ﬁgure 4.3 (obtained via a
straightforward RungeKutta fourthorder shooting method). These results indicate
that the option values all lie below the payoﬀ
5
(the repercussions of this will be
discussed below). Note also the slower decay to the [η[ → ∞ asymptotes as the
volatility increases due to the O(σ
2
K
2
) scaling that emerges from (4.8) and (4.9) in
these limits.
It is helpful to shift φ as follows:
φ = φ
∗
−
rK
2
,
which leads to the equation
φ
∗
−
η
2
φ
∗
η
−
σ
2
K
2
φ
∗
ηη
2
_
1 −λφ
∗
ηη
_
2
+
rK
2
[2H(−η) −1] = 0, (4.10)
which has the useful property of antisymmetry of φ
∗
with respect to η = 0 (and so
φ
∗
→
sgn(η)
2
rK as [η[ →∞). In addition, it enables us (with a little work) to deduce
5
Since from (4.7) and (3.4) it is evident that V (S, τ) −(K −S)
+
≡ τφ(η).
CHAPTER 4. FULLFEEDBACK MODEL 95
the results for calls from the results for puts, namely
φ
C
(η) = φ
P
(−η) + rK, (4.11)
i.e., we can recover the local solution for calls from that of puts. Note that this
symmetry of the local solutions is simply a manifestation of the putcall parity rela
tionship which still holds for all time, even in this highly nonlinear case (see section
4.1); provided that early exercise is not permitted. The standard BlackScholes put
call parity is given by
V
P
= V
C
−S +Ke
−r(T−t)
.
For the nonlinear problem, the chosen scaling for the inner region is given by
S = K +ητ
1
2
,
V
P
= −ητ
1
2
H(−η) +τφ
P
(η) +o(τ),
V
C
= ητ
1
2
H(η) +τφ
C
(−η) +o(τ),
e
−rτ
= 1 −rτ +o(τ).
Note that the scaling for a call was obtained by replacing η by −η in the put scaling.
Substitution thus gives, after a little rearranging
ητ
1
2
[1 −H(−η) −H(η)] +τ
_
φ
P
(η) −φ
C
(−η)
¸
= −rτK,
which reduces to
φ
C
(−η) = φ
P
(η) + rK,
which corresponds to the symmetry obtained for the inner equations, namely (4.11),
conﬁrming putcall parity for the nonlinear case.
The key observation in the above is the discontinuity in the delta (∆ =
∂V
∂S
) at η = 0
as indicated in (4.7), and it is the neglect of this that is undoubtedly responsible for
the apparent spurious results observed in ﬁgure 4.2. Another point to be noted is
that ﬁgure 4.3 indicates the possibility of negative put options values, a somewhat
undesirable property (although (4.11) indicates this is not the case with calls).
CHAPTER 4. FULLFEEDBACK MODEL 96
Before a consideration of the problem for calculations for nonsmall values of τ (i.e.
at times away from expiry), it turns out that yet another anomaly occurs, this time
in the limit as σ decreases (with other parameters held ﬁxed). For values of σ just
below 0.15 (taking the other parameters used in ﬁgure 4.2), the numerical treatment
applied to (4.8) and (4.9) failed, with the onset of negative roots in the computation.
0.04
0.035
0.03
0.025
0.02
0.015
0.01
0.005
0
0.6 0.4 0.2 0 0.2 0.4 0.6
PSfrag replacements
σ decreasing
η
φ
Figure 4.3: Local (τ → 0) solution of a fullfeedback put, K = 1, λ = 0.1, r = 0.04
and σ = 1, 0.95, . . ., 0.15.
To understand this, we rewrite (4.8) and (4.9) in the form of a quadratic in φ
ηη
;
2λ
2
ψφ
2
ηη
−
_
4λψ +σ
2
K
2
_
φ
ηη
+ 2ψ = 0,
where ψ = φ −
η
2
φ
η
+ rKH(−η). Using the quadratic formula we can write the
‘solution’ for φ
ηη
as
φ
ηη
=
1
λ
+
σ
2
K
2
4λ
2
ψ
_
1 −
_
1 +
8λψ
σ
2
K
2
_1
2
_
,
where we have taken the negative root in order to satisfy the condition that φ
ηη
→0
as [η[ → ∞. Indeed, this is the form that was taken as the basis of the numerical
treatment used to treat the results shown in ﬁgure 4.3, and inspection of the results
indicated that diﬃculties arose if
1 +
8λψ
σ
2
K
2
< 0.
CHAPTER 4. FULLFEEDBACK MODEL 97
Hence, we may expect this regime to arise for large values of the ratio λ/σ
2
K
2
, i.e.
for suﬃciently large λ, or suﬃciently small σ (since ψ must be an odd function about
η = 0, as evidenced by 4.10). We shall return to a consideration of this regime later
in section 4.7.1, which concerns itself with the full problem.
4.7 Numerical results  full problem
We now revisit the choice of parameters employed in ﬁgure 4.2, which led to the
aforementioned diﬃculties. The analysis in the previous section points to
6
a discon
tinuity in the delta (∆) at the strike price (S = K) of +1 in the case of a put option.
In order to incorporate this into our numerics, an alternative strategy was adopted,
based on the Keller (1978) scheme. This modiﬁed procedure involved writing (4.1) as
a system of two ﬁrstorder equations namely in V (S, τ) and V
1
(S, τ) = ∂V/∂S. The
grid was then chosen in such a manner that the strike price K coincided with the S
grid. At S = K two values of the option price and its delta were computed, namely
V
−
and V
−
1
(for S
−
= K) and V
+
and V
+
1
(for S
+
= K), such that V
−
= V
+
and
V
+
1
= V
−
1
−1. This latter condition eﬀectively builds the proposed jump in the delta
at the strike price into the numerical scheme. In the timewise direction, a standard
CrankNicolsontype scheme was adopted. Calculations performed in this manner
provided accurate and highly reliable results, as evidenced in ﬁgure 4.4, showing dis
tributions of V (S, τ) − max(K − S, 0), i.e. the diﬀerence between the option value
and payoﬀ, and as such can be compared directly with the smalltimetomaturity
solutions displayed in ﬁgure 4.3. Furthermore, ﬁgure 4.5 shows the corresponding
distributions of the delta (
∂V
∂S
), clearly indicating the jump in its value at S = K.
The computations shown are highly robust (i.e. grid independent), which adds sig
niﬁcant credence to the integrity of the results, in particular to the correctness of the
jump condition. These results also help to justify of the original form of the solution,
(4.7).
6
Since no inner solution with a continuous delta could be found.
CHAPTER 4. FULLFEEDBACK MODEL 98
There is, however, a further issue relating to the results observed in ﬁgure 4.4, namely
that this indicates the put option value (close to expiry) is always less then the option
payoﬀ. This has implications for the pricing of American options in this framework,
because if the European option value is always below the payoﬀ immediately prior to
expiry then, by a simple backward induction argument, the corresponding American
option will always be exercised immediately when the contract is initiated at t = 0 (or
τ = T), i.e. the solution to the American put will trivially correspond to the payoﬀ
for all time; this could also be regarded as a somewhat undesirable and unrealistic
feature of the model.
A corollary to the above remarks is that it can also be seen (from ﬁgure 4.4) that
the model permits negative values for put options. Whilst in certain extreme option
valuations, such as those involving storage costs, this may be acceptable, generally
this may be regarded as an unwanted facet of the model. It makes little ﬁnancial
sense to allow negative option values in any model incorporating market frictions,
at least under the dynamic hedging (replication) pricing paradigm. For example, in
transaction cost models the writer would not rehedge the portfolio (and hence incur
extra transaction costs) at times when he does not need to, provided the option is still
perfectly hedged. The same is true for liquidity, the price being modelled is the cost
of replicating the option by trading in the underlying. In doing this, we have freedom
in our hedging strategy, provided it perfectly replicates the option payoﬀ. Essentially
the hedging strategy should never force the hedger into an irrational position. This
could be avoided in practise by imposing the condition V ≥ 0 which eﬀectively creates
another free boundary on the PDE at S
b
where the conditions V (S
b
(τ), τ) = 0 and
∂V
∂S
(S
b
(τ), τ) = 0 should be applied. Note that Bakstein and Howison (2003) make a
similar observation and call this condition the ‘American’ constraint.
Figure 4.6 shows results (option value  payoﬀ) for the corresponding call. This
clearly reveals that call values not only remain positive, but are also always above
the payoﬀ and, hence, indicates that there is no value in early exercise.
CHAPTER 4. FULLFEEDBACK MODEL 99
0.04
0.035
0.03
0.025
0.02
0.015
0.01
0.005
0
0 0.5 1 1.5 2
PSfrag replacements
V

p
a
y
o
ﬀ
S
τ = 1
τ = .1
Figure 4.4: Full feedback put, K = 1, r = 0.04, σ = 0.2 and λ = 0.1; modiﬁed
numerical scheme.
1
0.8
0.6
0.4
0.2
0
0.2
0.96 0.98 1 1.02 1.04
PSfrag replacements
∆
S
τ = 1
τ = .1
Figure 4.5: Full feedback put, K = 1, r = 0.04, σ = 0.2 and λ = 0.1; modiﬁed
numerical scheme.
4.7.1 A second solution regime
Returning now to the other regime outlined in section 4.6, i.e. when 1 +
8λψ
σ
2
K
2
< 0,
which turns out to be even more problematic, since here even the τ ¸ 1 regime is
unclear. It was therefore decided to mount an homotopy type of approach in this
regime, speciﬁcally by considering a payoﬀ function of the form
V (S, 0) =
1
2
_
K −S +
_
(K −S)
2
+ρ
2
_
(4.12)
CHAPTER 4. FULLFEEDBACK MODEL 100
0
0.005
0.01
0.015
0.02
0.025
0.03
0.035
0.04
0 0.5 1 1.5 2
PSfrag replacements
V

p
a
y
o
ﬀ
S
τ = 1
τ = .1
Figure 4.6: Full feedback call, K = 1, r = 0.04, σ = 0.2 and λ = 0.1; modiﬁed
numerical scheme.
in conjunction with the full problem (4.1). In this way, it is possible to mimic a
standard put payoﬀ as the smoothing parameter ρ →0.
In Frey and Stremme (1997) and Frey (1998) this smoothed payoﬀ proﬁle was used
to represent an ‘idealised’ option payoﬀ, which represented a welldiversiﬁed portfo
lio containing a multitude of diﬀerent payoﬀs with diﬀerent strikes which combine
to produce a suﬃciently smooth payoﬀ to satisfy the smoothness assumptions im
posed for existence and uniqueness. Here its use is slightly diﬀerent, it is merely a
mathematical tool to investigate the limit of smoothness. Results corresponding to
the parameter choice of ﬁgure 4.4, but instead with σ = 0.1 and at a time shortly
before expiry (τ = 0.1) and for three choices of ρ are shown in ﬁgure 4.7. These
results were based on the method employed for ﬁgure 4.2, but were tested extensively
for numerical grid convergence and found to be numerically consistent on the scale
shown.
These calculations strongly indicate that in the limit as ρ → 0, the solution for the
put takes the trivial form:
V (S, τ) =
_
_
_
0 for S > K,
Ke
−rτ
−S for S < K,
(4.13)
CHAPTER 4. FULLFEEDBACK MODEL 101
0.0004
0.0003
0.0002
0.0001
0
1e04
0.0002
0.0003
0.0004
0.96 0.98 1 1.02 1.04
PSfrag replacements
ρ = .0005
ρ = .001
ρ = .00025
S
V

p
a
y
o
ﬀ
Figure 4.7: Full feedback put, smoothed payoﬀ, K = 1, r = 0.04, σ = 0.1, λ = 0.1
and τ = 0.01.
for all time, solutions which do (trivially) satisfy (4.1).
Note that this form of solution indicates discontinuous option values (compare the
small volatility analysis of Widdicks et al., 2005; Duck et al., 2008). Here the diﬀusion
term eﬀectively eliminates itself completely and the discontinuity at S = K and τ = 0
cannot propagate away from this point for τ > 0 since there is no diﬀusion. Note also
that (4.13) indicates that American options in this regime will always be exercised
immediately (at t = 0), for the same reasons expounded earlier for the other regime.
One interpretation of the above results is that the eﬀect of the nonlinearity, for
standard (nonsmooth) payoﬀ proﬁles, is to suppress the diﬀusion term of the equation
in regions of nonsmoothness, thereby failing to smooth out any discontinuities in the
derivative of the payoﬀ proﬁle, as would normally be the case with the BlackScholes
equation.
The following chapter investigates another breakdown of the nonlinear PDE, (4.1),
which occurs for smoothed payoﬀ proﬁles. Such a breakdown can be seen to be a
direct result of the singular nature of the diﬀusion coeﬃcient in the the governing
equation.
Chapter 5
Smoothed Payoﬀs  Another
Breakdown
The diﬃculties encountered in the previous chapter have been, so far, attributed
to the discontinuous delta (i.e. inﬁnite gamma) of the payoﬀ proﬁle. If we instead
assume smoothness in the payoﬀ (i.e. ﬁnite gamma), it can be seen that there is the
potential for further diﬃculties to arise due to the vanishing of the denominator in
(4.1), i.e. when
1
V
SS
=
1
λ
. (5.1)
Firstly, note that for standard put and call payoﬀ proﬁles, condition (5.1) must always
be satisﬁed somewhere in the domain T ⊆ R
+
[0, T]; since the solution must pass
from V
SS
= ∞ at (K, 0) to V
SS
→ 0 as τ → ∞. Secondly, note that for suﬃciently
smooth payoﬀ proﬁles, condition (5.1) may never be satisﬁed in the solution domain.
To illustrate the circumstances under which we should expect such singular behaviour,
we will once again consider the smoothed payoﬀ proﬁle
V (S, 0) =
1
2
_
K −S +
_
(K −S)
2
+ρ
2
_
(5.2)
where ρ ≥ 0. This function is smooth, but in the limit as ρ → 0 recovers the
discontinuous payoﬀ proﬁle of a put option. One can consider this as parameterising
1
Note that for simplicity here λ is a constant but can be generalised in what follows.
102
CHAPTER 5. SMOOTHED PAYOFFS  ANOTHER BREAKDOWN 103
the smoothness of the payoﬀ proﬁle by ρ, with the limit ρ → 0 representing highly
nonsmooth functions (in the sense of very large second derivatives in the region of the
strike) and conversely the limit ρ → ∞ representing increasingly smooth functions
(small second derivatives).
If we wish to prevent the denominator from vanishing then this can be seen as placing
a restriction on the size of the liquidity function λ(S, τ), i.e. we must have
λ(S, τ) ≤ sup
(S,τ)∈D
_
1
V
SS
_
,
or alternatively as placing a restriction on the payoﬀ proﬁle, i.e.
sup
(S,τ)∈D
¦V
SS
¦ ≤
1
λ(S, τ)
.
For the analysis in the remainder of this chapter λ(S, τ) will be considered constant
for simplicity.
Furthermore it can be shown, via a judicious application of the maximum principle,
outlined in appendix A, that the maximum of the second Sderivative of the solution
in the entire domain will coincide with the maximum at τ = 0, in other words,
sup
(S,τ)∈D
¦V
SS
¦ = sup
(S,τ)∈D
0
¦V
SS
¦ ,
where T
0
= R
+
¦0¦. In fact it is intuitively clear from the diﬀusive nature of
equation (4.1) for increasing τ that this should be so. With this in mind the crucial
property in determining the existence of singular behaviour will be the maximum of
the second derivative of the solution at τ = 0, i.e. the payoﬀ proﬁle. Returning to
the smoothed payoﬀ proﬁle (5.2) direct computation gives
V
S
(S, 0) = −
1
2
_
1 +
(K −S)
_
(K −S)
2
+ρ
2
_
, (5.3a)
V
SS
(S, 0) =
ρ
2
2
_
(K −S)
2
+ρ
2
_3
2
, (5.3b)
V
SSS
(S, 0) =
3ρ
2
(K −S)
2
_
(K −S)
2
+ρ
2
_5
2
, (5.3c)
V
SSSS
(S, 0) =
3ρ
2
2
_
_
4(K −S)
2
−ρ
2
_
(K −S)
2
+ρ
2
_7
2
_
_
. (5.3d)
CHAPTER 5. SMOOTHED PAYOFFS  ANOTHER BREAKDOWN 104
The maximum of V
SS
(S, 0) is determined by setting V
SSS
(S, 0) = 0, which yields
(obviously) that the maximum occurs at S = K. To check that this is indeed a
maximum of V
SS
(S, 0) we can see that
V
SSSS
(K, 0) =
−3
2ρ
3
≤ 0
since ρ ≥ 0 by deﬁnition. Therefore the maximum of the second derivative of the
payoﬀ proﬁle is given by
sup
D
0
¦V
SS
¦ = V
SS
(K, 0) =
1
2ρ
,
and we can exclude the denominator from vanishing if we place the restriction that
λ ≤ 2ρ, (5.4)
from which it can be seen that the smoother the payoﬀ proﬁle the more liquidity the
model can handle. Also a corollary to this result is that if we have a payoﬀ proﬁle
with a discontinuous ﬁrst derivative (delta), which includes the majority of the payoﬀ
proﬁles used in practise, then to restrict the denominator from vanishing it is required
to set λ = 0 and so this model cannot treat nonsmooth payoﬀ proﬁles.
Condition (5.4) may seem rather restrictive and indeed it is when considering standard
put and call payoﬀ proﬁles. In what follows we shall assume smooth payoﬀ proﬁles
and investigate the nature of the singularities that arise if this restriction is not
imposed, i.e. when we are in the regime that λ > 2ρ.
It should be noted at this stage that the results for existence and uniqueness of a
replicating portfolio provided by Frey (1998) only apply when the denominator is not
allowed to vanish (here we impose no such restriction). It should also be mentioned
that problems associated with the vanishing of the denominator have been highlighted
previously in the literature, but that this regime has deliberately been avoided. For
example Sircar and Papanicolaou (1998) set the option value to be the BlackScholes
price a small time prior to expiry, where is determined to be suﬃciently large
such that the denominator in the diﬀusion term is always positive  see section 7.3.
CHAPTER 5. SMOOTHED PAYOFFS  ANOTHER BREAKDOWN 105
In addition Lipton (2001) states that to avoid any undesirable behaviour of the the
option prices we have to limit the magnitude of the local volatility from above and
from below, although he gives no suggestions about how to do this. Along the same
lines Frey and Patie (2002) modify the diﬀusion term of the equation in an ad hoc
manner, more speciﬁcally they set
ˆ σ(S, τ) = max
_
δ
0
,
σS
1 −min ¦δ
1
, λV
SS
¦
_
for suﬃciently large δ
0
and suﬃciently small δ
1
. This is done in order to bypass
any problems associated with the limits σ → 0 or σ → ∞, hence ensuring the
denominator is nonzero and that the denominator does not become too large to
‘annihilate’ the diﬀusion term. Finally Liu and Yong (2005) suggest a form of the
liquidity function λ(S, τ) that is hoped to suppress such singular behaviour, this has
the eﬀect of fundamentally changing the option price dynamics close to expiry, see
section 7.5. Here we make no such modiﬁcations and attempt to fully investigate the
nature of these singularities.
We can expect singular behaviour of (4.1) when the singularity condition (5.1) is
satisﬁed. For the smooth payoﬀ proﬁle (5.2) we can calculate the explicit locations
of any singularities (denoted S
0
) at τ = 0 by equating the second derivative (5.3b)
to 1/λ. Doing so yields
S
0
= K ±
¸
_
λρ
2
2
_2
3
−ρ
2
.
Hence for the payoﬀ (5.2) there are two singularities, each equally spaced either side of
the strike K. Note that the solutions at τ = 0 are not themselves singular, but what
remains is to determine if and how singularities propagate through the solution for
τ > 0. To solve the full equation (4.1) near these singular points will be particularly
diﬃcult, both analytically and numerically, due to the inherent singular behaviour.
Instead a local similarity solution is to be attempted which reduces the PDE (4.1)
to a simpler ODE valid locally in the region close to the singularity. This analysis is
outlined in the next section.
CHAPTER 5. SMOOTHED PAYOFFS  ANOTHER BREAKDOWN 106
5.1 Local analysis about the singularities
We consider a general form of payoﬀ proﬁle, which we have determined to have an
isolated singularity in the payoﬀ proﬁle at S = S
0
; isolated in the sense that there are
no other singularities in the vicinity of S
0
.
2
A local expansion for small τ is sought
about the point where the equation becomes singular, i.e. where V
SS
= 1/λ. Since
we are assuming a smooth payoﬀ proﬁle, the solution in the vicinity of the singular
point S
0
must be analytic (at least for τ = 0) and so can be expressed in the form of
a Taylor series about S = S
0
, i.e.
V (S, 0) = V (S
0
, 0)+(S−S
0
)V
S
(S
0
, 0)+
(S −S
0
)
2
2
V
SS
(S
0
, 0)+
(S −S
0
)
3
6
V
SSS
(S
0
, 0)+. . . .
(5.5)
In addition, to remain close to the singular point at τ = 0 we also require that
V
SS
(S
0
, 0) =
1
λ
.
Next we seek a similarity solution for small τ of the form
V (S, τ) = V (S, 0) +τ
β
ˆ
V (η), (5.6)
with
η =
S −S
0
τ
α
,
where α and β are to be determined from the appropriate balancing of terms and
asymptotic matching (cf. section 3.1). Combining the Taylor series about S
0
(5.5)
and the small τ expansion (5.6) we therefore seek a similarity solution in the vicinity
of the singularity of the form
V (S, τ) = V
0
+τ
α
ηV
1
+τ
2α
η
2
2λ
+τ
β
ˆ
V (η) +. . . , (5.7)
where V
0
= V (S
0
, 0) and V
1
= V
S
(S
0
, 0); both of which are assumed to be known,
given the form of the payoﬀ proﬁle. Direct substitution of (5.7) into (4.1) gives
τ
β−1
_
β
ˆ
V −αη
ˆ
V
η
_
−
σ
2
S
2
0
_
λ
−1
+τ
β−2α
ˆ
V
ηη
_
2λ
2
τ
2β−4α ˆ
V
2
ηη
−rS
0
_
V
1
+
τ
α
η
λ
+τ
β−α
ˆ
V
ηη
_
+r
_
V
0
+τ
α
ηV
1
+
τ
2α
η
2
2λ
+τ
β
ˆ
V
_
= 0.
(5.8)
2
Note that for the case of the put or call payoﬀ proﬁle (i.e. ρ = 0) then the two singularities will
coincide, at the strike price, and so can not be thought of as isolated anymore and the following
analysis will not be appropriate.
CHAPTER 5. SMOOTHED PAYOFFS  ANOTHER BREAKDOWN 107
To balance the diﬀusion term and the time derivative term in (5.8) it is clear that we
require
β −1 = 4α −2β ⇒ 3β −4α = 1. (5.9)
To ﬁx the values of α and β, the matching of the inner and outer solution is used.
To be consistent, the form of the solution in the vicinity of the singularity (5.7) in
the limit as [η[ →∞ (i.e. as τ →0) must match with the Taylor series expansion of
the solution at τ = 0, i.e. (5.5). Hence
lim
η→∞
_
V
0
+τ
α
ηV
1
+τ
2α
η
2
2λ
+τ
β
ˆ
V (η) +. . .
_
= V
0
+(S−S
0
)V
1
+
(S −S
0
)
2
2λ
+
(S −S
0
)
3
6
V
3
+. . . ,
where we have deﬁned V
i
=
∂
i
V
∂S
i
(S
0
, 0). Since the ﬁrst three terms on both sides are
identical (by construction) this reduces to
lim
η→∞
_
τ
β
ˆ
V (η) +. . .
_
=
(S −S
0
)
3
6
V
3
+. . .
⇒ lim
η→∞
_
τ
β
ˆ
V (η)
_
=
τ
3α
η
3
6
V
3
+. . . .
For a nontrivial inner solution
ˆ
V (η) we require that
ˆ
V = O(1) as τ →0, this forces
us to set
β = 3α. (5.10)
Note that the above matching procedure has also provided us with the appropriate
boundary conditions (for large η) of the inner solution, i.e. that
ˆ
V (η) →
η
3
V
3
6
as [η[ →∞. (5.11)
Substituting (5.10) into (5.9) we ﬁnd that
α =
1
5
, β =
3
5
,
hence the appropriate form of the solution to try around the singularity is given by
V (S, τ) = V
0
+τ
1
5
ηV
1
+τ
2
5
η
2
2λ
+τ
3
5
ˆ
V (η) +. . . ,
which after substitution into (4.1) and evaluating in the limit τ →0 yields
3
ˆ
V −η
ˆ
V
η
−
5σ
2
S
2
0
2λ
3 ˆ
V
2
ηη
= 0. (5.12)
CHAPTER 5. SMOOTHED PAYOFFS  ANOTHER BREAKDOWN 108
Before attempting to solve (5.12) given the appropriate boundary conditions (5.11),
we ﬁrst look a little closer at the matching procedure performed above and the rele
vant boundary conditions that arise.
5.1.1 Asymptotic matching
Let us look again more closely at the boundary conditions (5.11) that were derived
from an asymptotic matching procedure. We can see that (5.11) is not a solution of
the inner equation (5.12), rather it is only the leading order term in the asymptotic
matching procedure, which must contain higher order matching terms. We can de
termine the next order correction by seeking a solution to the inner equation (5.12)
of the algebraic form
ˆ
V (η) =
η
3
V
3
6
+Aη
γ
, (5.13)
where A and γ are constants to be found. Note that the second term is a higher order
correction as [η[ →∞ iﬀ γ < 3. Substitution into (5.12) gives
(3 −γ)Aη
γ
−
5σ
2
S
2
0
2λ
3
_
ηV
3
+γ(γ −1)Aη
γ−2
_
2
= 0.
Since we are interested in the solution as [η[ →∞ we can approximate the denomi
nator as follows
(3 −γ)Aη
γ
−
5σ
2
S
2
0
2λ
3
η
2
V
2
3
_
1 +
γ(γ−1)A
V
3
η
γ−3
_
2
= 0,
⇒(3 −γ)Aη
γ
−
5σ
2
S
2
0
2λ
3
η
2
V
2
3
_
1 −
2γ(γ −1)A
V
3
η
γ−3
+. . .
_
= 0.
Hence in the limit [η[ →∞ we must have γ = −2, which leads to the equation
_
5A−
5σ
2
S
2
0
2λ
3
V
2
3
_
1
η
2
= 0(η
−5
),
therefore we have that the constant A must be given by
A =
σ
2
S
2
0
2λ
3
V
2
3
.
It is clear that we have an inﬁnite asymptotic series solution as [η[ → ∞ with each
subsequent term corresponding to a higherorder term in the Taylor series expansion
CHAPTER 5. SMOOTHED PAYOFFS  ANOTHER BREAKDOWN 109
of the outer solution. It appears that the inner solution is a power series in η starting
from η
3
and decreasing by powers of ﬁve; these further terms can be determined
by performing the same procedure adopted above. To summarise, the matching
condition (5.11) is modiﬁed to
ˆ
V (η) =
η
3
V
3
6
+
σ
2
S
2
0
2λ
3
V
2
3
η
2
+O(η
−7
) as [η[ →∞. (5.14)
The signiﬁcance of the extra term can be seen if we transform the large η behaviour
(5.14) back to the outer variables. Doing so we have that
ˆ
V (η) =
V
3
(S −S
0
)
3
6τ
3
5
+
σ
2
S
2
0
τ
2
5
2λ
3
V
2
3
(S −S
0
)
2
+O
_
τ
7
5
(S −S
0
)
−7
_
,
and the form of the similarity solution (5.7) in terms of the outer variable is given by
V (S, τ) =V
0
+ (S −S
0
)V
1
+
1
2λ
(S −S
0
)
2
+
V
3
6
(S −S
0
)
3
+
τσ
2
S
2
0
2λ
3
V
2
3
(S −S
0
)
2
+O
_
τ
2
(S −S
0
)
−7
_
.
(5.15)
This shows that the matching procedure results in integer powers of τ, suggesting
strongly that the scaling used is the correct scaling for this problem.
One ﬁnal point to note is that the sign of V
3
may aﬀect the qualitative behaviour of
the solution. V
3
is identiﬁed as the third derivative of the payoﬀ proﬁle evaluated at
the location of the singularity. Returning to the smoothed payoﬀ for a moment we
can see with a little work that
V
3
= ∓
3ρ
2
2
_
2
λρ
2
_5
3
¸
_
λρ
2
2
_2
3
−ρ
2
,
which shows that V
3
can be either positive or negative depending on which of the two
singularities we are seeking a local expansion near.
Now that we have the corrected boundary conditions (5.14) to the inner equation
(5.12) we can attempt to solve this nonlinear system, which is the focus of the next
section.
CHAPTER 5. SMOOTHED PAYOFFS  ANOTHER BREAKDOWN 110
5.1.2 Properties of the inner solution
For simplicity we shall rewrite the inner equation (5.12) as
ˆ
V
2
ηη
_
3
ˆ
V −η
ˆ
V
η
_
= κ (5.16)
where κ =
5σ
2
S
2
0
2λ
3
. The ﬁrst point to note is that κ can be scaled out of the problem
by setting
ˆ
V = κ
1
3
ˆ
V , however this scaling would then place the constant κ into the
boundary conditions (5.14). If we are a little more sophisticated we could perform
the transformation
ˆ
V = κ
3
5
ˆ
V ,
η = κ
1
5
η,
which would remove the constant from the equation and also from the leading term
of the boundary condition, however it would not be removed completely from the
boundary condition and consequently we shall not make any such scalings.
It is also noted that an exact solution of the ODE (5.16) exists. Trying a solution of
the form
ˆ
V (η) = Bη
χ
where B and χ are constants, leads to χ = 4/3 and to the exact solution
ˆ
V (η) =
_
243κ
80
_1
3
η
4
3
=
_
3
2
_5
3
(σS
0
)
2
3
λ
η
4
3
. (5.17)
However this clearly does not satisfy the matching condition (5.14), i.e. η
3
leading
order behaviour as [η[ →∞. Therefore in order to obtain a solution with the required
boundary conditions we must turn to numerical techniques, such as shooting or ﬁnite
diﬀerence methods.
Unfortunately, numerical solutions of (5.16) with the boundary conditions above
proved fruitless; shooting methods ﬂoundered and ﬁnitediﬀerence schemes failed to
converge. The previous statement suggests that a solution to equation (5.16) subject
to the boundary conditions (5.14) may not exist. The remainder of this chapter
CHAPTER 5. SMOOTHED PAYOFFS  ANOTHER BREAKDOWN 111
attempts to prove that this is, in fact, the case. An investigation of the phase portrait
of equation (5.16) using phaseplane analysis can provide us with such a proof and
also give us invaluable insight into the qualitative behaviour of the solutions to (5.16).
This analysis will be outlined in the following subsection.
5.1.3 Introduction to phaseplane analysis
A useful tool for the study of diﬀerential equations, especially if they are in two
dimensions, is the socalled phase portrait. Below we shall attempt to provide a
brief overview of the main properties of such phase portraits. For a more detailed
introduction see for example Jordan and Smith (1999) or Hirsch and Smale (1974)
and the references therein. Any general second order (autonomous) ODE of the form
u
xx
= g(u, u
x
)
can be expressed as two coupled ﬁrst order ODEs by deﬁning the new variable
v := u
x
,
hence
_
_
_
u
x
= v,
v
x
= g(u, v).
More generally we can have
u
x
= f(u, v), (5.18a)
v
x
= g(u, v). (5.18b)
The emphasis of phase portraits is on the general qualitative properties of diﬀerential
equations and their solutions rather than ﬁnding a closed form solution. This indeed
becomes extremely useful when such systems have no such closed form solutions.
Eliminating the x variable in (5.18) gives the following
dv
du
=
g(u, v)
f(u, v)
,
CHAPTER 5. SMOOTHED PAYOFFS  ANOTHER BREAKDOWN 112
where the variables u and v provide the axes for the phase portrait. This derivative
represents the ﬁeld direction in the phase portrait. The only problem that can arise
is if we ever have f(u, v) = g(u, v) = 0, since at these points the equation becomes
singular and nothing can be said about the direction of the ﬁeld at these points.
One might assume that a zero in the denominator, i.e. f(u, v) = 0 alone would cause
problems, however in this case we could simply shift the axis and consider
du
dv
=
f(u, v)
g(u, v)
,
which would give zero gradient in the (v, u) plane, which corresponds to a vertical
slope in the (u, v) plane.
The points where f(u, v) = g(u, v) = 0 are identiﬁed as ﬁxed points (also called
equilibrium or stationary points) of the system, so called because if a solution starts
at (or reaches) a ﬁxed point, then it remains at that point for all x since here u
x
=
u
xx
= 0. In addition, phase paths cannot normally intersect, if they do then this
would contradict uniqueness. In fact the only place where phase paths can intersect
is when either f(u, v) or g(u, v) are singular, which includes the ﬁxed points.
If we assume a linear system of the form
u
x
= au +bv,
v
x
= cu +dv,
which can be better represented in matrix form as
_
_
u
v
_
_
x
=
_
_
a b
c d
_
_
_
_
u
v
_
_
,
or more concisely
u
x
= A.u, (5.19)
then the only possible ﬁxed point of such systems are at u = v = 0. It can be shown
that the nature (and stability) of the ﬁxed point is determined by the solution of the
linear system (5.19) and, as it transpires, the determinant of the coeﬃcient matrix
CHAPTER 5. SMOOTHED PAYOFFS  ANOTHER BREAKDOWN 113
A. As an example, if we have a linear homogeneous system of the form (5.19) with
the eigenvalues of A given by ω
1,2
= α
1,2
+iβ
1,2
, then the general solution is given by
_
_
_
u(x) = u
0
e
α
1
x
(cos β
1
x +i sin β
1
x),
v(x) = v
0
e
α
2
x
(cos β
2
x +i sin β
2
x).
Three qualitatively diﬀerent classes of ﬁxed points can be identiﬁed, namely nodes,
spirals and saddle points, which are entirely determined by the values of α
1,2
and β
1,2
.
Nodes If β
1,2
= 0, i.e. the eigenvalues are real and if both α
1,2
have the same sign
then we have a nodal ﬁxed point. Furthermore if α
1,2
< 0 we have an stable
node (sink) since in the limit x → ∞ the solution will tend to the ﬁxed point
and if α
1,2
> 0 then we have a unstable node (source) since the solution will
move away from the ﬁxed point as x →∞.
Saddle Points If again β
1,2
= 0, i.e. real eigenvalues and α
1
and α
2
have opposite
sign then such a situation corresponds to a saddle point. A saddle point is stable
along one direction (corresponding to the eigenvector of the negative eigenvalue)
and unstable along another (corresponding to the positive eigenvalue)
Spirals If β
1,2
,= 0 we have a complex conjugate pair of eigenvalues, i.e. α
1
= α
2
= α
and β
1
= −β
2
= β. In this case we have a spiral ﬁxed point, socalled because,
due to the periodic functions (sin and cos) in the solution, the solutions will
spiral into or out of these ﬁxed points. Furthermore if α < 0 then we have a
stable spiral (sink) and if α > 0 at unstable spiral (source).
In addition if β
1,2
,= 0 but α
1,2
= 0 then the solutions become periodic but the
trajectories are closed, in this case the ﬁxed point is called a centre.
For a nonlinear system, however, the structure of the phase portrait is not obvious,
since there can be multiple ﬁxed points which may interact. However it can be
shown
3
that provided the system is structurally stable, then matters are nearly as
simple as in the linear case outlined above. When considering a nonlinear system,
3
See for example Peixoto (1997).
CHAPTER 5. SMOOTHED PAYOFFS  ANOTHER BREAKDOWN 114
multiple singularities can exist (but only a ﬁnite number), the combination of which
entirely determines the qualitative behaviour of the ODE and its solutions. Moreover
each singularity (ﬁxed point) is of the same elementary type as for linear systems
(e.g. nodes, spirals and saddles) and in this case linear stability implies nonlinear
stability (provided none of the eigenvalues have zero real parts, i.e. a centre).
For nonlinear systems, however, there is the additional possibility of socalled limit
cycles in the phase portrait. These are ﬁxed orbits that attract (or repel) nearby
paths and correspond to ﬁxed oscillatory solutions of the ODE. Furthermore, where
the nonlinear system is dependent on some parameter, the solution could undergo bi
furcations at critical values of the parameter, where the number of solutions increases
or decrease. Fortunately, however, the nonlinear system (5.16) does not exhibit such
nonlinear behaviour and so this will not be discussed further.
Consider the nonlinear system (5.16), the observant reader may have noticed that
this equation is not of the autonomous type (since the independent variable η appear
explicitly in the equation) and so will not have a two dimensional phase portrait.
However it turns out that we can produce an autonomous system by making an
appropriate transformation, which will be outlined in the following subsection.
5.1.4 Deriving an autonomous system
Equation (5.16) can be made autonomous. Firstly we make the variable transforma
tion
η = ±e
x
. (5.20)
Immediately we can see that the choice of the positive sign corresponds to a mapping
from x ∈ (−∞, ∞) to η ∈ (0, ∞) and the negative sign corresponds to the negative
semiinﬁnite plane in η. Hence we are eﬀectively making two separate transformations
on two diﬀerent Riemann surfaces. Substituting (5.20) into (5.16) yields (for both
transformations)
(V
xx
−V
x
)
2
(3V −V
x
) = κe
4x
,
CHAPTER 5. SMOOTHED PAYOFFS  ANOTHER BREAKDOWN 115
and further making the transformation
V = e
µx
u(x),
gives (after substitution)
_
µ(µ −1)u + (2µ −1)u
x
+u
xx
_
2
_
(3 −µ)u −u
x
_
e
3µx
= κe
4x
,
from which it is clear that the equation will become autonomous if we set µ =
4
3
,
resulting in
_
4
9
u +
5
3
u
x
+u
xx
_
2
_
5
3
u −u
x
_
= κ,
where x = ln η and u = e
−
4x
3
V . This can be rearranged to make u
xx
the argument,
i.e.
u
xx
= −
4
9
u −
5
3
u
x
±
_
κ
5
3
u −u
x
.
As outlined in the previous section, this second order ODE can be expressed as a
system of coupled ﬁrst order equations. Setting v = u
x
we arrive at the coupled
system
_
_
u
v
_
_
x
=
_
_
_
v
−
4
9
u −
5
3
v ±
_
κ
5
3
u−v
_
_
_
, (5.21)
from which we can eliminate the independent variable x to produce the equation for
the ﬁeld lines of the phase portrait, i.e.
dv
du
= −
4u
9v
−
5
3
±
1
v
_
κ
5
3
u −v
. (5.22)
Consistent with standard phaseplane theory, the behaviour of this nonlinear system
is entirely determined by the location and behaviour of its ﬁxed points, i.e. the
points where u
x
= u
xx
= 0; for further details see for example Peixoto (1997). These
points correspond to the singular points of (5.22), where the ﬁeld direction cannot
be determined. The nature and stability of the ﬁxed point can be determined by
investigating the linearised system in the vicinity of the ﬁxed point. The next section
outlines the linearisation procedure and the classiﬁcation of the ﬁxed points in more
detail. However before proceeding we note that a further simpliﬁcation can be made
CHAPTER 5. SMOOTHED PAYOFFS  ANOTHER BREAKDOWN 116
to the system (5.21), the parameter κ can be scaled out of the equation entirely by
another appropriate transformation. If we seek a solution of the form
u = κ
α
1
ˆ u, (5.23a)
v = κ
α
2
ˆ v, (5.23b)
then it can be shown that the equation in (ˆ u, ˆ v)space will be independent of the
parameter κ, if we choose
α
1
= α
2
=
1
3
.
Hence for all intents and purposes we can set κ = 1 in the original system (5.21) and
we recover the correctly scaled solution via the transformation (5.23).
Now that we have our autonomous system in the simplest possible form we can begin
to investigate the structure of the phase plane. Recall that the phase plane’s structure
is entirely determined by the location and nature of its ﬁxed points, hence the aim of
the next section is to ﬁnd and classify the ﬁxed points of the nonlinear system (5.21).
5.1.5 Behaviour of the ﬁxed points
We wish to ﬁnd and classify the ﬁxed points of the system
u
x
= f(u, v) = v,
v
x
= g(u, v) = −
4
9
u −
5
3
v +
1
_
5
3
u −v
,
(5.24)
where we have dropped the hats for simplicity of notation. The ﬁxed points are
deﬁned by f(u
0
, v
0
) = g(u
0
, v
0
) = 0 and so it can be seen from (5.24) that the ﬁxed
points correspond to v
0
= 0 with u
0
the solution of the following equation
4
u
3
0
=
243
80
.
4
Note that if we were to take the negative squareroot of the equation then the ﬁxed point equation
becomes
−
4u
9
−
3
5u
= 0,
which only has two solutions, both of which are complex, i.e.
u =
243
80
1
3
e
±
2iπ
3
.
Hence the positive root seems the correct choice.
CHAPTER 5. SMOOTHED PAYOFFS  ANOTHER BREAKDOWN 117
Obviously this has three roots, one real and a complex conjugate pair. They are
u
01
=
_
243
80
_1
3
, (5.25a)
u
02
=
_
243
80
_1
3
e
2iπ
3
, (5.25b)
u
03
=
_
243
80
_1
3
e
−
2iπ
3
. (5.25c)
Interestingly, the real part of the ﬁxed points above in (u, v)space corresponds to the
exact solution of (5.12) in (V, η)space, namely (5.17); however recall that this solu
tion does not satisfy the required boundary conditions (5.14). In order to determine
the nature of these ﬁxed points, we need to undertake analysis in the local neigh
bourhood of the ﬁxed point, and this can be performed by linearising the nonlinear
equation around these points, by setting
u = u
0
+¯ u, (5.26a)
v = v
0
+¯ v, (5.26b)
where is a small parameter (corresponding to a small perturbation) and performing
a Taylor series expansion on the functions f(u, v) and g(u, v) about the ﬁxed points
(u
0
, v
0
) which gives
f(u, v) = f(u
0
+¯ u, v
0
+¯ v) = f(u
0
, v
0
) +¯ u
∂f
∂u
(u
0
, v
0
) +¯ v
∂f
∂v
(u
0
, v
0
) +o(),
g(u, v) = g(u
0
+¯ u, v
0
+¯ v) = g(u
0
, v
0
) +¯ u
∂g
∂u
(u
0
, v
0
) +¯ v
∂g
∂v
(u
0
, v
0
) +o().
Hence the system becomes
_
_
u
0
+¯ u
v
0
+¯ v
_
_
x
=
_
_
f(u
0
, v
0
) +¯ u
∂f
∂u
(u
0
, v
0
) +¯ v
∂f
∂v
(u
0
, v
0
) +o()
g(u
0
, v
0
) +¯ u
∂g
∂u
(u
0
, v
0
) +¯ v
∂g
∂v
(u
0
, v
0
) +o()
_
_
,
_
_
u
0
v
0
_
_
x
+
_
_
¯ u
¯ v
_
_
x
=
_
_
f(u
0
, v
0
)
g(u
0
, v
0
)
_
_
+
_
_
∂f
∂u
(u
0
, v
0
)
∂f
∂v
(u
0
, v
0
)
∂g
∂u
(u
0
, v
0
)
∂g
∂v
(u
0
, v
0
)
_
_
_
_
¯ u
¯ v
_
_
+o().
The ﬁrst term in the above equation disappears as it is a derivative of a constant
(ﬁxed point). In addition the other O(1) term also disappears as it is the functions
CHAPTER 5. SMOOTHED PAYOFFS  ANOTHER BREAKDOWN 118
f and g evaluated at the ﬁxed points which by deﬁnition is equal to zero. This leads
to the following linear system for (¯ u, ¯ v),
_
_
¯ u
¯ v
_
_
x
=
_
_
∂f
∂u
(u
0
, v
0
)
∂f
∂v
(u
0
, v
0
)
∂g
∂u
(u
0
, v
0
)
∂g
∂v
(u
0
, v
0
)
_
_
_
_
¯ u
¯ v
_
_
.
The linearisation performed here does not always work however; the long term be
haviour of the linearised system near a ﬁxed point can diﬀer qualitatively from the
long term behaviour near a ﬁxed point of the fully nonlinear system. Fortunately
however there are only two situations where this can occur. One is when the ﬁxed
point of the linearised system is a centre and the other when the linearised system
has zero as an eigenvalue. In all other cases the local picture of the nonlinear system
near a ﬁxed point looks like its linearisation. For the nonlinear system (5.24) we have
∂f
∂u
= 0,
∂f
∂v
= 1,
∂g
∂u
= −
4
9
−
5
6
_
5
3
u −v
_
−
3
2
,
∂g
∂v
= −
5
3
+
1
2
_
5
3
u −v
_
−
3
2
.
We now evaluate these derivatives at each ﬁxed point in turn. Considering ﬁrst u
01
,
at this point, the system becomes
_
_
¯ u
¯ v
_
_
x
=
_
_
0 1
−
2
3
−
23
15
_
_
_
_
¯ u
¯ v
_
_
.
The behaviour of this linear system is determined by its eigenvalues, which are de
termined by the solution to the following characteristic equation
det
_
_
0 −ω 1
−
2
3
−
23
15
−ω
_
_
= 0,
which has two (complex) solutions
ω
1,2
=
1
30
_
−23 ±i
√
71
_
. (5.28)
CHAPTER 5. SMOOTHED PAYOFFS  ANOTHER BREAKDOWN 119
Systems with complex eigenvalues correspond to spiral node ﬁxed points and since
the real part of the eigenvalue is negative (see section 5.1.3) we have a stable ﬁxed
point, usually called a spiral sink.
Next consider the ﬁxed point u
02
, at this point the local linear system corresponds to
_
_
¯ u
¯ v
_
_
x
=
_
_
0 1
−
2
9
−
9
5
_
_
_
_
¯ u
¯ v
_
_
which has two real, negative eigenvalues
ω
1
= −
2
15
, (5.29a)
ω
2
= −
5
3
. (5.29b)
It also turns out that the ﬁnal ﬁxed point u
03
has the same linearised system as above
and so the same eigenvalues. Recall that two negative real eigenvalues correspond
to a stable node. However since these ﬁxed points are in the complex domain we are
strictly required to perform a full (fourdimensional) complex stability analysis about
these ﬁxed points (i.e. in the complex domain) to fully determine their behaviour.
However since we are only really interested in real solutions to the nonlinear system
(5.24) the two complex ﬁxed points can be omitted from our analysis. To conclude
we have determined that the (real) ﬁxed point is a (stable) spiral sink.
Figure 5.1 shows the phase portrait of the the autonomous system (5.24) (with κ
scaled out of the problem completely). The results were obtained using MATLAB and
the pplane ODE software package
5
which employed the DormandPrince modiﬁcation
to the standard RungeKutta shooting technique, (cf. Dormand and Prince, 1980).
The important point to note is that in the absence of any other ﬁxed points, every
path (each corresponding to a diﬀerent boundary condition) passes through this ﬁxed
point. What this implies for the solution of (5.12) subject to the boundary condition
(5.14) is that it too must pass through this ﬁxed point, and so will be unable to
satisfy the boundary conditions as η → ∞ and η → −∞. Further, assume that we
5
Copyright John C. Polking. For more information see http://math.rice.edu/∼dfield/.
CHAPTER 5. SMOOTHED PAYOFFS  ANOTHER BREAKDOWN 120
8
6
4
2
0
2
4
6
8
0 2 4 6 8 10
P
S
f
r
a
g
r
e
p
l
a
c
e
m
e
n
t
s
u
v
Complex Region
Figure 5.1: Phase portrait of the autonomous system (5.24). Note the ﬁxed point at
u =
_
243
80
_1
3
, v = 0 and the ﬁeld direction lines. The dotted line represents an analytic
envelope for the phase portrait close to the singular line v =
5u
3
, cf. equation (5.31).
have obtained the solution for η ∈ (−∞, 0) and ‘jumped’ to the current phase plane.
Solving from η = 0 corresponds to shooting in the current phase plane (Riemann
surface) from x = −∞. This path (like all paths) must pass through the ﬁxed point
and by deﬁnition it must remain there for all x as x →∞, corresponding to η →∞.
However, since the solution at the ﬁxed point is ﬁxed as η → ∞, and in fact takes
the form V ∼ η
4
3
(i.e. (5.17)), we have no hope of satisfying the boundary condition
(5.14) as η →∞.
Before we can conclude that no smooth solution to the ODE (5.12) exists (and hence
no smooth inner solution about the point S
0
) satisfying the boundary condition (5.14),
we must ﬁrst check that all ﬁxed points of the system have been found and indeed
that every path must pass through the ﬁxed point we have previously found. The
following subsection investigates the structure of the phase plane in yet more detail.
5.1.6 Structure of the phase portrait
Firstly it is clear that the equation (5.24) only has real solutions for
5u
3
−v > 0,
CHAPTER 5. SMOOTHED PAYOFFS  ANOTHER BREAKDOWN 121
hence the line v =
5u
3
is a limiting point for imaginary solutions and is in fact a
singular line, which corresponds to solutions of the form V ∼ η
3
in the original
variables. Similarly solutions of the form V ∼ η
4
3
correspond to the line v = 0 on the
phase plane and solutions of the form V ∼ const. correspond to the line v = −
4u
3
on
the phase plane.
The singular line thus corresponds to where we must apply the boundary conditions
of the ODE (5.12) and so it should be clear that we have no chance of applying the
boundary condition on this singular line. However the extra term in the matching
condition (5.14) means that the boundary condition is to be applied slightly below
this singular line. To see this recall that the boundary behaviour (5.14) for large η
has the form
V = A
0
η
3
+
A
1
η
2
+. . .
where A
0
and A
1
are known constants. We wish to determine whereabouts on the
phase plane this condition corresponds to. We ﬁrst transform variables to give
u = A
0
e
5x
3
+A
1
e
−
10x
3
,
and so
v = u
x
=
5A
0
3
e
5x
3
+
10A
1
3
e
−
10x
3
. (5.30)
Clearly this cannot be expressed explicitly in terms of just u and v, but we can make
an approximation for large x that
x =
3
5
ln
_
u
A
0
_
,
which after substituting into (5.30) gives
v =
5u
3
−
10A
1
A
2
0
3u
2
.
Recalling that A
0
=
V
3
6
and A
1
=
1
5V
2
3
yields
v =
5u
3
−
1
54u
2
. (5.31)
Hence this is just below the singular line, and so there is a chance that shooting
methods will work here.
CHAPTER 5. SMOOTHED PAYOFFS  ANOTHER BREAKDOWN 122
It can also be seen from ﬁgure 5.1 that the ﬁeld lines appear to be approaching v = 0
in a square root fashion, for example v ∼ (u −u
1
)
1
2
, where u
1
is the v = 0 intercept.
It can be shown that this is indeed the correct behaviour by seeking a solution of the
form
v = C
2
(u −u
1
)
α
2
to give
α
2
C
2
2
(u −u
1
)
2α
2
−1
= −
4
9
u −
5
3
C
2
(u −u
1
)
α
2
+
1
_
5
3
u −C
2
(u −u
1
)
α
2
.
We are interested in the limit u → u
1
, here we have that u
1
¸ (u −u
1
)
α
2
provided
α
2
> 0. This gives
α
2
C
2
2
(u −u
1
)
2α
2
−1
≈ −
4
9
u
1
+
_
3
5u
1
,
and so we must have
2α
2
−1 = 0 ⇒α
2
=
1
2
,
which also gives the value of C
2
to be
C
2
= ±
_
−
8
9
u
1
+
_
12
5u
1
_
1
2
. (5.32)
The constant C
2
found in equation (5.32) only has real solutions provided
u
1
<
_
243
80
_1
3
= u
0
,
the location of the ﬁxed point. If we are interested in calculating the behaviour to
the right of the ﬁxed point then a solution of the form
v = C
2
(u
1
−u)
α
2
should be used.
5.1.7 Other ﬁxed points
Recall that in phase plane analysis, once we have determined the location and be
haviour of all the ﬁxed points of the system we have entirely determined the qualita
tive behaviour of the solution. However, there may still be ﬁxed points which we have
CHAPTER 5. SMOOTHED PAYOFFS  ANOTHER BREAKDOWN 123
still not considered, namely those at [u[ = [v[ = ∞. To investigate any ﬁxed points at
inﬁnity it is convenient to transform the problem into plane polar coordinates. This
is done via the transformation
u = r cos θ,
v = r sin θ,
and diﬀerentiation gives
r
x
=
uu
x
+vv
x
r
,
θ
x
=
uv
x
−vu
x
r
2
.
Substitution of equation (5.24) into the above yields
r
x
=
5
9
r cos θ sin θ −
5
3
r sin
2
θ +
sin θ
r
1
2
_
5
3
cos θ −sin θ
_1
2
, (5.33a)
θ
x
= −
4
9
cos
2
θ −
5
3
cos θ sin θ +
cos θ
r
3
2
_
5
3
cos θ −sin θ
_1
2
−sin
2
θ. (5.33b)
Next in order to investigate the behaviour at inﬁnity we make the transformation
6
ρ =
1
r
,
φ = −θ,
and diﬀerentiating gives
ρ
x
= −ρ
2
r
x
,
φ
x
= −θ
x
.
This leads to the following system in (ρ, φ)space
ρ
x
= f(ρ, φ) =
5
9
ρ cos φsin φ +
5
3
ρ sin
2
φ +
ρ
5
2
sin φ
_
5
3
cos φ + sin φ
_1
2
, (5.34a)
φ
x
= g(ρ, φ) =
4
9
cos
2
φ −
5
3
cos φsinφ −
ρ
3
2
cos φ
_
5
3
cos φ + sin φ
_1
2
+ sin
2
φ. (5.34b)
6
Note that this corresponds to the transformation ˆ z = z
−1
in complex space z = re
iθ
.
CHAPTER 5. SMOOTHED PAYOFFS  ANOTHER BREAKDOWN 124
Clearly the only ﬁxed point of this system is when ρ = 0 and when φ satisﬁes the
following equation
φ
x
=
4
9
cos
2
φ −
5
3
cos φsin φ + sin
2
φ = 0. (5.35)
Equation (5.35) can be factorised as follows
_
4
3
cos φ −sin φ
__
1
3
cos φ −sin φ
_
= 0.
Hence this has solutions when either tan φ =
4
3
or tan φ =
1
3
which have inﬁnitely
many solutions, however we are only interested in the principle branch when 0 ≤ φ ≤
2π, and so we have only two solutions. Transforming these ﬁxed points back to the
original (u, v) coordinates it can be shown that these two ﬁxed points corresponds to
the point at inﬁnity along the lines u = −
4
3
v and u = −
1
3
v.
All that remains is to determine the nature of these ﬁxed points. For simplicity we
shall remain in the transformed space (ρ, φ). The nature of the ﬁxed points in this
space will remain unchanged under the transformation back to the original coordinate
system. Again to investigate the nature of the ﬁxed point it is required to perform a
linearisation about the ﬁxed points. Doing so leads to the following linear system in
(¯ ρ,
¯
φ)
_
_
¯ ρ
¯
φ
_
_
x
=
_
_
∂f
∂ρ
(ρ
0
, φ
0
)
∂f
∂φ
(ρ
0
, φ
0
)
∂g
∂ρ
(ρ
0
, φ
0
)
∂g
∂φ
(ρ
0
, φ
0
)
_
_
_
_
¯ ρ
¯
φ
_
_
.
Evaluating the partial derivatives of our system yields
∂f
∂ρ
=
5
9
cos φsin φ +
5
3
sin
2
φ +
5ρ
3
2
sin φ
2
_
5
3
cos φ + sin φ
_1
2
,
∂f
∂φ
=
10ρ
3
cos φsin φ +
5ρ
9
_
cos
2
φ −sin
2
φ
_
−
ρ
5
2
_
1 + sin
2
φ +
5
3
cos φsin φ
_
2
_
5
3
cos φ + sin φ
_3
2
,
∂g
∂ρ
= −
3ρ
1
2
cos φ
2
_
5
3
cos φ + sin φ
_1
2
,
∂g
∂φ
=
10
9
cos φsin φ −
5
3
_
cos
2
φ −sin
2
φ
_
+
ρ
3
2
_
1 + sin
2
φ +
5
3
cos φsin φ
_
2
_
5
3
cos φ + sin φ
_3
2
.
CHAPTER 5. SMOOTHED PAYOFFS  ANOTHER BREAKDOWN 125
First we will consider the ﬁxed point at (ρ, φ) = (0, arctan
4
3
) ≈ (0, 0.9273). Here the
linearised system becomes
_
_
¯ ρ
¯
φ
_
_
x
=
_
_
4
3
0
0 1
_
_
_
_
¯ ρ
¯
φ
_
_
which, since the matrix is diagonal clearly has eigenvalues
ω
1
=
4
3
, (5.36a)
ω
2
= 1, (5.36b)
with eigenvectors along the ρdirection and the φdirection. These eigenvalues are
both real and positive which corresponds to a nodal source ﬁxed point, which is
unstable.
Finally considering the ﬁxed point at (ρ, φ) = (0, arctan
1
3
) ≈ (0, 0.3218) leads to the
following linearised system
_
_
¯ ρ
¯
φ
_
_
x
=
_
_
1
3
0
0 −1
_
_
_
_
¯ ρ
¯
φ
_
_
which has eigenvalues
ω
1
=
1
3
, (5.37a)
ω
2
= −1. (5.37b)
Here we have real eigenvalues, but of opposite sign, corresponding to a saddle node,
which has a stable direction (here corresponding to the φdirection) and an unstable
direction (the ρdirection). In other words any perturbation in the ρdirection will
result in the solutions being pushed away from this ﬁxed point.
One ﬁnal point of interest is that the analysis of the ﬁxed points at inﬁnity has
revealed a path which does not terminate at the ﬁxed point near the origin. We can
move from the nodal source at inﬁnity along the φ direction and arrive at the saddle
node (also at inﬁnity) provided there is no movement in the radial direction. This path
however is not realistic in the context of numerical solutions of the original ODE, as
CHAPTER 5. SMOOTHED PAYOFFS  ANOTHER BREAKDOWN 126
any solution method would introduce small numerical perturbations and the solution
would always terminate at the spiral sink near the origin. Hence we have shown, via
phase plane analysis, that every numerically simulated path in the phase space of the
nonlinear system under consideration, corresponding to solutions of the second order
ODE with any given boundary condition (at a point) will always terminate at the
ﬁxed node near the origin. Hence any solution of a numerical shooting method given
any condition at any boundary will asymptote to the ﬁxed point solution u =
_
243
80
_1
3
,
which corresponds to the solution
V =
_
243κ
80
_1
3
η
4
3
in the original variables. Hence the ﬁxed point in u corresponds to the exact solution
of the original equation.
In this section we have used phaseplane analysis to show that it does not appear
possible to resolve the singular behaviour of equation (5.12), even using smallscale
analysis, suggesting that, despite applying a smoothed payoﬀ proﬁle, nonsmoothness
has been ‘induced’ into the solution for τ > 0. The corollary to this is, therefore, that
there is insuﬃcient ﬁnancial modelling in (4.1), for standard option payoﬀ proﬁles,
to prevent such behaviour, indicating (another) failure in the underlying modelling.
Finally, it should be pointed out that strictly the parameter values taken in ﬁgure
4.7 are in the range ρ <
λ
2
, as discussed in the present chapter. However, there is
a further subtlety as ρ → 0 (which we do not explore), insofar as in this limit yet
further asymptotic analysis is applicable, involving another small parameter, namely
ρ itself. Note too that as ρ → 0, the two values of S
0
will coincide and in this limit
the problems associated with the vanishing of the denominator are in some ways
mediated by the problems of the inﬁnite gamma. Figure 4.7 is still useful, however,
in guiding the asymptotics described earlier.
Chapter 6
Perpetual Options
Explicit solutions to parabolic freeboundary problems are rare. The situation is quite
diﬀerent, however, if we consider perpetual American options. For these options the
dependence on time, or rather on time left to maturity, is removed, so the partial
diﬀerential equation is reduced to an ordinary diﬀerential equation. This chapter
investigates such perpetual options in the context of the nonlinear models described
in chapter 2.
Although section 4.7 has demonstrated that the fullfeedback model with early exer
cise leads to what amounts to a trivial problem for puts, the question that naturally
arises (given the results of the previous chapter) is what of other payoﬀ conditions,
in particular those which do not have discontinuous deltas (and assuming the dif
ﬁculties raised in chapter 5 can be bypassed). The next set of results (obtained
using a straightforward PSOR scheme), shown in ﬁgure 6.1, correspond to a calcu
lation obtained taking the smoothed payoﬀ condition (5.2). This set of results (for
an Americanstyle put option) corresponds to the payoﬀ condition with ρ = 0.15
(together with K = 1, r = 0.04, σ = 0.2, λ = 0.25), in this parameter regime the
denominator does not vanish since λ < 2ρ (cf. (5.4)). To be consistent with the ﬁnal
payoﬀ conditions, the earlyexercise condition was imposed by taking
V = max
_
1
2
_
K −S +
_
(K −S)
2
+ρ
2
_
, V
PDE
_
,
127
CHAPTER 6. PERPETUAL OPTIONS 128
at all S and τ at each iteration of the PSOR algorithm, where V
PDE
is the solution to
(4.1). The computation was permitted to continue until a near steady state had been
attained (i.e. the asymptote to a perpetual valuation). Figure 6.1 clearly indicates
that the computation could be extended, unabated, for long maturities. This does
emphasise, of course, that much of the diﬃculty reported above with standard payoﬀ
conditions is associated with the vanishing of the denominator in (4.1), which will
certainly be the case for standard payoﬀ functions on account of the discontinuous
deltas.
0
0.2
0.4
0.6
0.8
1
1.2
0 0.5 1 1.5 2 2.5 3
PSfrag replacements
τ = 0
τ = 10
S
V
Figure 6.1: Full feedback American put, K = 1, r = 0.04, σ = 0.2, λ = 0.25, ρ = 0.15
(smoothed payoﬀ), τ = 0, 1, . . . , 10. Note that we are in the regime λ < 2ρ and so
we should expect no singular behaviour.
There is a subtlety with the application of the smoothed payoﬀ proﬁle (5.2) to per
petual options that should be noted and that we shall attempt to outline below. If
we apply the smoothed payoﬀ proﬁle to the standard BlackScholes equation, (3.2),
and evaluate V
τ
at ﬁnal maturity (directly from the PDE) then this will provide
us with an indication of whether there will exist any earlyexercise regions or not,
more speciﬁcally if V
τ
ever changes sign. For a smoothed put numerical investigation
shows that V
τ
at expiry can be both positive and negative, indicating that there is
an earlyexercise region. However for ρ large enough then this is not the case and
V
τ
always remains negative and thus the option will always be exercised immediately
CHAPTER 6. PERPETUAL OPTIONS 129
(at t = 0). Hence for the BlackScholes equation the earlyexercise boundary exists
only for
0 ≤ ρ < ρ
max
.
Numerical investigations reveal that the same behaviour is also seen for the nonlinear
equation (4.1) where we can now identify four regimes
I. ρ = 0,
II. 0 < ρ ≤
λ
2
,
III.
λ
2
< ρ < ρ
max
,
IV. ρ > ρ
max
.
For regime I, i.e. kinked payoﬀ proﬁles, we have demonstrated in chapter 4, using
a local expansion about τ = 0, that American options are always early exercised.
Regime IV is of little interest (since we would never optimally exercise the option) and
also ﬁgure 6.1 indicates that in regime III there exists a wellposed American option
problem. However, the behaviour of the solution to the American option problem
in regime II still remains unclear, since the (CrankNicolson) ﬁnite diﬀerence scheme
successfully employed to the system in regime III proved unsuccessful in regime II.
We shall not investigate this regime further but recall however, that in chapter 5 it
was shown that we should expect nonsmooth behaviour for the European option in
this regime and hence it is likely that the American counterpart will exhibit similar
solution diﬃculties.
Given that longterm solutions to (4.1) (with early exercise) can exist under certain
parameter regimes, it is of some interest to investigate the behaviour of this system
with the time variation omitted, i.e.
1
2
σ
2
S
2
V
SS
(1 −λV
SS
)
2
+rSV
S
−rV = 0, (6.1)
subject to (the standard earlyexercise put conditions) V → 0 as S → ∞, and
V = K −S,
dV
dS
= −1 on the free boundary S = S
f
. This system was solved using a
straightforward RungeKutta algorithm, which performed an iteration procedure to
CHAPTER 6. PERPETUAL OPTIONS 130
evaluate S
f
. Results, based on (6.1) are shown in ﬁgure 6.2 for a range of values of the
parameter λ, with K = 1, σ = 0.2, r = 0.04. The location of the free boundary is also
clearly marked, and thus reveals yet another interesting feature, namely the approach
of the free boundary towards S = 0. For λ 1.1, for the choice of parameters taken
above, it would appear that it is never optimal to early exercise the perpetual option
(the free boundary reaches S = 0 at λ ≈ 1.1).
As a ﬁnal cautionary note on the numerical solution of the nonlinear ODE (6.1), it
was observed that multiple solution branches could be found using certain numerical
techniques, such as ﬁnite diﬀerence methods and the socalled bodyﬁtted coordinate
system (described further in section 9.3). These solutions exhibited nonsmooth be
haviour and were thought to be a possible steadystate solution of the time dependent
PDE (4.1). However these solution branches appear to be merely a numerical artifact
of the equation (and the solution technique) as increasing the resolution of the grid
saw these solution branches collapse down onto the ‘stable’ branch, corresponding to
the smooth solutions shown in ﬁgure 6.2.
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
PSfrag replacements
λ = 0
λ = 1.1
S
V
Figure 6.2: Perpetual fullfeedback American put, K = 1, r = 0.04, σ = 0.2, λ =
0, 0.1, 0.2, . . . , 1.1; freeboundary location as indicated.
CHAPTER 6. PERPETUAL OPTIONS 131
6.1 Analytic solutions and perturbation methods
It is well known that the BlackScholes perpetual American put option admits an
exact analytical solution
V
P
(S) = (K −S
f
)
_
S
S
f
_
−α
(6.2)
where α =
2r
σ
2
> 0 and S
f
is the (ﬁxed) location of the free boundary determined by
S
f
=
αK
α + 1
. (6.3)
Furthermore, as stated in section 1.3.5 the American call option value is coincident
with the European call option value and so there exists no (optimal) earlyexercise
region for the American call. Hence the value of a perpetual American call option
will be trivially equal to the current value of the stock. If, however, we include the
payment of a constant dividend yield in the underlying then the optimal exercise
boundary becomes nontrivial and thus its perpetual counterpart will have a non
zero value. In fact the value of the perpetual American call option on an underlying
paying a constant dividend yield D is given by
V
C
(S) = (S
f
−K)
_
S
S
f
_
β
, (6.4)
where β > 0 is given by
β =
1
σ
2
_
_
−
_
r −D −
1
2
σ
2
_
+
¸
_
r −D −
1
2
σ
2
_
2
+ 2rσ
2
_
_
,
and the free boundary S
f
given by
S
f
=
βK
β −1
.
Interestingly we can see that as D → 0, we have β → 1 and so the free boundary
for the American call tends to inﬁnity and the solution reduces to the trivial solution
V
C
(S) ≡ S.
Unfortunately the nonlinear ODE (6.1) has no analytical solution. As an alternative
to resorting to fully numerical solutions we can utilise perturbation methods to obtain
CHAPTER 6. PERPETUAL OPTIONS 132
an approximation of the solution for small values of the parameter λ. It is useful to
rewrite (6.1) as
1
2
σ
2
S
2
V
SS
+ (rSV
S
−rV ) (1 −λV
SS
)
2
= 0 (6.5)
and we proceed by trying a regular expansion of V in powers of λ, i.e.
V = V
0
+λV
1
+λ
2
V
2
+. . . .
Substituting the above into equation (6.5) and collecting together the powers of λ
gives the following asymptotic set of equations (cf. sections 1.8 and 4.4)
O(λ
0
) :
1
2
σ
2
S
2
V
0SS
+rSV
0S
−rV
0
= 0, (6.6a)
O(λ
1
) :
1
2
σ
2
S
2
V
1SS
+rSV
1S
−rV
1
= 2V
0SS
(rSV
0S
−rV
0
) , (6.6b)
O(λ
2
) :
1
2
σ
2
S
2
V
2SS
+rSV
2S
−rV
2
= 2V
0SS
(rSV
1S
−rV
1
)
−
_
V
2
0SS
−2V
1SS
_
(rSV
0S
−rV
0
) . (6.6c)
The solution to the leading order equation (6.6a) is simply the solution to the Black
Scholes perpetual option, (6.2), or more generally
V
0
(S) = AS +BS
−α
,
with α as previously deﬁned and constants A and B are to be determined from the
appropriate boundary conditions. Note that we are required to use this solution in
order to solve the next order equation (6.6b), now a nonhomogeneous BlackScholes
equation, which can also be solved analytically. Hence equation (6.6b) becomes
1
2
σ
2
S
2
V
1SS
+rSV
1S
−rV
1
= −2rα(α + 1)
2
B
2
S
−2α−2
,
⇒S
2
V
1SS
+αSV
1S
−αV
1
= −2α
2
(α + 1)
2
B
2
S
−2α−2
. (6.7)
The general solution to this equation is thus
V
1
(S) = CS +DS
−α
,
where again the constants C and D are to be determined from the boundary condi
tions. To deal with the nonhomogeneity we seek a particular solution of the form
V
1
(S) = kS
−2α−2
, where k is to be determined. Substitution into (6.7) gives
(2α + 2) (2α + 3) k −α(2α + 2) k −αk = −2α
2
(α + 1)
2
B
2
,
CHAPTER 6. PERPETUAL OPTIONS 133
which can be solved for k to yield
k = −
2α
2
(α + 1)
2
B
2
(2α + 3) (α + 2)
.
Hence the solution of the ﬁrst order correction is
V
1
(S) = CS +DS
−α
−
2α
2
(α + 1)
2
B
2
(2α + 3) (α + 2)
S
−2α−2
.
The constant B is found from the boundary conditions on the leading order equation
V
0
and C and D are found from the boundary conditions on the ﬁrst order correction
V
1
; these shall now be determined for the case of a (nondividend paying) perpetual
American put option. In this case the boundary conditions are given by
V (S) →0 as S →∞, (6.8a)
V (S
f
) = K −S
f
, (6.8b)
V
S
(S
f
) = −1. (6.8c)
First we must also apply an asymptotic expansion to the location of the free boundary
S
f
, namely
S
f
= S
f
0
+λS
f
1
+. . . .
Along with the perturbation in V (S) the boundary condition (6.8b) thus becomes
V
0
(S
f
0
+λS
f
1
+. . .) +λV
1
(S
f
0
+λS
f
1
+. . .) +. . . = K −S
f
0
−λS
f
1
−. . . .
Exploiting the smallness of λ and recalling Taylor’s theorem we have
V
0
(S
f
0
+λS
f
1
+. . .) = V
0
(S
f
0
) +λS
f
1
V
0S
(S
f
0
) +O(λ
2
),
V
1
(S
f
0
+λS
f
1
+. . .) = V
1
(S
f
0
) +λS
f
1
V
1S
(S
f
0
) +O(λ
2
),
hence equating powers of λ we can see that this boundary condition becomes
O(λ
0
) : V
0
(S
f
0
) = K −S
f
0
, (6.9a)
O(λ
1
) : V
1
(S
f
0
) = −S
f
1
_
1 +V
0S
(S
f
0
)
_
. (6.9b)
CHAPTER 6. PERPETUAL OPTIONS 134
A similar application of Taylor’s theorem to the smooth pasting condition (6.8c)
yields
O(λ
0
) : V
0S
(S
f
0
) = −1, (6.10a)
O(λ
1
) : V
1S
(S
f
0
) = −S
f
1
V
0SS
(S
f
0
). (6.10b)
Interestingly we can see that the smooth pasting condition (6.10a) when substituted
into (6.9b) results in the condition at the free boundary for the ﬁrstorder correction
reducing to zero. This implies that the solution of V
1
(S) can be determined without
knowledge of the correction to the free boundary, S
f
1
. However we still have the
smooth pasting condition on this correction which can be exploited to give us an
estimate of the correction to the free boundary, therefore rearranging (6.10b) gives
S
f
1
= −
V
1S
(S
f
0
)
V
0SS
(S
f
0
)
.
Bringing things together we have that the leading order system is given by
_
¸
¸
¸
¸
¸
¸
¸
_
¸
¸
¸
¸
¸
¸
¸
_
V
0
(S) = AS +BS
−α
, General solution;
V
0
(S
f
0
) = K −S
f
0
, Boundary condition 1;
V
0
(S →∞) = 0, Boundary condition 2;
V
0S
(S
f
0
) = −1, Smooth pasting.
This coincides exactly with the BlackScholes Perpetual Put whose solution was given
by (6.2), hence
A = 0,
B = (K −S
f
0
) S
α
f
0
,
S
f
0
=
αK
α + 1
.
The system for the ﬁrst order correction V
1
is
_
¸
¸
¸
¸
¸
¸
¸
_
¸
¸
¸
¸
¸
¸
¸
_
V
1
(S) = CS +DS
−α
−
2α
2
(α+1)
2
B
2
(2α+3)(α+2)
S
−2α−2
, General solution;
V
1
(S
f
0
) = 0, Boundary condition 1;
V
1
(S →∞) = 0, Boundary condition 2;
S
f
1
= −
V
1S
(S
f
0
)
V
0SS
(S
f
0
)
, Perturbed free boundary.
CHAPTER 6. PERPETUAL OPTIONS 135
Clearly from the boundary condition at inﬁnity we have C = 0 and the condition at
the free boundary leads to
D =
2(α + 1)
2
S
α
f
0
(2α + 3)(α + 2)
.
To determine the position of the perturbed free boundary involves evaluating the
derivatives of V
0
and V
1
which, after some laborious algebra, results in
S
f
1
= −
2(α + 1)
2α + 3
.
Hence our approximate solution can be written as
V (S) =
S
f
0
α
_
S
S
f
0
_
−α
+
2λ(α + 1)
2
(2α + 3)(α + 2)
_
_
S
S
f
0
_
−α
−
_
S
S
f
0
_
−2α−2
_
+O
_
λ
2
_
,
(6.11)
with the free boundary now located at
S
f
= S
f
0
−
2λ(α + 1)
2α + 3
+O
_
λ
2
_
.
Figure 6.3 shows the ﬁrst order correction term, i.e. λV
1
and compares it to the
diﬀerence of the (numerical) solution to the full problem (6.1) and the BlackScholes
value, i.e. V − V
BS
; it can be seen that there is a good agreement in the solutions.
Note that V
1
has only been calculated in the region S ∈ (S
f
0
, ∞) since it is not
entirely clear whether the approximate solution (6.11) is valid in the region S < S
f
0
.
Figure 6.4 shows the same comparison as in ﬁgure 6.3 but for increasing values of the
parameter λ, expectedly the agreement with the ‘exact’ solution worsens for larger
values of λ.
CHAPTER 6. PERPETUAL OPTIONS 136
0
0.005
0.01
0.015
0.02
0.025
0 0.5 1 1.5 2 2.5 3 3.5
P
S
f
r
a
g
r
e
p
l
a
c
e
m
e
n
t
s
S
λ
V
1
S
f
0
S
f
Figure 6.3: The ﬁrst order correction to the BlackScholes perpetual American put
option (solid line) compared to the diﬀerence of the fully numerical option value with
the BlackScholes (dotted line). K = 1, r = 0.04, σ = 0.2 and λ = 0.1.
0
0.05
0.1
0.15
0.2
0.25
0 0.5 1 1.5 2 2.5 3 3.5 4
P
S
f
r
a
g
r
e
p
l
a
c
e
m
e
n
t
s
S
λ
V
1
λ = 0.1
λ = 0.5
λ = 1.0
S
f
0
Figure 6.4: The ﬁrst order correction to the BlackScholes perpetual American put
option (solid line) compared to the diﬀerence of the fully numerical option value with
the BlackScholes (dotted line) for various values of λ. K = 1, r = 0.04, σ = 0.2 and
λ = 0.1, 0.5, 1.
Chapter 7
Other Models
We have shown that signiﬁcant diﬃculties arise when the form of the function λ(S, t)
in (4.1) is taken to be constant, the model as introduced by Sch¨ onbucher and Wilmott
(2000). There have been alternative models proposed, including those which involve
nonconstant λ(S, t) as was brieﬂy discussed in chapter 2. We now discuss some of
these, in particular by investigating their small τ behaviour, from which we will be
able to ascertain whether or not the diﬃculties outlined in the preceding chapters (for
constant λ) are also present. We will not present derivations of the models below;
the interested reader is referred to the appropriate references.
7.1 Frey (1998, 2000)
The model of Frey (1998, 2000) is the most similar to that discussed earlier, and leads
to the same PDE as (2.10) but with λ(S, τ) =
ˆ
λS where
ˆ
λ ∈ R. In fact it can be
seen that the Frey model is in some sense a more consistent model of price impact
as this form of λ(S, τ) eﬀectively models the price impact on the percentage price
change rather than the absolute price change, which is more consistent when using
geometric Brownian motion as the reference process.
The same scaling as was employed for the fullfeedback model as τ →0, namely (4.7)
can be used here. The solution turns out to be very similar, and for a put is given
137
CHAPTER 7. OTHER MODELS 138
by the equation
φ −
η
2
φ
η
−
σ
2
K
2
φ
ηη
2
_
1 −
ˆ
λKφ
ηη
_
2
+rKH(−η) = 0,
and for a call
φ −
η
2
φ
η
−
σ
2
K
2
φ
ηη
2
_
1 −
ˆ
λKφ
ηη
_
2
−rKH(η) = 0.
It is clear that this model will exhibit the same behaviour as the model discussed
previously, since the addition of S is not suﬃcient to alter the qualitative behaviour
of the solution  it merely leads to a rescaling of λ. The same argument holds in the
case of call options.
Equally, the arguments expounded earlier (in section 3.3 and chapter 5) regarding
the zero in the denominator of (2.10) remain applicable. However, one interesting
diﬀerence in the model of Frey (1998, 2000) is when we consider ﬁrstorder feedback.
When considering the location of the vanishing of the denominator similar to the
analysis in section 3.3 the denominator of the ﬁrstorder Frey (1998, 2000) model
vanishes when
1 −
λe
−
1
2
d
1
(S
∗
,τ)
2
σ
√
2πτ
= 0 (7.1)
where
d
1
(S, τ) =
log
_
S
K
_
+
_
r +
1
2
σ
2
_
τ
σ
√
τ
,
and S
∗
(τ) is the location of the singular denominator. Similar to the Sch¨ onbucher
and Wilmott (2000) model equation (7.1) can be solved explicitly to obtain
S
∗
(τ) = Ke
−(r+
1
2
σ
2
)τ
exp
_
±σ
¸
τ log
_
λ
2
2πσ
2
τ
_
_
, (7.2)
Figure 7.1 shows these locations for the same parameters as ﬁgure 3.7. Again it can
be seen that there exists no solution past a critical value of τ. This can be seen
directly from equation (7.2) since there exists no real solution when
λ
2
2πσ
2
τ
< 1,
hence when τ > τ
crit
where
τ
crit
=
λ
2
2πσ
2
.
CHAPTER 7. OTHER MODELS 139
For the parameters used in ﬁgure 3.7, i.e. λ = 0.1 and σ = 0.2, then we have
τ
crit
≈ 0.039789.
0.97
0.98
0.99
1
1.01
1.02
0 0.005 0.01 0.015 0.02 0.025 0.03 0.035 0.04 0.045
P
S
f
r
a
g
r
e
p
l
a
c
e
m
e
n
t
s
S
τ
Figure 7.1: Location of the vanishing of the denominator of the Frey (1998, 2000)
(solid line) and Sch¨ onbucher and Wilmott (2000) (dotted line) model with λ = 0.1,
K = 1, r = 0.04 and σ = 0.2.
7.2 Frey and Patie (2002)
Here an asset dependent liquidity is introduced in order to reproduce the volatility
smile with
λ(S, t) =
ˆ
λ
_
1 + (S −S
0
)
2
_
a
1
I
{S≤S
0
}
+a
2
I
{S≥S
0
}
__
, (7.3)
where I denotes the indicator function and
ˆ
λ, a
1
and a
2
are found by minimising the
squared distance of the observed price from the model. This form of the liquidity
structure incorporates so called liquidity drops, i.e. that λ(S, t) increases if the stock
price drops. However this additional structure is not seen on the small scale close to
expiry where the scaling (4.7) reduces (7.3) to
λ(S, τ) =
ˆ
λ
_
1 + τ(η −η
0
)
2
_
a
1
I
{η≤η
0
}
+a
2
I
{η≥η
0
}
__
,
which as τ → 0 reduces to
ˆ
λ, i.e. the model of Sch¨ onbucher and Wilmott (2000),
where the previous analysis is thus applicable.
CHAPTER 7. OTHER MODELS 140
7.3 Sircar and Papanicolaou (1998)
Another interesting model is that of Sircar and Papanicolaou (1998), who arrived at
the following PDE
V
τ
−
1
2
σ
2
S
2
_
1 −
ˆ
λV
S
1 −
ˆ
λV
S
−
ˆ
λSV
SS
_
2
V
SS
−rSV
S
+rV = 0, (7.4)
where again
ˆ
λ ∈ R. With a little rearranging we can see that this equates to (2.10)
with
λ(S, τ; V
S
) =
ˆ
λS
1 −
ˆ
λV
S
.
The authors state that there will be diﬃculties with the PDE when the denominator
passes through zero (as has been outlined in chapter 5 of the present study). To
circumvent this diﬃculty, Sircar and Papanicolaou set the value of the option close
to expiry to be equal to the BlackScholes analytical value. More speciﬁcally this is
done in the region 0 < τ < , where is chosen such that for τ > the denominator
of (7.4) remains positive. This is eﬀectively introducing a smoothing in the payoﬀ
function and, indeed, is switching oﬀ the eﬀect of the price impact close to expiry.
They do, however, oﬀer the ﬁnancial argument that transaction costs act as a natural
smoothing close to strike and close to expiry and so the cost of replication (hence the
price) would naturally be smoothed.
It turns out that the small τ analysis considered earlier is quite similar to (7.4), in
particular using the scaling (4.7) leads to the following smallτ equation for a put:
φ −
η
2
φ
η
−
σ
2
K
2
φ
ηη
2
_
1 −
_
ˆ
λK
1+
ˆ
λH(−η)
_
φ
ηη
_
2
+rKH(−η) = 0, (7.5)
and, similarly, for a call
φ −
η
2
φ
η
−
σ
2
K
2
φ
ηη
2
_
1 −
_
ˆ
λK
1−
ˆ
λH(η)
_
φ
ηη
_
2
−rKH(η) = 0. (7.6)
The two equations above are subject to the same boundary conditions as employed
in section 4.6. Indeed, note that these are the same form considered in section
4.6 but with a discontinuous jump in the value of the elasticity/liquidity parameter
CHAPTER 7. OTHER MODELS 141
(equivalent to λ) at S = K (η = 0). Another interesting point to note is that
the symmetry seen in the fullfeedback model of section 4.6 has now been broken,
φ
call
(η) ,= φ
put
(−η) + rK; this is as a consequence of the inclusion of V
S
into the
function λ(S, τ).
Equations (7.5) and (7.6) were then solved in a manner similar to that employed on
(4.8) and (4.9), and results for a call and a put are presented for a range of values
of
ˆ
λ in ﬁgures 7.2 and 7.3 respectively, with K = 1, σ = 0.2, r = 0.04. Although in
both cases there appears to be very little variation with
ˆ
λ, it was found for values in
excess of those shown (up to 0.3 in the cases of puts, 0.2 in the case of calls) that
the calculation failed, in a manner described in section 4.6, with the occurrence of
negative square roots in the calculation, suggesting yet again another solution regime.
This matter was not pursued further but it would appear that a treatment along the
lines of section 4.6 is again relevant, as is our discussion regarding the vanishing of
the denominator in (2.10).
0
0.005
0.01
0.015
0.02
0.025
0.03
0.035
0.04
0.4 0.2 0 0.2 0.4
PSfrag replacements
η
φ
ˆ
λ = 0
ˆ
λ = 0.2
Figure 7.2: Local (τ →0) call solution of the Sircar and Papanicolaou (1998) model
K = 1, r = 0.04, σ = 0.2, and
ˆ
λ = 0, 0.05, . . . , 0.2.
7.4 Bakstein and Howison (2003)
Bakstein and Howison (2003) developed a model for liquidity eﬀects which results in
CHAPTER 7. OTHER MODELS 142
0.04
0.035
0.03
0.025
0.02
0.015
0.01
0.005
0
0.4 0.2 0 0.2 0.4
PSfrag replacements
η
φ
ˆ
λ = 0
ˆ
λ = 0.3
Figure 7.3: Local (τ →0) put solution of the Sircar and Papanicolaou (1998) model
K = 1, r = 0.04, σ = 0.2, and
ˆ
λ = 0 0.05, . . ., 0.3.
the following PDE,
V
τ
−
1
2
σ
2
S
2
V
SS
−
ˆ
λσ
2
S
3
V
2
SS
−
1
2
ˆ
λ
2
(1 −α)
2
σ
2
S
4
V
3
SS
−rSV
S
+rV = 0, (7.7)
where
ˆ
λ ∈ R is a measure of the market’s liquidity and α is a measure of the price
slippage impact of a trade felt by all market participants, α = 1 corresponding to no
slippage. The ﬁrst point to note is that this equation looks uncannily like a small
ˆ
λ
expansion of the previously discussed models, but we shall return to this later. First
we shed a little light onto the term price slippage.
When a large order has been placed, the large trader will inevitably obtain a worse
price for the order than the quoted prices. However, a question that may be asked is
whether this large order should have a permanent aﬀect on the price process or not.
Assuming that a market maker provides the quotes, and that he has also provided
the other side of the large trade, then it is not unreasonable to expect that in the
next period the market maker will quote diﬀerent prices in order to neutralise his
position, hence a permanent impact will be felt by the market; this is what Bakstein
and Howison (2003) term price slippage.
It can be shown that the nonsmooth scaling (4.7) can be applied to the Bakstein and
Howison (2003) model to obtain a valid local solution, and indeed this shall be done
CHAPTER 7. OTHER MODELS 143
in the following subsection. However it transpires that, unlike in the Sch¨ onbucher
and Wilmott (2000), Frey (1998, 2000) and Sircar and Papanicolaou (1998) cases, a
suitable scaling (applicable to standard put and call payoﬀ proﬁles) can be found in
the class of smooth functions. The following outlines the procedure adopted to ﬁnd
such a scaling.
For the standard payoﬀ proﬁle (puts and calls) we have that, due to the increasing
gamma in the region close to strike and expiry, the cubic V
SS
term in equation (7.7)
will dominate over the other lower order terms in V
SS
. Hence to ﬁnd an appropriate
scaling we require a balancing between this cubic term and the time derivative, i.e.
V
τ
∼
1
2
ˆ
λ
2
(1 −α)
2
σ
2
S
4
V
3
SS
. (7.8)
We seek a solution of the form
V (S, τ) = τ
γ
1
f(η) where η =
S −K
τ
γ
2
where we require γ
1
= γ
2
= γ in order to obtain an O(1) inner solution and correctly
match with the standard payoﬀ proﬁles (cf. section 3.1). Substituting this into
equation (7.8) we have
γτ
γ−1
(f −ηf
η
) ∼
1
2
ˆ
λ
2
(1 −α)
2
σ
2
K
4
f
3
ηη
τ
−3γ
we can see that in order to obtain an appropriate balance of terms (for all τ) we are
forced to set
γ −1 = −3γ ⇒ γ =
1
4
.
Hence applying this scaling to the full equation (7.7) it can be shown that in the limit
τ →0 the equation for the inner solution becomes the cubically nonlinear ODE
1
f
3
ηη
+ν
1
(ηf
η
−f) = 0 (7.9)
where ν
1
=
_
2
ˆ
λ
2
(1 −α)
2
σ
2
K
4
_
−1
. The boundary conditions arise from an asymp
totic match which is the same as for the corresponding BlackScholes (
ˆ
λ = 0) case,
1
Note that ν
1
can be scaled out by setting f(η) =
√
ν
1
ˆ
f(η) but we will not do so as this would
place the constant ν
1
into the boundary conditions.
CHAPTER 7. OTHER MODELS 144
i.e.
f →0 as η →∞, f →−η as η →−∞ for calls, (7.10a)
f →η as η →∞, f →0 as η →−∞ for puts. (7.10b)
The system (7.9) and (7.10) can be solved using standard ﬁnitediﬀerence techniques.
Figure 7.4 shows sample results for the inner put solution with σ = 0.2, r = 0.04,
K = 1, α = 1.5 and with λ = 0.01, 0.5, 1, 5. Note that as λ is decreased (i.e. the
parameter ν
1
increased) the solution appears to collapse down onto the standard put
option payoﬀ proﬁle, with the solution becoming increasingly focused around η = 0.
0.1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
1 0.5 0 0.5 1
P
S
f
r
a
g
r
e
p
l
a
c
e
m
e
n
t
s
f
(
η
)
η
λ ↓, ν
1
↑
Figure 7.4: Solution to equation (7.9) for a put option with λ = 0.01, 0.5, 1, . . ., 5,
σ = 0.2, r = 0.04, K = 1, and α = 1.5.
However there is another subtlety that arises with the above scaling in the limit
α → 1, i.e. in the absence of price slippage impact. Here the inner equation (7.9)
becomes degenerate and looking again at the full PDE (7.7) it is clear that the scaling
used above is no longer appropriate when α →1. In this limit, the smallness of 1−α
must be considered, in this case the quadratic V
2
SS
term in the equation can no longer
be neglected and will be of comparable size to the cubic term for certain ranges of
τ, suggesting that an asymptotic breakdown occurs when τ = O
_
(1 −α)
8
_
, this can
be seen by balancing the quadratic and cubic V
SS
terms in equation (7.7). Therefore
the limit α →1 is not a trivial one. When α = 1 however, the cubic term in equation
CHAPTER 7. OTHER MODELS 145
(7.7) disappears and the equation is reduced to
2
V
τ
−
1
2
σ
2
S
2
V
SS
−
ˆ
λσ
2
S
3
V
2
SS
−rSV
S
+rV = 0. (7.11)
In this case the appropriate balancing of terms is now between the quadratic term in
V
SS
and the time derivative. Performing the same power balancing analysis as above
leads to the conclusion that the appropriate scaling for the problem when α = 1 is
given by
V (S, τ) = τ
1
3
g(ξ) where ξ =
S −K
τ
1
3
,
leading to the equation (in the limit τ →0)
3
g
2
ξξ
+ν
2
(ξg
ξ
−g) = 0 (7.12)
where ν
2
=
_
3
ˆ
λσ
2
K
3
_
−1
and it is once again coupled with the boundary conditions
(7.10). Note that if we diﬀerentiate equation (7.12) then we obtain
g
ξξ
(2g
ξξξ
+ν
2
ξ) = 0
which suggests two solutions to equation (7.12)
g(ξ) = −
ν
2
ξ
4
48
+
C
1
ξ
2
2
+C
2
ξ +
C
2
1
ν
2
,
g(ξ) = D
1
ξ,
however neither of these satisfy the required boundary conditions (7.10). In addition,
attempts to solve the system numerically using the same techniques employed on
equation (7.9) proved ineﬀective. The results were not consistent with grid reﬁnement
and kinked solutions started to appear. This suggests that either the chosen form of
the solution is inappropriate or that a smooth solution to the equation may not exist.
Furthermore the numerics indicate that the solution most likely takes on the form
of the (nonsmooth) payoﬀ proﬁle. This shall not be investigated further, rather we
leave this as a subject of future research.
2
It is (very) interesting to note that this PDE also arises in the quadratic transaction cost model
of Cetin et al. (2004).
3
Note again that ν
2
can be scaled out by setting g(ξ) = ν
2
ˆ g(ξ) but we will not do this.
CHAPTER 7. OTHER MODELS 146
Finally, let us recall the equation arising from the Frey (2000) model, i.e.
V
τ
−
σ
2
S
2
V
SS
2
_
1 −
ˆ
λSV
SS
_
2
−rSV
S
+rV = 0. (7.13)
Making the assumption that
ˆ
λ is small and further that
ˆ
λSV
SS
¸ 1 then we can
make the approximation
(1 −
ˆ
λSV
SS
)
−2
≈ 1 + 2
ˆ
λSV
SS
+. . .
and hence equation (7.13) can be approximated by
V
τ
−
1
2
σ
2
S
2
V
SS
−
ˆ
λσ
2
S
3
V
2
SS
−rSV
S
+rV = 0,
which coincides with the Bakstein and Howison (2003) model for α = 1, i.e. equation
(7.11). Clearly the assumption that
ˆ
λSV
SS
¸ 1 prohibits the denominator of the
equation (7.13) vanishing, but also leads to a regime in which standard put and call
payoﬀ proﬁles are not permitted due to their inﬁnite second derivatives.
7.4.1 Nonsmooth solutions to the Bakstein and Howison
(2003) model
As mentioned in the previous section, if we apply the scaling (4.7) to equation (7.7),
we can obtain a valid local solution and furthermore doing so for a put we obtain
φ −
η
2
φ
η
−
1
2
σ
2
K
2
φ
ηη
−
ˆ
λσ
2
K
3
(φ
ηη
)
2
−
ˆ
λ
2
2
(1 −α)
2
σ
2
K
4
(φ
ηη
)
3
+rKH(−η) = 0,
(7.14)
and
φ −
η
2
φ
η
−
1
2
σ
2
K
2
φ
ηη
−
ˆ
λσ
2
K
3
(φ
ηη
)
2
−
ˆ
λ
2
2
(1 −α)
2
σ
2
K
4
(φ
ηη
)
3
−rKH(η) = 0,
(7.15)
for a call, with the boundary conditions described in section 4.6.
The cubic nonlinearity in the secondorder derivative φ
ηη
demands a somewhat dif
ferent numerical approach from that adopted for the Sircar and Papanicolaou (1998)
CHAPTER 7. OTHER MODELS 147
model. Consequently, we followed a treatment which considered the problem as a
system in φ and φ
1
= φ
η
, thereby resulting in two ﬁrstorder equations. Secondorder
ﬁnite diﬀerencing, coupled with Newton iteration was then employed. Sample results
are shown in ﬁgures 7.5 and 7.6 for puts and calls respectively, with σ = 0.2, r = 0.04,
K = 1, α = 1.5, and with
ˆ
λ = −5, 4.75, . . . , 5. Bakstein and Howison (2003) do
discuss possible values for these latter two parameters (it appears that [
ˆ
λ[ is likely
very small). Other computations performed suggested that although the variation
of option values with α is generally quite small (with ﬁxed
ˆ
λ), in some regimes the
numerical scheme failed to converge, suggesting the possibility of regimes for which
solutions of (7.14) and (7.15) do not exist (which in turn suggests the possibility of
a regime akin to that described in section (4.6)).
0.04
0.035
0.03
0.025
0.02
0.015
0.01
0.005
0
0.6 0.4 0.2 0 0.2 0.4 0.6
PSfrag replacements
η
φ
ˆ
λ = 0
ˆ
λ = 5
ˆ
λ = −5
Figure 7.5: Local (τ → 0) put solution of the Bakstein and Howison (2003) model
K = 1, r = 0.04, σ = 0.2, α = 1.5, and
ˆ
λ = −5, 4.75, . . ., 5.
7.4.2 New nonsmooth solutions to the BlackScholes equa
tion
It should be noted that when
ˆ
λ = 0, the Bakstein and Howison model degenerates
to the standard BlackScholes equation and, somewhat intriguingly, ﬁgures 7.5 and
7.6 indicate that a nontrivial solution to the classical BlackScholes problem does
CHAPTER 7. OTHER MODELS 148
0
0.005
0.01
0.015
0.02
0.025
0.03
0.035
0.04
0.6 0.4 0.2 0 0.2 0.4 0.6
PSfrag replacements
η
φ
ˆ
λ = 0
ˆ
λ = 5
ˆ
λ = −5
Figure 7.6: Local (τ → 0) call solution of the Bakstein and Howison (2003) model
K = 1, r = 0.04, σ = 0.2, α = 1.5, and
ˆ
λ = −5, 4.75, . . ., 5.
exist with a discontinuous delta, quite distinct from the classical solution. Although
this new solution is found as the limiting case of the nonlinear Bakstein and Howison
model it can be illustrated by directly applying the scaling (4.7) to the BlackScholes
equation, i.e. seeking a solution of the BlackScholes equation (3.2) of the form
V
BS
= −τ
1
2
ηH(−η) +τφ
BS
(η).
Doing so we arrive at the equation for the inner solution of a put
φ
BS
−
η
2
φ
BS
η
−
1
2
σ
2
K
2
φ
BS
ηη
+rKH(−η) = 0,
i.e. equation (7.15) with
ˆ
λ = 0. To illustrate the nonsmooth behaviour of these
solutions ﬁgure 7.7 gives sample numerical solutions of the BlackScholes equation,
(3.2), using the Keller (1978) scheme (described in section 4.7) to build in the jump
in the ﬁrst derivative.
Note that this does not contradict the standard and well known uniqueness results
of the (linear) BlackScholes equation as these results are only valid in a restricted
class of smooth (classical) solutions to the PDE. In fact it should also be noted
that there also exists inﬁnitely many smooth solutions to the BlackScholes equation
but ones which do not satisfy the required growth conditions on the coeﬃcients of
CHAPTER 7. OTHER MODELS 149
0.2
0
0.2
0.4
0.6
0.8
1
0 0.5 1 1.5 2
PSfrag replacements
V
S
τ = .25
τ = 1
Figure 7.7: Nonsmooth solution of the BlackScholes equation. K = 1, r = 0.04,
σ = 0.2.
the PDE, namely the volatility term. Uniqueness of the BlackScholes PDE relies
on its ability to be reduced to the heat equation where uniqueness results are well
known, furthermore uniqueness of the heat equation is only ensured if we prescribe
the behaviour of the solution for large [x[. Also Widder (1975) shows that there is at
most one solution that is nonnegative for t ≥ 0 and all x, a reasonable assumption
when u(x, t) models absolute temperature or option prices.
7.5 Liu and Yong (2005)
The model of Liu and Yong (2005), previously discussed in chapter 2, attempts to
overcome the undesirable asymptotic behaviour at expiry by eﬀectively ‘switching oﬀ’
the eﬀect of liquidity as we approach it. This is achieved by adding a time dependency
to the function λ(S, τ) of the form
λ(S, τ) =
ˆ
λ(1 −e
−βτ
) (7.16)
where
ˆ
λ, β ∈ R. The authors’ rationale behind this choice of the liquidity function is
stated that as time passes, the private information about the asset value is gradually
revealed so that the price impact gradually decreases to zero at maturity, which will
CHAPTER 7. OTHER MODELS 150
also prevents any stock price manipulation at maturity. The nonsmooth scaling
(4.7) is no longer appropriate for this PDE and in fact it transpires that the standard
BlackScholes scaling (3.3) is the relevant one to use as τ →0, leading to
σ
2
K
2
f
ηη
+ηf
η
−f = 0.
Note that this has the same structure as the standard BlackScholes model as τ →0,
i.e. equation (3.5) and hence the asymptotic behaviour of the Liu and Yong (2005)
model close to expiry will be close to the BlackScholes model. In addition, it appears
that this form of the liquidity function λ(S, τ) may not completely circumvent the
issues associated with the vanishing of the denominator, as we shall outline next.
7.5.1 Vanishing of the denominator
First we consider the ﬁrstorder feedback case, i.e. equation (3.1) with λ(S, τ) deﬁned
by (7.16). To determine if the denominator vanishes in this simple case reduces to
determining whether there exists a real solution S
∗
(τ) to the equation
1 −
ˆ
λ(1 −e
−βτ
)
σS
∗
√
2πτ
e
−
1
2
d
1
(S
∗
,τ)
2
= 0,
where d
1
(S, τ) has been deﬁned previously. The ﬁrst point to note is that the Liu
and Yong (2005) model reduces to the Sch¨ onbucher and Wilmott (2000) model in
the limit β →∞; as such the location of the vanishing denominator in this limit was
shown in ﬁgure 3.7. Conversely when considering the limit β → 0, the denominator
will never vanish since here the denominator becomes uniformly unity. Figure 7.8
shows the location of the singular denominator for various values of β, hence we can
see that the denominator does still vanish for suﬃciently large values of β. This will
not cause any problems in the ﬁrstorder feedback model (as was outlined in chapter
3), but it may for the full feedback case which we now consider.
In the full feedback case of the Liu and Yong (2005) model at τ = 0 the denominator
never vanishes because the denominator reduces to 1 at τ = 0. However, it is of
interest to check to see if the denominator vanishes for τ > 0. The best we can do
CHAPTER 7. OTHER MODELS 151
0.97
0.98
0.99
1
1.01
1.02
0 0.005 0.01 0.015 0.02 0.025 0.03 0.035 0.04 0.045
P
S
f
r
a
g
r
e
p
l
a
c
e
m
e
n
t
s
τ
S
β increasing
Figure 7.8: Location of the vanishing of the denominator for the Liu and Yong (2005)
model for various value of β. K = 1, r = 0.04, σ = 0.2, λ = 0.1, and β = 1 10
5
,
2 10
5
, . . ., 1 10
6
.
is investigate the region close to strike and expiry, i.e. τ ¸ 1 and S ≈ K. In this
region the Liu and Yong (2005) model is identical to the BlackScholes local solution
and hence we can approximate V
SS
in this region by the BlackScholes local solution,
i.e.
V
SS
=
e
−
(S−K)
2
2τσ
2
K
2
σK
√
2πτ
,
valid for both puts and calls. Using this, and the knowledge that in this region
e
−βτ
≈ 1 −βτ the denominator will thus vanish if and when
1 −
βλ
σK
_
τ
2π
e
−
(S−K)
2
2τσ
2
K
2
= 0.
The exponential is bounded above by one and so the denominator will not vanish
provided
β <
σK
λ
_
2π
τ
,
in other words there is a potential for model failure (breakdown) in the region
τ = O(β
−2
),
if this is also in the region τ ¸1, hence if β is suﬃciently large. Liu and Yong (2005)
use the value β = 100 in the numerics giving τ = O(10
−5
) which is within the region
where τ ¸1.
CHAPTER 7. OTHER MODELS 152
7.6 Jonsson and Keppo (2002)
Another model (in addition to Bakstein and Howison, 2003) that can be found to
be free of the problems with nonsmooth solutions for standard put and call payoﬀ
proﬁles, is the Jonsson and Keppo (2002) model. Here the price observed in the
market, S, is dependent on an exogenous Brownian motion (ds = µsdt +σsdW
t
) via
an exponential price impact function, i.e.
dS
S
= e
gf(S,t)
ds
s
, (7.17)
where f(S, t) again denotes the number of stocks held by the hedger and g a so called
eﬀect parameter. This leads to a very diﬀerent nonlinear pricing PDE, namely
V
τ
−
1
2
σ
2
S
2
e
2aV
S
V
SS
−rSV
S
+rV = 0, (7.18)
where a ∈ R. It transpires that the relevant scaling in this case is also that of
BlackScholes (3.3), which leads to the equation
σ
2
K
2
e
2afη
f
ηη
+ηf
η
−f = 0, (7.19)
subject to
f
η
→1 as η →∞, f →0 as η →−∞,
for a call and
f
η
→−1 as η →−∞, f →0 as η →∞,
for a put. Note that in this case there is no discontinuity in the derivatives and we
appear to have a classical solution. However it is clear that this model will encounter
diﬃculties if we have discontinuous payoﬀ proﬁles, such as binary options. The above
system was solved using a straightforward fourthorder shooting scheme; results for
calls and puts are shown in ﬁgures 7.9 and 7.10 respectively, with σ = 0.2, K = 1,
and for a = −1, 0.9, . . ., 1. Note that, in agreement with proposition 3.2 of Jonsson
and Keppo (2002), the call option value is monotonically increasing in the parameter
a, and conversely the put monotonically decreasing. We therefore conclude that
equation (7.18) admits wellbehaved solutions at times close to expiry. However the
CHAPTER 7. OTHER MODELS 153
modelling assumptions used to arrive at this equation may be questioned; there does,
however, exist a link to the other models discussed in this chapter, a link which is
outlined below.
0
0.1
0.2
0.3
0.4
0.5
0.6
0.4 0.2 0 0.2 0.4
PSfrag replacements
η
f
a = 1
a = −1
Figure 7.9: Local (τ → 0) call solution of the Jonsson and Keppo (2002) model
K = 1, σ = 0.2, and a = −1, 0.9, . . ., 1.
0
0.1
0.2
0.3
0.4
0.5
0.6
0.4 0.2 0 0.2 0.4
PSfrag replacements
η
f
a = 1
a = −1
Figure 7.10: Local (τ → 0) put solution of the Jonsson and Keppo (2002) model
K = 1, σ = 0.2, and a = −1, 0.9, . . ., 1.
CHAPTER 7. OTHER MODELS 154
7.6.1 Connections with the other modelling frameworks
A connection of the Jonsson and Keppo (2002) model with the models of Frey (1998,
2000) etc. can be seen if we consider small values of the eﬀect parameter g, or more
speciﬁcally if gf(S, t) ¸1. In this case equation (7.17) can be approximated to
dS
S
≈
_
1 +gf(S, t)
_
ds
s
=
ds
s
+gf(S, t)
ds
s
= µdt +σdW
t
+gf(S, t)
ds
s
.
It is now possible to equate this model to the reduced form SDE model, (2.2), and
in order for the two processes to agree we require that
gf(S, t)
ds
s
= λ(S, t)
df(S, t)
S
,
in other words the liquidity function λ(S, t) can be chosen such that
λ(S, t)
S
= g
f
s
ds
df
.
If we identify f as the number of shares traded and s as their price, then it is possible
to deﬁne the quantity
s
f
df
ds
= L as the price elasticity of the market to trades (cf.
section 1.4). The models, therefore, become equivalent for small gf(S, t) if we allow
the parameter λ to be
λ(S, t) =
g
L
S. (7.20)
This is intuitive since, if the elasticity of the market increases, i.e. L → ∞ the
relationship (7.20) gives that λ → 0 and hence the liquidity of the market also
increases; in line with our understanding of price elasticity.
Chapter 8
Explaining the Stock Pinning
Phenomenon
This chapter aims to illustrate one of the most documented forms of price impact,
that of the stock pinning phenomenon. Stock pinning is deﬁned as the tendency of
stock prices to move to the strike price of heavily traded option contracts as the
options approach expiry. It transpires that the models outlined in this thesis can
be used to partly explain and quantify such interesting market behaviour. First,
however, we start by discussing some of the empirical evidence for stock pinning.
The most comprehensive empirical investigation of stock price pinning was under
taken by Ni et al. (2005), who provided striking evidence that the presence of options
perturbs the prices of underlying stocks, more speciﬁcally that on expiration dates,
clustering appeared when nonoptionable stocks became optionable and disappeared
when optionable stocks became nonoptionable. Prior to this there had been little
indication of any signiﬁcant impact, despite understandable interest from market pro
fessionals. Ni et al. (2005) suggested that this pinning eﬀect is most likely due to
delta hedgers moving the price and possibly due to intentional market manipulation.
Krishnan and Nelken (2001) observed Microsoft stocks (during the period 19902001)
and concluded that the probability of the stock price pinning at strike (i.e. being
within a small of strike) on expiration days was 23.29%, and on nonexpiration
155
CHAPTER 8. EXPLAINING THE STOCK PINNING PHENOMENON 156
days was only 13.52%. In addition, Avellaneda and Lipkin (2003) described anecdo
tal evidence of the pinning of J.D. Edwards stocks in 2003.
8.1 Linear price impact
Avellaneda and Lipkin (2003) developed a model in which stock trading, undertaken
to maintain delta hedges on existing netpurchasedoption positions is allowed to
impact the price and as a consequence pushes the stock price toward the strike price
as expiration approaches, i.e. pinning. The model assumes that the price impact is
proportional to the ‘rate of change’ of the BlackScholes delta, i.e.
dS
S
∼
∂∆
∂t
dt.
Furthermore they argue, counter to the assumptions made in chapter 2, that the net
position (in stock) of delta hedgers in the market is short rather than long. The
rational behind this is as follows: If large institutions such as hedge funds sell options
to market makers (deﬁned in section 1.6), then the market makers will inevitably
hedge their position. Since the market makers will be long options, they must become
short the underlying in order to hedge their position. If the large institution is not
also hedging the sale of the options, then the net position (of the buyer and seller
of the option) in the market for the underlying will be short. With this in mind
Avellaneda and Lipkin (2003) therefore make the assumption that the underlying
stochastic process is modiﬁed to
dS
S
=
_
µ −nE
∂∆
∂t
_
dt +σdW
t
, (8.1)
where n is the open interest and E a constant price elasticity term; note the negative
sign. The focus of the work by Avellaneda and Lipkin (2003) is not on option pricing,
but rather on calculating the socalled pinning probability, deﬁned as the probability,
denoted P(S, t), that the stock price will end up at S = K at t = T. This can be
calculated via the Kolmogorov backward equation which is given by
1
∂P
∂t
+µ(S, t)
∂P
∂S
+
1
2
σ
2
(S, t)
∂
2
P
∂S
2
= 0, (8.2)
1
See for example proposition 5.11 of Bj¨ ork (2004)
CHAPTER 8. EXPLAINING THE STOCK PINNING PHENOMENON 157
where µ(S, t) is the drift of the process and σ(S, t) the volatility. For the model of
(8.1) this corresponds to solving the system
∂P
∂t
+
_
µS −nE
∂∆
∂t
_
∂P
∂S
+
1
2
σ
2
S
2
∂
2
P
∂S
2
= 0, (8.3a)
P(S, T) =
_
_
_
1, if S = K;
0, otherwise.
(8.3b)
Furthermore, we assume for simplicity (and to be consistent with the original paper)
that the delta is given by the delta of a call option, which can calculate directly as
∂∆
∂t
=
_
log(S/K) −
_
r +
1
2
σ
2
_
(T −t)
2σ(T −t)
3
2
√
2π
_
exp
_
−
_
log(S/K) +
_
r +
1
2
σ
2
_
(T −t)
_
2
2σ
2
(T −t)
_
.
(8.4)
Avellaneda and Lipkin (2003) showed that, for the special case µ = 0 and r+
1
2
σ
2
= 0,
the system (8.3) admits an exact solution given by
P(S, t) = 1 −exp
_
−
nE
σ
_
2π(T −t)
exp
_
−
_
log(S/K)
_
2
2σ
2
(T −t)
__
. (8.5)
Figure 8.1 shows the solution to equation (8.5) from which it is clear that there is
a greater pinning probability when nE is increased. For the case when µ ,= 0, and
more importantly r +
1
2
σ
2
,= 0 the authors showed that there still exists a positive
probability of pinning. To determine this probability however we are required to solve
(8.3) numerically. Doing so we see that an increase in r results in an increased pinning
probability below the strike price and, correspondingly, a decreased probability above;
this might be expected since, for an increase in r, the process will drift upwards more
on average.
Subsequent to the work of Avellaneda and Lipkin (2003), Jeannin et al. (2006) com
mented on the similarities of the model with those outlined in this thesis. In fact,
the above model can be seen as simply an approximation to the form of price impact
considered in this thesis with any terms from the quadratic variation of the process
and the eﬀects on the volatility ignored. This can be seen more clearly by considering
CHAPTER 8. EXPLAINING THE STOCK PINNING PHENOMENON 158
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0.7 0.8 0.9 1 1.1 1.2 1.3
P
S
f
r
a
g
r
e
p
l
a
c
e
m
e
n
t
s
S
P
nE increasing
Figure 8.1: The pinning probability (8.5) for values of nE = 0.5, 1, . . ., 5. T −t = 0.1,
K = 1, and σ = 0.2.
the following
dS
S
= µdt +σdW
t
−nEd∆(S, t)
= µdt +σdW
t
−nE
_
∂∆
∂t
dt +
∂∆
∂S
dS +
1
2
∂
2
∆
∂S
2
(dS)
2
_
=
1
1 +nES
∂∆
∂S
_
µ −nE
_
∂∆
∂t
+
σ
2
S
2
2
_
1 +nES
∂∆
∂S
_
2
∂
2
∆
∂S
2
__
dt +
σdW
t
1 +nES
∂∆
∂S
,
⇒
dS
S
= ˆ µ(S, t)dt + ˆ σ(S, t)dW
t
. (8.6)
Hence the model of Avellaneda and Lipkin (2003) only includes the time derivative
term of the change in delta, whereas the model outlined above includes the time
derivative and the delta convexity term.
2
Jeannin et al. (2006) solve the associated
Kolmogorov equation (numerically) to determine the pinning probabilities for the
more accurate model (8.6) and showed that the pinning probability was higher in
this case than the corresponding Avellaneda and Lipkin (2003) model. This can
be attributed to the fact that not only does the price impact aﬀect the drift of
the process it also aﬀects the volatility, which for the case investigated above (i.e.
hedging a long call option position) is decreased in the region of the strike. The
2
Note that for (8.6) the problems associated with the vanishing of the denominator highlighted
in this thesis are not present, since 1 +nES
∂∆
∂S
> 0 in the entire domain; this is due to the net long
option position assumption.
CHAPTER 8. EXPLAINING THE STOCK PINNING PHENOMENON 159
combined eﬀect is that the drift pushes the underlying towards the strike and the
reduced volatility eﬀectively keeps the underlying in the vicinity of the strike. Figure
8.2 shows comparisons of the pinning probabilities for the model (8.6), obtained from
the associated Kolmogorov backward equation, and the Avellaneda and Lipkin (2003)
model, for the same value of nE.
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.8 0.85 0.9 0.95 1 1.05 1.1 1.15 1.2
P
S
f
r
a
g
r
e
p
l
a
c
e
m
e
n
t
s
S
P
Figure 8.2: Comparing the pinning probability associated with (8.6) (solid line) with
the model of Avellaneda and Lipkin (2003) (dotted line) for nE = 0.1, T − t = 0.1,
K = 1, σ = 0.2, and r +
1
2
σ
2
= 0.
8.2 Nonlinear price impact
Recently, empirical work by Kempf and Korn (1999) brought into question the reli
ability of the assumption of linear price impact using empirical evidence on German
index futures. This motivated several studies to determine quantitatively how a mar
ket order of a given size would aﬀect the price of the underlying. Furthermore, Plerou
et al. (2002) showed that this relationship is sublinear (i.e. concave) and further still,
Bouchaud et al. (2002) studied the distribution of the order ﬂow and the resulting
average order book, ﬁnding that the average order book has its maximum away from
the best bid (or ask), possibly helping to explain the concavity of the price impact
function. Lillo et al. (2003) concluded that price impact is well described by a power
CHAPTER 8. EXPLAINING THE STOCK PINNING PHENOMENON 160
law where the change in price (dS) caused by an order of volume w is given by
dS(w) ∼ w
p
where p is found to be between 0.1 and 0.5; again a concave function. However the
value of p is greatly contested in the literature and there has been many attempts to
determine its value. Almgren et al. (2005) proposed a 3/5 power dependence whereas
Gabaix et al. (2003) hypothesised that p = 1/2. In addition, recent work by Potters
and Bouchaud (2003) suggested that this relationship may be better described by a
logarithmic price impact function. This still remains an open question in the empirical
literature.
In recent years a number of computersimulated, artiﬁcial ﬁnancial markets have been
constructed; see LeBaron (2000) for a review of work in the ﬁeld. These agentbased
models aim to reproduce the main stylized facts observed in real ﬁnancial markets,
such as fattailed distributions of returns and volatility clustering. Jeannin et al.
(2006) proposed such an agentbased model (based on a double auction) in an attempt
to model stock pinning whilst also incorporating the aforementioned nonlinear price
impact. Simulations of the model give the same qualitative eﬀect on the drift as
the model described by (8.6) but with increased volatility around strike rather than
decreased, which is at odds with their numerical solutions of the model (8.6).
More recently, Avellaneda et al. (2007) attempted to incorporate the empirical evi
dence of nonlinear price impact into the linear model ﬁrst introduced in Avellaneda
and Lipkin (2003). They do so by assuming a price impact of the form
dS
S
∼ sgn
_
∂∆
∂t
_¸
¸
¸
¸
∂∆
∂t
¸
¸
¸
¸
p
dt
hence they assume the stochastic process for the underlying in the presence of delta
hedgers hedging a long call position to take the form
dS
S
=
_
µ −nE sgn
_
∂∆
∂t
_¸
¸
¸
¸
∂∆
∂t
¸
¸
¸
¸
p
_
dt +σdW
t
.
CHAPTER 8. EXPLAINING THE STOCK PINNING PHENOMENON 161
It is clear that the pinning probability under such assumptions will be given by the
solution to the following Kolmogorov backward equation
∂P
∂t
+
_
µS −nE sgn
_
∂∆
∂t
_¸
¸
¸
¸
∂∆
∂t
¸
¸
¸
¸
p
_
∂P
∂S
+
1
2
σ
2
S
2
∂
2
P
∂S
2
= 0, (8.7a)
P(S, T) =
_
_
_
1, if S = K;
0, otherwise,
(8.7b)
in conjunction with the previously calculated delta (8.4). Avellaneda et al. (2007)
showed that there is a zero probability of pinning if p ≤
1
2
and conversely a nonzero
probability if p >
1
2
. Figure 8.3 shows the numerical solution to (8.7) for ﬁve values
of the exponent p, it can be clearly seen that the pinning probability is monotonic
increasing in p.
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.8 0.85 0.9 0.95 1 1.05 1.1 1.15 1.2
P
S
f
r
a
g
r
e
p
l
a
c
e
m
e
n
t
s
S
P
p = 1
Figure 8.3: Solution to (8.7) for p = 0.8, 0.9, . . . , 1.2, T − t = 0.1, K = 1, σ = 0.2
and r +
1
2
σ
2
= 0.
8.3 A new nonlinear price impact model
The aim of this section is to suggest a possible extension to the model which would
incorporate the empirically observed powerlaw price impact into the more accurate
(linear) price impact model (8.6). The most consistent way to extend the model
CHAPTER 8. EXPLAINING THE STOCK PINNING PHENOMENON 162
would be to assume a price impact of the form
dS
S
∼ (d∆)
p
dS
S
∼
_
∂∆
∂t
dt +
∂∆
∂S
dS +
1
2
∂
2
∆
∂S
2
(dS)
2
_
p
however clearly this makes little mathematical sense. We instead choose to model
nonlinear price impact as
3
dS
S
= sgn(ˆ µ) [ˆ µ[
p
1
dt + sgn(ˆ σ) [ˆ σ[
p
2
dW
t
where ˆ µ(S, t) and ˆ σ(S, t) are given by (8.6) and the parameters p
1
and p
2
model the
degree of nonlinearity. Note that for p
1
= p
2
= 1 this reduces to the model of Jeannin
et al. (2006), namely (8.6). The corresponding Kolmogorov backward equation is thus
given by
∂P
∂t
+ sgn(ˆ µ) [ˆ µ[
p
1
∂P
∂S
+
1
2
sgn(ˆ σ) [ˆ σ[
2p
2
∂
2
P
∂S
2
= 0, (8.8a)
P(S, T) =
_
_
_
1, if S = K;
0, otherwise.
(8.8b)
Figure 8.4 shows the solution to equation (8.8) for various combinations of p
1
and p
2
.
It appears that p
2
has a greater aﬀect on the solution than p
1
; further investigation
of this model, however, is left as the subject of future research.
3
Note that this is eﬀectively assuming that the change of the log price with respect to time and
the Brownian motion is given by
d log S
dt
∼ sgn(ˆ µ)  ˆ µ
p1
,
d log S
dW
t
∼ sgn(ˆ σ) ˆ σ
p2
.
CHAPTER 8. EXPLAINING THE STOCK PINNING PHENOMENON 163
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.8 0.85 0.9 0.95 1 1.05 1.1 1.15 1.2
P
S
f
r
a
g
r
e
p
l
a
c
e
m
e
n
t
s
P
S
p
1
= 0.8
p
1
= 1
p
1
= 1.2
p
2
=
0
.
8
p
2
=
1
p
2
=
1
.
2
(a) p
2
= 1, p
1
= 0.8, 1 and 1.2.
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.8 0.85 0.9 0.95 1 1.05 1.1 1.15 1.2
P
S
f
r
a
g
r
e
p
l
a
c
e
m
e
n
t
s
P
S
p
1
=
0
.
8
p
1
=
1
p
1
=
1
.
2
p
2
= 0.8
p
2
= 1
p
2
= 1.2
(b) p
1
= 1, p
2
= 0.8, 1 and 1.2.
Figure 8.4: Solution to equation (8.8) (solid line) compared to (8.5) (dotted line) for
T = 0.1, K = 1, σ = 0.2, and r +
1
2
σ
2
= 0.
Chapter 9
The British Option
The most important questions of life are, for the most part, really only
problems of probability.
 Pierre Simon Laplace (17491827)
Th´eorie Analytique des Probabilit´es (1812)
In this chapter we aim to investigate the mathematical properties and ﬁnancial mo
tivations of the newly introduced British option (see Peskir and Samee, 2008a,b), a
new nonstandard class of early exercise option; such options can help to mediate the
illiquidity eﬀects discussed in the preceding chapters.
9.1 Introduction
The noarbitrage price of an earlyexercise option uses the implicit assumption that
the holder of the option will act optimally in the sense of following the optimal
strategy of exercising upon ﬁrst hitting the rational exercise free boundary (cf. section
1.3.5). If the holder of the option chooses to deviate from this strategy, then the
expected discounted payout (under the riskneutral measure) will be less than the
amount the writer received for the option at the start of the contract.
164
CHAPTER 9. THE BRITISH OPTION 165
Recall that the riskneutral pricing measure is used to determine the noarbitrage
price of the option, since the writer of the option can completely hedge away risk.
Indeed, if we make the assumption that the writer of an earlyexercise option will
hedge away the risk exposure associated with selling such an option, then from the
writer’s perspective at least, the exercise strategy followed by the holder is irrelevant.
For the purposes of this chapter, we shall assume that writers of earlyexercise options
perfectly hedge their positions, hence eliminating any risk associated with the options.
The holder of such an option is assumed not to be hedging the position; instead, he
is interested in maximising proﬁt given his view on the market. Peskir and Samee
(2008b) deﬁne such investors as true buyers, i.e. those who have no ability or desire
to sell the option; in short holders of ‘naked’ (unhedged) options. In this case, the
exercise strategy for the true buyer is of paramount importance. Furthermore, since
the true buyer has no interest in hedging his position, the real world drift of the
underlying will play a role in determining the rational exercise strategy.
Let us suppose initially that the holder of an earlyexercise option knows with cer
tainty the true drift of the underlying µ, and that it diﬀers from the riskneutral
drift rate r. In such a situation, the optimal exercise boundary for the true buyer
would deviate from the optimal exercise boundary under the assumption of a riskfree
drift. In fact, the true buyer’s rational exercise boundary would be given by the free
boundary of the following optimal stopping problem
1
V (S, t; µ) = sup
t≤γ≤T
E
P
S,t
_
e
−r(γ−t)
(K −S
T
)
+
¸
,
recall that the indices of the expectation denotes that the process is started at S at
time t and that the expectation is taken under the real world probability measure
P. Furthermore, such a buyer would presumably not invest in a put option unless
the true drift µ were less than the riskfree interest rate r, otherwise the true buyer
would purchase a call option.
1
Note that here we are making the assumption of a riskneutral investor (in the utility sense).
CHAPTER 9. THE BRITISH OPTION 166
Unfortunately, the drift of a stochastic process is notoriously diﬃcult to measure, and
so investors would not be able to know the true drift with any degree of certainty.
They can however have an estimate of the true drift, or a ‘view’ on the future direction
of the underlying, which they wish to take advantage of by buying such options. In
this case the true buyer has become a speculator in the sense of seeking to maximise
gains or minimise losses given his particular sentiment of future market conditions.
In other words a speculator who chooses to invest in an earlyexercise option must
be under the belief that the true drift of the underlying process is less than the
riskneutral drift r.
The British option, recently proposed by Peskir and Samee (2008a,b), is a new class of
earlyexercise option that attempts to utilise the idea of optimal prediction in order to
provide the true buyer with an inherent protection mechanism should the true buyer’s
beliefs on the future price movements not transpire. More speciﬁcally, at any time γ
during the term of the contract, the investor can choose to exercise the option, upon
which he receives the best prediction of the European put payoﬀ (K − S
T
)
+
(given
all the information up to the stopping time γ) under the assumption that the drift
of the underlying is µ
c
, the socalled contract drift, for the remaining term of the
contract. Hence the (now timedependent) payoﬀ proﬁle of the earlyexercise British
put option is given by
G(S, γ; µ
c
) = E
R
_
(K −S
T
)
+
[T
γ
¸
,
where the expectation is taken with respect to a new probability measure R, under
which the stock price evolves according to
dS
t
= (µ
c
−D)S
t
dt +σS
t
dW
R
t
,
where we have included (in anticipation of what will follow) a constant dividend yield
D. The value of the contract drift is chosen by the holder of the option at the start of
the contract and is selected to represent the level of protection (from adverse realised
drifts) that the holder requires; a higher µ
c
corresponds to a lower level of protection.
CHAPTER 9. THE BRITISH OPTION 167
9.2 The noarbitrage price
Let us ﬁx a probability space (Ω, T, Q), where Ω describes a ﬁnancial market with a
ﬁltration (T
t
)
t≥0
, which represents the information structure of the ﬁnancial market
(see Harrison and Kreps, 1979) with the unique riskneutral measure Q. Assume that
S
t
is an T
t
adapted stochastic process that describes the stock price process.
Analogous with the American option deﬁned in subsection 1.3.5, the noarbitrage
price of the British put option is given by the supremum over all stopping times γ
(adapted to the ﬁltration T
t
generated by the process S
t
) of the expected discounted
future payoﬀ. In contrast with an American option, the future payoﬀ is now itself an
expectation, i.e.
V (S, t) = sup
t≤γ≤T
E
Q
S,t
_
e
−r(γ−t)
E
R
_
(K −S
T
)
+
[T
γ
¸¸
.
Recall that the future payoﬀ is deﬁned as the best prediction of the European payoﬀ,
conditional on all the information available up to the stopping time γ. There are
numerous approaches to evaluating this expectation, for example we could directly
use the probability density function of the process under the measure R, to give
E
R
_
(K −S
T
)
+
[T
γ
¸
=
_
∞
0
(K −z)
+
f
R
(S, γ; z, T)dz,
where f
R
(S, γ; z, T) is the transitional probability density function of the process
started at time γ at the position S and ﬁnishing at time T at the position z and is
given by
2
f
R
(S, γ; z, T) =
1
σz
_
2π(T −γ)
exp
_
−
_
log
_
z
S
_
−
_
µ
c
−D−
1
2
σ
2
_
(T −γ)
_
2
2σ
2
(T −γ)
_
.
Tackling the above integral would provide us with the required expectation; however,
we shall adopt an alternative approach in order to give some intuition behind the
resulting expression. We can evaluate the expectation using direct integration with
respect to the probability measure R over the probability space Ω, i.e.
E
R
_
(K −S
T
)
+
[T
γ
¸
=
_
Ω
(K −S
T
)
+
dR.
2
See appendix C for a derivation.
CHAPTER 9. THE BRITISH OPTION 168
The random variable is a real valued function whose domain is given by S
T
∈ [0, ∞)
and so the integral can be written as
_
∞
0
(K −z)
+
dR(z) =
_
∞
0
(K −z)
+
R(dz) =
_
∞
0
(K −z)
+
R(z ∈ dz[T
γ
),
where the probability is now conditional on the ﬁltration of information up to the
stopping time γ. Further simpliﬁcation gives
_
K
0
(K −z)R(z ∈ dz[T
γ
).
In order to evaluate this integral we rewrite the function K−z as an integral to give
_
K
z=0
_
K−z
y=0
dyR(dz) =
_
K
y=0
_
K−y
z=0
R(dz)dy,
where we have also changed the order of integration. We can now exploit the fact
that (see appendix C)
_
K−y
0
R(z ∈ dz[T
γ
) = P
R
[z ≤ K −y[T
γ
] ,
= Φ
_
_
log
_
K−y
Sγ
_
−
_
µ
c
−D −
1
2
σ
2
_
(T −γ)
σ
√
T −γ
_
_
,
where Φ denotes the standard normal distribution function
Φ(z) =
1
√
2π
_
z
−∞
e
−
1
2
y
2
dy.
Finally we have our expression for the conditional expectation
E
R
_
(K −S
T
)
+
[T
γ
¸
=
_
K
0
Φ
_
_
log
_
K−y
Sγ
_
−
_
µ
c
−D−
1
2
σ
2
_
(T −γ)
σ
√
T −γ
_
_
dy
= S
γ
_ K
Sγ
0
Φ
_
log u −
_
µ
c
−D −
1
2
σ
2
_
(T −γ)
σ
√
T −γ
_
du,
where at the ﬁnal step we have made the substitution u =
K−y
Sγ
. The British option
value is thus given by
V (S, t) = sup
t≤γ≤T
E
Q
S,t
_
e
−r(γ−t)
S
γ
_ K
Sγ
0
Φ
_
log z −
_
µ
c
−D −
1
2
σ
2
_
(T −γ)
σ
√
T −γ
_
dz
_
.
CHAPTER 9. THE BRITISH OPTION 169
Deﬁning the function
G(S, t) = S
_ K
S
0
Φ
_
log z −
_
µ
c
−D −
1
2
σ
2
_
(T −t)
σ
√
T −t
_
dz,
we have that the problem now reads
V (S, t) = sup
t≤γ≤T
E
Q
S,t
_
e
−r(γ−t)
G(S, γ)
¸
.
This is the standard form of an Americantype option (cf. equation 1.11) and so we
can directly see the links with the existing American option theory.
3
It remains to
ﬁnd an analytical expression for G(S, t) which can be done easily with integration by
parts to give
G(S, t) = KΦ
_
log
_
K
S
_
−
_
µ
c
−D−
1
2
σ
2
_
(T −t)
σ
√
T −t
_
−Se
(µc−D)(T−t)
Φ
_
log
_
K
S
_
−
_
µ
c
−D +
1
2
σ
2
_
(T −t)
σ
√
T −t
_
.
Alternatively this can be written as
G(S, t) = KΦ
_
d
1
(S, t; µ
c
)
_
−Se
(µc−D)(T−t)
Φ
_
d
2
(S, t; µ
c
)
_
, (9.1)
where
d
1
(S, t; µ
c
) =
log
_
K
S
_
−
_
µ
c
−D −
1
2
σ
2
_
(T −t)
σ
√
T −t
, (9.2a)
d
2
(S, t; µ
c
) =
log
_
K
S
_
−
_
µ
c
−D +
1
2
σ
2
_
(T −t)
σ
√
T −t
= d
1
(S, t; µ
c
) −σ
√
T −t. (9.2b)
Note that the dependency on the contract drift rate µ
c
has been stated explicitly.
General optimal stopping theory can now be applied to this problem analogous with
the American option problem (cf. section 1.3.5) we have that
( = ¦(S, t) : V (S, t) > G(S, t)¦ (continuation set),
T = ¦(S, t) : V (S, t) = G(S, t)¦ (stopping set),
with the optimal stopping time deﬁned as
γ
∗
= inf¦t ∈ [0, T] : S
t
∈ T¦,
3
For the standard American put option the gain function is simply the (time homogeneous)
payoﬀ proﬁle i.e. G
A
(S, t) = (K −S)
+
.
CHAPTER 9. THE BRITISH OPTION 170
i.e. the ﬁrst time that the stock price enters the stopping region. It can be shown (see
Peskir and Samee, 2008b) that the stopping and continuation regions are separated
by a smooth function b(t), the free boundary, and hence ( = ¦(S, t) : S ∈ (b(t), ∞)¦.
Now applying standard optimal stopping and Markovian arguments, again analo
gous to the American put option, the problem can be conveniently expressed as the
following freeboundary problem:
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
V
t
+
1
2
σ
2
S
2
V
SS
+ (r −D)SV
S
−rV = 0 for S ∈ [b(t), ∞),
V (S, t) = G(S, t) on S = b(t),
V
S
(S, t) = G
S
(S, t) on S = b(t) (smooth ﬁt),
V > G in (,
V = G in T,
(9.3)
with the gain function G(S, t) given by equation (9.1). Before we continue, let us
ﬁrst take a closer look at the gain function.
9.2.1 The gain function
This section brieﬂy comments on the links between the gain function G(S, t) and the
analytical expressions of the corresponding European option values. Note that the
BlackScholes European put value is given by
V
P
E
(S, t) = Ke
−r(T−t)
Φ
_
d
1
(S, t; r)
_
−Se
−D(T−t)
Φ
_
d
2
(S, t; r)
_
,
where d
1
(S, t; r) emphasises the dependence on the parameter r and is given by
equation (9.2a) for µ
c
= r. Note that this looks very much like the gain function
and indeed if µ
c
= r, then it is clear that
G(S, t)[
µc=r
= e
r(T−t)
V
P
E
(S, t).
This is intuitive as the payoﬀ is the best prediction received now and so involves no
discounting, unlike the European option value. The above identity can also be used
to check that we are calculating G(S, t) correctly. Also note that as t → T we have
that Φ(d
2
) → Φ(d
1
) → 0 for S > K and conversely Φ(d
2
) → Φ(d
1
) → 1 for S < K
CHAPTER 9. THE BRITISH OPTION 171
which indicates that in this limit the gain function takes on the form of the standard
put payoﬀ condition, namely
G(S, T) = (K −S)
+
,
which is consistent with the fact that at t = T clearly the best prediction of the put
payoﬀ will be the payoﬀ itself.
In addition, this suggests that the British option may be considered as a compound
option, i.e. an option on an option. This type of option has been studied extensively,
compare Geske (1977), Geske (1979), Whaley (1982) and Hodges and Selby (1987),
and often leads to trivial solutions. For example, a standard American option writ
ten on an underlying European option has simply the same value as the underlying
European option, i.e. there exists no optimal earlyexercise region. However the in
novation with the British option considered here is that the two options are priced
under diﬀerent measures, resulting in nontrivial solutions.
9.3 Numerical treatment
The (nonlinear) system (9.3) can be solved numerically via a number of diﬀerent
methods. By far the easiest to implement is the Projected Successive Over Relax
ation (PSOR) algorithm, which attempts to solve the appropriately discretised and
linearised system (typically based on a CrankNicolson discretisation scheme), subject
to the constraint
V (S, t) ≥ G(S, t)
at every node and iteration. In addition it is advantageous to make the transformation
τ =
√
T −t, (9.4)
which has the eﬀect of concentrating the grid points close to expiry, the region where
the solution is changing most rapidly. Furthermore, transforming to logspace via
the transform
ˆ
S = log
_
S
K
_
will simplify the resulting equations. In fact, the entire
CHAPTER 9. THE BRITISH OPTION 172
system (9.3) can be nondimensionalised via an appropriate set of transformations,
which for the interested reader can be found in appendix B. Nondimensionalising
makes the system more parsimonious, but to retain ﬁnancial intuition, in what follows
we choose to retain the dimensional form.
Despite the PSOR algorithm’s ease of implementation, its major drawback (for the
purposes of this exposition) is that in order to determine the location of the free
boundary accurately, some form of interpolation is needed, and further if a more
accurate estimate of the freeboundary location is needed, then the number of grid
points must be increased, increasing the computation time at a disproportionate rate
to the increased accuracy gained in the freeboundary estimate. For these reasons a
more sophisticated method, the bodyﬁtted coordinate system, was employed for the
results given in the present chapter. Bodyﬁtted coordinates where ﬁrst proposed
by Landau (1950) and later applied to ﬁnitediﬀerence schemes by Crank (1957).
More recently Widdicks (2002) adapted the scheme to solve the standard American
option problem and Johnson (2007) to more complex options (involving multiple free
boundaries). The idea behind this technique is to make the transform (in addition
to the time change (9.4)),
ˆ
S =
S
b(τ)
. (9.5)
This eﬀectively maps the continuation region S ∈ [b(τ), ∞) onto the ﬁxed domain
ˆ
S ∈ [1, ∞), where the free boundary now becomes an additional variable in the prob
lem. Making these changes of variables yields the following modiﬁed ﬁxed boundary
problem
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
1
2τ
V
τ
−
1
2
σ
2
ˆ
S
2
V
ˆ
S
ˆ
S
−
_
r +
1
2τb(τ)
∂b(τ)
∂τ
_
ˆ
SV
ˆ
S
+rV = 0 for
ˆ
S ∈ [1, ∞),
V (
ˆ
S, t) = G(
ˆ
S, t) on
ˆ
S = 1,
V
ˆ
S
(
ˆ
S, t) = G
ˆ
S
(
ˆ
S, t) on S = 1 (smooth ﬁt),
V > G for
ˆ
S ∈ (1, ∞),
V = G for
ˆ
S ∈ [0, 1],
(9.6)
where the gain function must also be transformed via (9.4) and (9.5). The drawback
of this method is that we now have a more complicated equation to solve, but it
CHAPTER 9. THE BRITISH OPTION 173
is now in a ﬁxed domain and standard ﬁnitediﬀerence methods for such problems
are well understood and easily applicable. Note that this method provides highly
accurate results for the freeboundary locations, since no interpolation to calculate
the free boundary is required.
0.5
0.6
0.7
0.8
0.9
1
1.1
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
P
S
f
r
a
g
r
e
p
l
a
c
e
m
e
n
t
s
S
t
µ
c
increasing
Figure 9.1: The British put option free boundary for varying values of the contract
drift. T = 1, K = 1, σ = 0.4, r = 0.1, D = 0, and µ
c
= 0.11, 0.115, 0.12, . . ., 0.16.
0.65
0.7
0.75
0.8
0.85
0.9
0.95
1
1.05
1.1
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
P
S
f
r
a
g
r
e
p
l
a
c
e
m
e
n
t
s
S
t
σ increasing
Figure 9.2: The British put option free boundary for varying volatilities. T = 1,
K = 1, µ
c
= 0.125, r = 0.1, D = 0, and σ = 0.05, 0.1, . . ., 0.5.
Figure 9.1 shows the location of the British put option free boundary for varying val
ues of the contract drift µ
c
and in addition ﬁgure 9.2 the variation with the volatility
CHAPTER 9. THE BRITISH OPTION 174
(for a ﬁxed µ
c
). Note that some parameters (such as volatility) have been chosen
artiﬁcially high in order to illustrate the distinct behaviour of the British option free
boundaries. It can be seen that as the contract drift parameter increases the free
boundary (and its value at expiry) collapses downwards monotonically to zero. In
fact the exact location to which the free boundary asymptotes and its behaviour close
to expiry is determined in section 9.5. Also the volatility appears to have a signiﬁcant
inﬂuence on the shape of the free boundary.
The most striking diﬀerence with the American option free boundary is that, for some
values of the contract drift at least, the boundary behaviour is no longer monotonic.
This nonmonotonic behaviour can lead to complications in the convergence of the
numerical schemes employed and in addition can result in subtle diﬃculties when
trying to prove the regularity and smoothness of the free boundary (see Peskir and
Samee, 2008b). In what follows, in particular in the asymptotic analysis, regularity
and smoothness shall be implicitly assumed.
Another interesting point to note is that for a given investor at t = 0 with current
stock price S
0
, there exists a value of µ
c
below which all British option free boundaries
at t = 0 lie above the current stock price, i.e. b(0) > S
0
and so the investor is
automatically placed in the exercise region at the initiation of the contract, hence the
optimal investor would choose to exercise immediately.
Finally, numerical investigations showed that varying µ
c
but keeping the diﬀerence
between µ
c
and interest rate r constant resulted in distinct free boundaries. Further
more, varying µ
c
and keeping the ratio of the contract drift to interest rate constant,
i.e.
r
µc
= const., also resulted in diﬀering free boundaries, although in the latter case
all free boundaries asymptoted to the same value at expiry. More information about
the relationship between the parameters could be gleaned from a consideration of
the fully nondimensionalised problem (see appendix B) but this shall be left as the
subject of future research.
CHAPTER 9. THE BRITISH OPTION 175
9.4 Free boundary analysis far from expiry
Far away from expiry, i.e. T −t →∞(eﬀectively the perpetual limit), the numerical
results showed in ﬁgure 9.1 suggest that the free boundary for the British put option
either tends to inﬁnity or tends to zero in this limit, dependent on the choice of the
parameter µ
c
. The following analysis will try to shed some light on these two distinct
regimes of behaviour. First we introduce the following function
H(S, t) = L
_
G(S, t)
_
= G
t
+
1
2
σ
2
S
2
G
SS
+ (r −D)SG
S
−rG, (9.7)
which will be of use in our analysis, since it is a known analytic function which we
shall calculate shortly. It is a standard result from optimal stopping theory that the
free boundary cannot be contained in the region in which H(S, t) > 0.
4
This can
be easily seen by directly applying Itˆ o’s formula to the discounted gain function and
taking expectations to obtain
E
Q
_
e
−r(γ
B
−t)
G(S
γ
B
, γ
B
)
¸
= G(S, t) +E
Q
__
γ
B
−t
0
e
−ru
H(S
t+u
, t +u)du
_
, (9.8)
where γ
B
is deﬁned as
γ
B
= inf¦t ∈ [t, T] : (S
t
, t) / ∈ B¦,
i.e. the ﬁrst exit time of the process from the domain B deﬁned as an arbitrary
small halfcircular domain around the point (S, t). From (9.8) it is clear that the
expected future gain (the lefthandside) must be greater than the current gain G(S, t)
if H(S, t) > 0, indicating it would not be optimal to stop at (S, t) and so the free
boundary cannot be located in the region where H(S, t) > 0. As such the solution to
H(S
h
, t) = 0 will act as an analytical proxy for the free boundary, or at least provide
an upper bound on the location of the free boundary at any given instant. Since this
section is concerned with the large time to expiry behaviour of the free boundary, we
will ultimately be interested in the behaviour of S
h
in this limit. Making extensive
use of the identity,
Φ
(d
2
) =
K
S
e
−(µc−D)(T−t)
Φ
(d
1
), (9.9)
4
See for example Peskir and Shiryaev (2006).
CHAPTER 9. THE BRITISH OPTION 176
where primes denote derivatives, we can directly compute each term in the Hfunction
as
G
t
= −
σKe
−
1
2
d
2
1
2
_
2π(T −t)
+ (µ
c
−D)Se
(µc−D)(T−t)
Φ(d
2
), (9.10)
G
S
= −e
(µc−D)(T−t)
Φ(d
2
),
G
SS
=
Ke
−
1
2
d
2
1
σS
2
_
2π(T −t)
,
which after substitution into (9.7) yields
5
H(S, t) = µ
c
Se
(µc−D)(T−t)
Φ(d
2
) −rKΦ(d
1
). (9.11)
This pleasingly simple expression for the Hfunction can now be used to provide
us with an upper bound on the location of the free boundary. To do this we are
interested in the solution to the equation
µ
c
S
h
e
(µc−D)(T−t)
Φ
_
d
2
(S
h
, t)
_
−rKΦ
_
d
1
(S
h
, t)
_
= 0 (9.12)
for S
h
= S
h
(t), and furthermore in the limit T − t → ∞. It is hypothesised that
if S
h
→ 0 as T − t → ∞ then since the free boundary must lie below S
h
(t) then
the free boundary must also tend to zero as T − t → ∞. Note that the converse
is not necessarily true. Figure 9.3 shows the solution of equation (9.12) obtained
using standard NewtonRaphson iteration. The ﬁrst point to note is that S
h
is not
monotonic for some values of the contract drift µ
c
, the same qualitative behaviour
that can be seen for the true free boundary. It can also be seen that for large enough
values of the contract drift, S
h
(t) appears to tend to zero with no visible turning point,
suggesting a critical value of µ
c
which separates two distinct regions of asymptotic
behaviour of S
h
(t), namely tending to inﬁnity or to zero. This is an important point,
since if we can show under what circumstances the value of S
h
(t) tends to zero for
large times to maturity, we are able to infer that the free boundary must also tend
to zero in that limit.
Guided by numerical diﬀerentiation it appears that the solution for S
h
for large values
5
Note also that the dividends do not occur anywhere but in the exponents.
CHAPTER 9. THE BRITISH OPTION 177
0
0.5
1
1.5
2
0 10 20 30 40 50
P
S
f
r
a
g
r
e
p
l
a
c
e
m
e
n
t
s
S
h
t
µ
c
increasing
Figure 9.3: The zero of the Hfunction, i.e. S
h
(t), for varying values of the contract
drift. µ
c
= 0.102, 0.104, . . . , 1. T = 50, K = 1, r = 0.1, D = 0, and σ = 0.4.
of T −t takes on an exponential form, i.e.
S
h
(t) ∼ Ae
β(T−t)
as T −t →∞. (9.13)
Therefore making this ansatz we can see that
d
1,2
(S
h
, t) =
log
_
K
A
_
−
_
µ
c
−D ±
1
2
σ
2
+ β
_
(T −t)
σ
√
T −t
,
which as T −t →∞ the second term dominates giving
lim
T−t→∞
d
1,2
(S
h
, t) = −
1
σ
_
µ
c
−D ±
1
2
σ
2
+β
_
√
T −t.
The equation for S
h
(t) in the limit T −t →∞ thus becomes
µ
c
Ae
(µc−D+β)(T−t)
Φ
_
−
(µ
c
−D +
1
2
σ
2
+β)
σ
√
T −t
_
= rKΦ
_
−
(µ
c
−D −
1
2
σ
2
+β)
σ
√
T −t
_
.
(9.14)
For convenience in what follows we shall deﬁne
ν
±
=
µ
c
−D ±
1
2
σ
2
+β
σ
.
Now it is clear that in the limit T − t → ∞, the functions d
1,2
will tend to ±∞
depending on the sign of the parameter ν
±
. In order to provide an appropriate
CHAPTER 9. THE BRITISH OPTION 178
balancing of terms in the following analysis we assume that both functions d
1,2
→−∞
in the limit. This can only be achieved if ν
±
> 0, i.e. if
β >
1
2
σ
2
−µ
c
+D > −
1
2
σ
2
−µ
c
+D, (9.15)
a condition which can (and will) be checked a posteriori.
Next we use the well known result (see for example Abramowitz and Stegun, 1968)
that for large negative values of the arguments z, the cumulative normal distribution
function Φ(z) has the following asymptotic expansion
Φ(z) = −
e
−
1
2
z
2
z
√
2π
_
1 −
1
z
2
+
3
z
4
+. . . +
(−1)
n
1.3 . . . (2n −1)
z
2n
+. . .
_
; (9.16)
note that a similar expression exists for large positive z. Equation (9.16) gives to
leading order that
Φ
_
−ν
±
√
T −t
_
=
e
−
1
2
ν
2
±
(T−t)
ν
±
_
2π(T −t)
+. . . ,
therefore equation (9.14) becomes
µ
c
Ae
(µc−D+β)(T−t)
ν
+
_
2π(T −t)
e
−
1
2
ν
2
+
(T−t)
=
rK
ν
−
_
2π(T −t)
e
−
1
2
ν
2
−
(T−t)
+. . . . (9.17)
Finally considering the exponent of the exponentials on the lefthandside we can see
that
µ
c
−D +β −
1
2
ν
2
+
= µ
c
−D +β −
(µ
c
−D +
1
2
σ
2
+β)
2
2σ
2
= −
(µ
c
−D −
1
2
σ
2
+β)
2
2σ
2
= −
1
2
ν
2
−
,
which remarkably allows us to cancel through all exponential terms leaving the much
simpler equation
µ
c
A
ν
+
=
rK
ν
−
. (9.18)
This, however, only provides us with one equation for two unknown constants A and
β. In order to determine their value uniquely we can obtain another equation by
diﬀerentiating (9.12), noting that S
h
is a function of t, to obtain
µ
c
e
(µc−D)(T−t)
_
dS
h
dt
−(µ
c
−D)S
h
_
Φ(d
2
) +KΦ
(d
1
)
_
(µ
c
−r)
∂d
1
∂t
+
σµ
c
2
√
T −t
_
= 0.
CHAPTER 9. THE BRITISH OPTION 179
Once again making the ansatz (9.13) and performing a similar analysis to the above,
we can arrive at the following expression, again provided ν
±
> 0
µ
c
ν
+
(µ
c
−D +β) A =
K
2
_
(µ
c
−r)ν
−
+σµ
c
_
. (9.19)
Now eliminating the unknown A from equations (9.18) and (9.19) leads to the fol
lowing
2r (µ
c
−D +β) = (µ
c
−r)ν
2
−
+σµ
c
ν
−
,
⇒2r (µ
c
−D +β) = (µ
c
−r)
_
µ
c
−D −
1
2
σ
2
+β
_
2
+σµ
c
_
µ
c
−D−
1
2
σ
2
+β
_
.
(9.20)
Finally solving the (quadratic) equation (9.20) for β yields
β =
1
2
σ
2
_
µ
c
+r
µ
c
−r
_
−µ
c
+D (9.21)
where we have taken the positive root in order not to violate the assumption (9.15).
Also note that since
µc+r
µc−r
> 1 for µ
c
> r, the positive root, i.e. equation (9.21) is
consistent with the initial assumption (9.15), adding credence to the obtained result.
Also note that this result is consistent with the fact that as µ
c
→r the free boundary
appears to be tending to inﬁnity at an increasing rate, suggesting that it is always
optimal to early exercise immediately. Equation (9.21) is also consistent with the
observation (as seen in ﬁgure 9.2) that the volatility appears to greatly aﬀect the
behaviour of the free boundary. Finally, substitution of the found value for β, (9.21),
back into equation (9.18) gives immediately that A = K, and so to summarise we
have found that for T −t →∞ the zero of the Hfunction, i.e. S
h
, will behave as
S
h
(t) = Ke
β(T−t)
where β is given by equation (9.21).
The more useful corollary of this result however is that clearly if β > 0 then S
h
(t) →
∞for large values of T −t and conversely if β < 0 then S
h
→0. The critical value of
the parameter µ
c
which separates these two distinct regimes is given by the solution
to the equation
β =
1
2
σ
2
_
µ
∗
c
+r
µ
∗
c
−r
_
−µ
∗
c
+D = 0,
CHAPTER 9. THE BRITISH OPTION 180
hence
µ
∗
c
=
1
2
σ
2
_
µ
∗
c
+r
µ
∗
c
−r
_
+D,
or
µ
∗
c
=
1
2
_
r +D +
1
2
σ
2
_
+
1
2
_
_
r +D +
1
2
σ
2
_
2
+ 4r
_
1
2
σ
2
−D
_
_1
2
. (9.22)
If D <
1
2
σ
2
then we are required to take the positive square root in order to obtain
a positive µ
∗
c
. If D >
1
2
σ
2
then we could have two positive solutions, however only
the positive root will be greater than r. In addition the square root will always stay
positive for all value of D and so we will always have a critical value.
Knowledge about the large time to expiry behaviour of the Hfunction is not only
useful in determining diﬀerent regimes of behaviour of the free boundary, but can also
be used to improve the eﬃciency of the numerical calculations of the option value
and its corresponding freeboundary location. The large time to expiry behaviour
can be used to make an appropriate transformation that removes the observed blow
up to inﬁnity of the free boundary for large T − t. We have shown that in the limit
S
h
(t) takes on the form Ke
β(T−t)
, suggesting that the transformation
ˆ
S = Se
−β(T−t)
(9.23)
may help remove any exponentially growing behaviour. In fact as the function S
h
(t)
provides an upper bound for the location of the free boundary, it is clear that in (
ˆ
S, t)
space, all free boundaries, regardless of the value of the contract drift µ
c
, will tend
to ﬁnite or zero values. Furthermore any numerical treatment of the freeboundary
problem in (
ˆ
S, t)space will be better behaved in the absence of any numerical break
downs.
9.5 Analysis close to expiry
Also of great interest is the behaviour of the free boundary close to expiry. We must
determine the value to which the free boundary asymptotes and also the functional
CHAPTER 9. THE BRITISH OPTION 181
form in which it asymptotes to that value. In addition to being interesting in its own
right, knowledge of the free boundary behaviour close to expiry can be exploited to
improve the eﬃciency of numerical schemes used in determining the free boundary.
For example when using the bodyﬁtted coordinate system described in section 9.3,
providing the correct location of the free boundary at expiry can be crucial in the
scheme’s success.
When investigating the small time to expiry behaviour of free boundaries arising from
derivative securities with earlyexercise features, there are a multitude of approaches
that can be taken. For example the analysis of Kuske and Keller (1998), which in
vestigates the standard American put option, exploits the Green’s function for the
heat equation to convert the resulting boundary value problem to an integral equa
tion, which is then solved asymptotically for times close to expiry. Alternatively the
diﬀerential form of the freeboundary problem can be tackled directly and matched
asymptotic expansions can be used to investigate the solution behaviour close to
expiry, for example see Wilmott et al. (1995), Johnson (2007) or the recent survey
article by Howison (2005).
The value to which the British option free boundary asymptotes can be obtained
simply via a cursory inspection of the Hfunction deﬁned in the previous section and
given by (9.11). Recall that the region in which H(S, t) > 0 cannot contain the free
boundary. Using this information as t → T we have that the Hfunction is trivially
zero for S > K, and for S < K is given by
H(S, T) = µ
c
S −rK.
Hence we have that
S
h
(T) =
rK
µ
c
,
and that the Hfunction is positive in the region S >
rK
µc
and so we can conclude that
the free boundary must lie at or below this value. In order to convince ourselves that
it should in fact coincide with S
h
(T) we consider a situation in which the process is
an arbitrarily small time away from expiry and that the current price is below the
CHAPTER 9. THE BRITISH OPTION 182
value
rK
µc
. In this region H(S, t) < 0 and so the investor is accumulating negative
gain, with no possibility of the process reaching the region in which H(S, t) > 0,
since the process is extremely close to expiry; hence the optimal investor would stop
in this region leading to the conclusion that
b(T) =
rK
µ
c
. (9.24)
Note that the inclusion of dividends does not aﬀect the location to which the free
boundary asymptotes as one might expect; this fact is easily conﬁrmed numerically.
The above result can be shown using a diﬀerent approach which we shall brieﬂy
outline below. One of the hallmarks of the existence of an earlyexercise region is
the existence of a region in which the value of the earlyexercise option’s European
counterpart (i.e. with no early exercise) lies below the payoﬀ function (gain function).
Close to expiry this amounts to determining if the dynamics of the PDE move the
option price below the gain function at a small time prior to expiry. More speciﬁcally
if the quantity
∂
∂t
_
V (S, T −δt) −G(S, T −δt)
_
ever becomes positive. A simple Taylor expansion leads to
∂
∂t
_
V (S, T −δt) −G(S, T −δt)
_
≈ V (S, T) −G(S, T) +δt
_
∂V
∂t
(S, T) −
∂G
∂t
(S, T)
_
.
By deﬁnition V (S, T) = G(S, T) and we can also directly evaluate the expression for
G
t
, namely (9.10), at t = T. In addition rearranging the governing PDE directly to
obtain an expression for V
t
and using the fact that V (S, T) = (K −S)
+
leads to
∂
∂t
_
V (S, T −δt) −G(S, T −δt)
_
≈ δt(rK −µ
c
S)
for S < K and trivially zero otherwise. Hence there exists a (possible) earlyexercise
region when
rK −µ
c
S > 0 ⇒S <
rK
µ
c
,
in agreement with (9.24).
CHAPTER 9. THE BRITISH OPTION 183
However, of more interest is the functional form in which the free boundary ap
proaches (9.24). It transpires that the answer is pleasingly simple, the leadingorder
behaviour is identical to that of the standard American put option with dividends,
the proof of which is given below.
If we make a subtle change of variable, namely
ˆ
S = Se
(µc−D)(T−t)
, (9.25)
then the governing PDE becomes
V
t
+
1
2
σ
2
ˆ
S
2
V
ˆ
S
ˆ
S
+ (r −µ
c
)
ˆ
SV
ˆ
S
−rV = 0,
and the gain function is transformed to
G(
ˆ
S, t) = KΦ
_
d
1
(
ˆ
S, t)
_
−
ˆ
SΦ
_
d
2
(
ˆ
S, t)
_
,
with
d
1
(
ˆ
S, t) =
log
_
K
ˆ
S
_
+
1
2
σ
2
(T −t)
σ
√
T −t
,
d
2
(
ˆ
S, t) = d
1
(
ˆ
S, t) −σ
√
T −t.
Notice that the parameter µ
c
has been taken out of the gain function (and hence
the boundary conditions) and placed into the PDE, where it appears as a pseudo
dividend. Also note that the actual dividend yield D has been completely removed
from the problem by the scaling (9.25), indicating that dividends have a rather benign
aﬀect on the dynamics of the British put option.
Since we are interested in the behaviour of the free boundary as we approach expiry,
we are therefore interested in the behaviour of the gain function as we approach
expiry, i.e. for small values of T − t. The ﬁrst point to note is that in this limit
d
2
→d
1
and so it will suﬃce to determine the behaviour of d
1
in this limit. It is clear
that as T −t →0 (i.e. t →T) the ﬁrst term in d
1
(
ˆ
S, t) will dominate giving
lim
t→T
d
1
(
ˆ
S, t) = lim
t→T
1
σ
√
T −t
log
_
K
ˆ
S
_
.
CHAPTER 9. THE BRITISH OPTION 184
Since log
_
K
ˆ
S
_
< 0 for
ˆ
S > K, then in this region it is clear that d
1
→−∞ as t →T
and conversely for
ˆ
S < K then d
1
→+∞as t →T. In the region
ˆ
S > K if d
1
→−∞
then we can exploit the asymptotic expansion of the cumulative normal distribution
function (9.16) to give
Φ(d
1
) = −
e
−
1
2
d
2
1
d
1
√
2π
+. . . ,
⇒Φ(d
1
) = −σ
_
T −t
2π
_
log
_
K
ˆ
S
__
−1
exp
_
−
1
2σ
2
(T −t)
_
log
_
K
ˆ
S
__
2
_
+. . . .
Similarly for the region
ˆ
S < K where d
1
→+∞ we can use the symmetry of Φ() to
obtain
Φ(d
1
) = 1 −
e
−
1
2
d
2
1
d
1
√
2π
+. . . ,
⇒Φ(d
1
) = 1 −σ
_
T −t
2π
_
log
_
K
ˆ
S
__
−1
exp
_
−
1
2σ
2
(T −t)
_
log
_
K
ˆ
S
__
2
_
+. . . .
Using the above and further the fact that Φ(d
1
) →Φ(d
2
) as t →T, the gain function
can be approximated by
G(
ˆ
S, t) = (K−
ˆ
S)
+
−σ
_
T −t
2π
_
log
_
K
ˆ
S
__
−1
exp
_
−
1
2σ
2
(T −t)
_
log
_
K
ˆ
S
__
2
_
(K−
ˆ
S)+. . .
(9.26)
in the region close to expiry. Note that the secondorder term in (9.26) decays as
(T − t)
1
2
e
−
A
T−t
in the limit t → T, where A must be a positive constant. Clearly
this decays much faster than the terms arising from the asymptotic (power series)
expansion of the PDE and consequently the gain function can be approximated to
6
G(
ˆ
S, t) =
_
K −
ˆ
S
_
+
,
i.e. the standard put option payoﬀ.
Hence to leading order as t → T we have that the British put option freeboundary
problem (9.3), under the transformation (9.25), is identical to the standard American
put option freeboundary problem with dividends, where the dividend amount (yield)
6
Note that there is a breakdown if S < Ke
−B
√
T−t
or S > Ke
+B
√
T−t
, which as t →T tend to 0
and ∞ respectively, such that these breakdowns do not interfere with the free boundary at expiry.
CHAPTER 9. THE BRITISH OPTION 185
is given by the contract drift rate µ
c
. Hence the behaviour of the British put option
free boundary will coincide with that of the American Put option with dividends.
Fortuitously the asymptotic form of the American put with a constant dividend yield
close to expiry has been studied extensively, see for example Evans et al. (2002), and
the leading order behaviour is given by
b(t) ∼
rK
ˆ
D
_
1 −σα
0
_
2(T −t)
_
for constant dividend yield
ˆ
D > r, where the constant α
0
is the solution to the
transcendental equation
α
3
0
e
α
2
0
_
∞
α
0
e
−u
2
du =
1
4
_
2α
2
0
−1
_
,
which can be calculated to give α
0
≈ 0.4517. Therefore the asymptotic form of the
British put free boundary, after transforming back to the original variables via (9.25),
is given by
b(t) ∼
rK
µ
c
_
1 −σα
0
_
2(T −t)
_
e
−(µc−D)(T−t)
, (9.27)
i.e. the free boundary approaches expiry parabolically. Note, however, that to lead
ing order the behaviour of the free boundary remains unchanged by the transform
(9.25). Figure 9.4 shows the above approximation compared with the free boundary
obtained via a full numerical treatment using the techniques described in section 9.3.
In addition, numerical investigations performed by the author indicate that (9.27)
becomes a better approximation (over the same scale) as the volatility is decreased.
This is most likely due to the fact that for larger σ the turning point of the free bound
ary occurs closer to expiry and the free boundary would diverge from the parabolic
approximation over a shorter timescale.
Note that if we are interested in higherorder asymptotics then the behaviour will
diﬀer to that of the American put with dividends. This is due to the various terms
we have neglected as t → T from the gain function and also from the exponential
transform. Also note that for the American put the free boundary takes on a very
diﬀerent functional form for diﬀerent values of the dividend parameter, i.e. for
ˆ
D = r
CHAPTER 9. THE BRITISH OPTION 186
0.775
0.78
0.785
0.79
0.795
0.8
0 0.002 0.004 0.006 0.008 0.01
P
S
f
r
a
g
r
e
p
l
a
c
e
m
e
n
t
s
b
(
t
)
t
Figure 9.4: The asymptotic approximation for the British put option free boundary
close to expiry, i.e. (9.27) (dotted line) compared with fully numerical value (solid
line). T = 0.01, σ = 0.4, r = 0.1, µ
c
= 0.125, and D = 0.
or
ˆ
D < r, however for the British put option we have the natural restriction that
µ
c
> r and so the parabolic regime is the only one of interest. Furthermore, the
value of the dividend D does not qualitatively change the asymptotic behaviour of
the British put free boundary since the scaling (9.25) has removed the parameter
completely from the system.
9.6 Financial analysis of the British put option
The protection feature of the British option that was so crucial in motivating its in
troduction led Peskir and Samee (2008a,b) to investigate further the motivations of a
British option investor, in particular the potential return on investment in the British
(and other) options. Here we reproduce this analysis in more detail by exploiting the
advanced numerical techniques employed to obtain relativereturn surfaces of the op
tions. Any speculator (as opposed to a hedger) who chooses to invest in a putstyle
option will inevitably hold the view that the underlying will experience downward
price movements at some time in the future, or more speciﬁcally that the realised
drift will be less than the risk neutral rate r. Indeed, it was with these very traders
CHAPTER 9. THE BRITISH OPTION 187
in mind that the British option was developed. Furthermore, as noted by Peskir
and Samee (2008a,b), the British option provides an instrument that can be used to
‘milk’ proﬁts from the speculators view of future price movements, whilst maintaining
a protection feature should these price movements not transpire.
To further illustrate possible situations in which a British option might be favourable
over other options available in the market, let us construct an idealised market con
sisting solely of a European, American and British put option written on the same
underlying, with the same expiry date, say one year and strike price K which for
simplicity we shall assume is equal to the current stock price and scaled to unity, i.e.
S
0
= K = 1 and hence all options are ‘at the money’. Now an investor in such a
market has the choice of investing in either of the three options and as such has to
pay the corresponding option price. Of great interest from the point of view of such
an investor is the expected return on their investment for all possible future stock
prices at any point in time prior to maturity.
The rational strategy of such an investor would be to choose the option that has
the greatest return under the future price movements that are most inline with the
sentiment of that particular investor. For example, if an investor believes that the
stock price will fall by approximately 20% in the next 6 to 9 months, then which
available option will provide the greatest return on the price paid for the option,
deﬁned as
R(S, t) =
money received at (S, t)
option price paid at t = 0
100%.
The idea of this section is to produce a return surface for all the available options in
the market, given the current stock price S
0
. The option price of all three options
is easily calculated at t = 0 given the current price S
0
, but there is some ambiguity
in the value of the money received at a future time. For a European option we are
unable to exercise the option until expiry and so there can be no payout of the option
at any time prior to maturity. Alternatively for an American option we can choose
to exercise at any time prior to maturity and receive the payoﬀ (K −S
t
)
+
, however
any rational investor would never choose to exercise the option ‘out of the money’,
CHAPTER 9. THE BRITISH OPTION 188
i.e. if S
t
> K for any time before maturity. However, despite the option having a
zero gain function in such a situation, it still has some intrinsic value, precisely the
noarbitrage price of the American option at that particular price and time. For this
reason the most consistent measure of the expected return on an American option is
given by
R
A
(S, t) =
max
_
(K −S)
+
, V
A
(S, t)
_
V
A
(S
0
, 0)
100%,
hence we are allowing the investor to sell the option in the market and receive the
current ‘noarbitrage’ value of the option in addition to allowing for the exercise of
the option upon which the payoﬀ is received. Note that this assumption assumes that
the investor has access to the market in order to sell and that he incurs no transaction
or liquidation costs. In some sense this corresponds to a ‘best case’ scenario for the
American and European option, since if the investor is exposed to market frictions
then they will inevitably obtain a worse price than the noarbitrage values used in
the above calculations. The British put on the other hand will not incur such costs
(in the exercise region) as there is no need for the investor to enter the market. Also
note that for these purposes we are assuming the investor to behave rationally in the
sense that the option should be exercised at a stock price in the rational exercise
region and to sell the option at a stock price in the continuation region. Similar
assumptions allow us to deﬁne the return on the British option as
R
B
(S, t) =
max
_
G
B
(S, t), V
B
(S, t)
_
V
B
(S
0
, 0)
100%,
and the return on the European option as
R
E
(S, t) =
V
E
(S, t)
V
E
(S
0
, 0)
100%,
hence the European option investor’s only choice is to sell the option at times prior
to maturity.
We note that at maturity the return on all three investments are not equal, the money
received will be the same, i.e. (K−S
T
)
+
, but that the initial price paid for the option
will have varied, more speciﬁcally we must have V
A
(S
0
, 0) > V
B
(S
0
, 0) > V
E
(S
0
, 0).
CHAPTER 9. THE BRITISH OPTION 189
This is natural since if an investor waits until maturity to exercise an earlyexercise
option, then they have ‘wasted’ the earlyexercise premium priced into the American
and British option prices. Figure 9.6 shows the diﬀerence in returns of the British
put option and the American put option, i.e. R
B
(S, t) −R
A
(S, t), for a contract drift
of µ
c
= 0.125. This surface indicates the diﬀerences in the returns (as a percentage
of the initial investment) should the investor close out their position by exercising or
selling, whichever provides the greatest return (or is permitted). Similarly ﬁgure 9.7
compares the return of the British put to the European put and ﬁnally for reference
ﬁgure 9.8 compares the American and European put option. For comparison the
rational exercise boundaries can be found in ﬁgure 9.5.
0.65
0.7
0.75
0.8
0.85
0.9
0.95
1
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
P
S
f
r
a
g
r
e
p
l
a
c
e
m
e
n
t
s
b
(
t
)
t
µ
c
= 0.125
µ
c
=
0
.
1
3
5
Figure 9.5: Location of the free boundary for the British (solid line) and American
(dotted line) put option under investigation in ﬁgures 9.6, 9.7 and 9.8. T = 1, K = 1,
σ = 0.4, r = 0.1, µ
c
= 0.125, and D = 0.
The most striking feature of ﬁgure 9.6 is that the British put option appears to be
providing a greater expected return in the majority of the stopping regions. The
American option only provides the investor with a better return on their investment
if it transpires that the investor chooses to stop and close out their position in a
‘wedge’ shaped region below the current stock price and extending just beyond half
way to maturity. Even then the diﬀerence in the American put return is no greater
than 20%, whereas in some regions the British put option can provide up to 60%
CHAPTER 9. THE BRITISH OPTION 190
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
20
0
20
40
60
80
P
S
f
r
a
g
r
e
p
l
a
c
e
m
e
n
t
s
t
S
Figure 9.6: The diﬀerence in the percentage return of the British put option and
the American put option at every possible stopping location. The solid lines denote
contours at increments of 10% from 10% to 60%. The dotted line represents the zero
contour. S
0
= 1, T = 1, K = 1, σ = 0.4, r = 0.1, D = 0, and µ
c
= 0.125.
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
20
0
20
40
60
80
P
S
f
r
a
g
r
e
p
l
a
c
e
m
e
n
t
s
t
S
Figure 9.7: The diﬀerence in the percentage return of the British put option and the
European put option. Again the solid lines denote contours at increments of 10%
from 0% to 70%. The dotted line represents the zero contour. S
0
= 1, T = 1, K = 1,
σ = 0.4, r = 0.1, D = 0, and µ
c
= 0.125.
CHAPTER 9. THE BRITISH OPTION 191
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
80
60
40
20
0
20
40
P
S
f
r
a
g
r
e
p
l
a
c
e
m
e
n
t
s
t
S
Figure 9.8: The diﬀerence in the percentage return of the American put option and
the European put option. The solid lines denote contours at increments of 10% from
70% to 30%. The dotted line represents the zero contour. S
0
= 1, T = 1, K = 1,
σ = 0.4, r = 0.1, and D = 0. Note the change of orientation.
extra return over the American put option (although these regions are many stan
dard deviations away from the current price of the underlying). More importantly,
numerical investigations have shown that the return proﬁle seen in ﬁgure 9.6 and
indeed the region in which the expected return on the American is greater than the
British put option remains relatively constant in shape and size for varying values of
the contract drift µ
c
. Note that the above observations are in total agreement with
Peskir and Samee (2008a,b).
When comparing the British with the corresponding European put option (ﬁgure
9.7) the return diﬀerentials are the greatest for low values of the stock price and large
times to expiry, as we might expect. However it can be seen that rather surprisingly
the British put option will provide a greater return for the vast majority of stopping
locations below strike K; with the only exception being relatively close to maturity
where the European option would provide (at most) a 10% greater return.
Figure 9.9 attempts to encapsulate all of the above observations into schematic form.
It plots the regions in which the ordering of the expected returns (given any stopping
CHAPTER 9. THE BRITISH OPTION 192
0
0.2
0.4
0.6
0.8
1
1.2
1.4
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
P
S
f
r
a
g
r
e
p
l
a
c
e
m
e
n
t
s
R
B
> R
A
> R
E
R
B
> R
E
> R
A
R
A
> R
B
> R
E
R
E
> R
B
> R
A
t
S
Figure 9.9: Schematic representation of the regions in which atthemoney European,
American and British put option would provide the greatest return on an investment.
The dotted lines represent the free boundaries of the American and British put option
for reference. T = 1, K = 1, σ = 0.4, r = 0.1 and D = 0.
location) for the British, American and European put option are diﬀerent.
7
The free
boundary location of the British and American put have been included for reference.
It appears that there is a region, centred around the American earlyexercise bound
ary in which the American option would provide a greater return on the initial invest
ment than the other two options. This region gets smaller as maturity approaches
and disappears at approximately threeﬁfths of the way to maturity.
This also highlights what could be described as an ‘unexpected’ additional beneﬁt of
the British option, which is somewhat counter to the protection feature for which the
option was originally created. If the stock price falls sharply, then the British option is
able to ‘milk’ more money out of such a situation than the holder of the corresponding
American option. In fact any stopping location below strike in the second half of the
contract will inevitably result in a higher return for the British option holder, over the
American option holder. Considering this fact, in conjunction with the the empirical
observation that the majority of American options are exercised in the second half of
7
The three curves correspond to the zero contours of the three ﬁgures 9.6, 9.7, and 9.8 obtained
for the parameter value µ
c
= 0.125, although the same qualitative behaviour is seen for other values.
CHAPTER 9. THE BRITISH OPTION 193
the contract’s term,
8
the British option can be considered an attractive alternative.
In addition, if the options are far out of the money, i.e. S
t
¸K then the European
option would provide the greatest return, however the values of all three options are
so small in this region that any comparison becomes a purely academic issue.
Finally, ﬁgure 9.10 shows how the location of the wedgeshaped region (in which the
American put option provides greater returns over the British put option) changes
as the initial value of the stock price is varied, whilst keeping the strike constant. In
other words as we increase the moneyness of the option. Figure 9.10(a) shows that
as we increase the moneyness by decreasing the ratio S
0
/K from 1 to 0.7 in steps of
0.1 this region becomes progressively smaller until it actually disappears for a ratio
of 0.6, at which point the British option is guaranteed to provide a greater return
on an investment. Figure 9.10(b) shows the shape of this region when we continue
to increase the moneyness by changing the ratio from 0.5 down to 0.2. In this case
the region grows steadily, however this time any immediate downwards stock price
movements will result in the British option providing the greatest return, whereas
an upward movement would be better served by an American put option. This is
the opposite of the behaviour seen in ﬁgure 9.10(a). It should be noted that for the
majority of the values of S
0
/K presented in ﬁgure 9.10 the option holder is placed
immediately in the exercise region, and so should ‘optimally’ exercise. However, the
comparisons are still of interest, since the investor need not follow the optimally
strategy.
9.7 The British call option
We now turn our attention to the British call option (cf. Peskir and Samee, 2008a).
Unlike the American call option (without dividends) the British call option (with
or without dividends) no longer has a trivial solution; there exists an earlyexercise
8
See for example Diz and Finucane (1993), who investigate the earlyexercise behaviour of options
on the the S&P 100 index. They show that over 82% of all call options and 77% of all put options
that are exercised are done so during the ﬁnal week before maturity (inclusive of maturity).
CHAPTER 9. THE BRITISH OPTION 194
0
0.2
0.4
0.6
0.8
1
1.2
1.4
0 0.2 0.4 0.6 0.8 1
P
S
f
r
a
g
r
e
p
l
a
c
e
m
e
n
t
s
t
S
Increasing Moneyness
(a)
S0
K
= 1, 0.9, 0.8 and 0.7.
0
0.2
0.4
0.6
0.8
1
1.2
1.4
0 0.2 0.4 0.6 0.8 1
P
S
f
r
a
g
r
e
p
l
a
c
e
m
e
n
t
s
t
S
Increasing Moneyness
(b)
S0
K
= 0.5, . . ., 0.2.
Figure 9.10: Figures representing the region in which American put options would
provide a greater expected return that its British option counterpart, for increasing
moneyness. T = 1, K = 1, σ = 0.4, r = 0.1 and D = 0.
boundary for all µ
c
< r. The noarbitrage pricing procedure is identical to that of the
British put described in section 9.2 and in fact leads to the following freeboundary
problem representation of the British call option price:
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
V
t
+
1
2
σ
2
S
2
V
SS
+ (r −D)SV
S
−rV = 0 for S ∈ (0, b(t)],
V (S, t) = G(S, t) on S = b(t)
V
S
(S, t) = G
S
(S, t) on S = b(t) (smooth ﬁt)
V > G in (
V = G in T
(9.28)
where now the gain function is given by
G
C
(S, t) = K
_
Φ
_
d
1
(S, t; µ
c
)
_
−1
¸
−Se
(µc−D)(T−t)
_
Φ
_
d
2
(S, t; µ
c
)
_
−1
¸
(9.29)
where all of the notation is as previously deﬁned. Note the relationship between the
call and put option gain functions
G
C
(S, t) = Se
(µc−D)(T−t)
−K +G
P
(S, t)
where G
C
and G
P
denote the gain functions of the British call and put respectively.
9
Again we can note that the gain function approaches the standard call option payoﬀ
as t →T, i.e.
G(S, T) = (S −K)
+
.
9
Notice the similarity of this expression with the putcall parity relationship described in section
4.1, i.e. V
C
E
= Se
−D(T−t)
−Ke
−r(T−t)
+V
P
E
.
CHAPTER 9. THE BRITISH OPTION 195
Figure 9.11 shows the variation of the free boundary for varying values of the contract
drift µ
c
showing again a monotonic collapse of the free boundary as µ
c
is increased
to r. Again note that the free boundary of the British call option, like the British
put, is no longer monotonic. Figure 9.12 also shows its variation with the volatility
parameter σ, again showing that the nature of the free boundary is highly dependent
the volatility of the underlying stock.
0.8
1
1.2
1.4
1.6
1.8
2
2.2
2.4
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
P
S
f
r
a
g
r
e
p
l
a
c
e
m
e
n
t
s
b
(
t
)
t
µ
c
increasing
Figure 9.11: The British call option free boundary for varying values of the contract
drift. T = 1, K = 1, σ = 0.4, r = 0.1, D = 0, and µ
c
= 0.05, 0.055, 0.06, . . ., 0.09.
0.8
0.9
1
1.1
1.2
1.3
1.4
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
P
S
f
r
a
g
r
e
p
l
a
c
e
m
e
n
t
s
b
(
t
)
t
σ increasing
Figure 9.12: The British call option free boundary for varying volatilities. T = 1,
K = 1, µ
c
= 0.08, r = 0.1, D = 0, and σ = 0.05, 0.1, . . ., 0.5.
CHAPTER 9. THE BRITISH OPTION 196
9.7.1 Analysis far from expiry
For the British call we can apply the same analysis as for the British put in order to
determine the large time from expiry behaviour of the free boundary. For the British
call the Hfunction can be shown to be
H(S, t) = µ
c
Se
(µc−D)(T−t)
[Φ(d
2
) −1] −rK [Φ(d
1
) −1] , (9.30)
and hence we are interested in the large T−t behaviour of the solution to H(S
h
, t) = 0,
hence
µ
c
S
h
e
(µc−D)(T−t)
_
Φ
_
d
2
(S
h
, t)
_
−1
¸
−rK
_
Φ
_
d
1
(S
h
, t)
_
−1
¸
= 0. (9.31)
Again making the ansatz that the solution for S
h
(t) behaves as
S
h
(t) =
¯
Ae
¯
β(T−t)
as t →T we wish to solve for the constants
¯
A and
¯
β. Substitution into the equation
(9.30) yields
µ
c
¯
Ae
(µc−D+
¯
β)(T−t)
_
Φ
_
−¯ ν
+
√
T −t
_
−1
_
= rK
_
Φ
_
−¯ ν
−
√
T −t
_
−1
_
,
where we have deﬁned
¯ ν
±
=
µ
c
−D ±
1
2
σ
2
+
¯
β
σ
for convenience of algebra. We wish to look at the asymptotic behaviour as T −t →
∞ and it transpires that there are three diﬀerent regimes in which the solution is
quantitatively diﬀerent. If ¯ ν
±
> 0 then both the cumulative normal distributions
tend to zero in this limit and the leading order terms become
µ
c
¯
Ae
(µc−D+
¯
β)(T−t)
= rK +. . . ,
and since in this regime
¯
β >
1
2
σ
2
− µ
c
+ D, the exponent of the exponential term
must be positive, indicating that there can be no solution to this equation for
¯
β in the
limit, hence our ansatz for the form of the solution does not permit
¯
β >
1
2
σ
2
−µ
c
+D.
The second regime corresponds to a situation when ¯ ν
+
> 0 and ¯ ν
−
< 0, i.e. when
−
1
2
σ
2
−µ
c
+D <
¯
β <
1
2
σ
2
−µ
c
+D ⇒−
1
2
σ
2
<
¯
β <
1
2
σ
2
;
CHAPTER 9. THE BRITISH OPTION 197
in this regime the ﬁrst cumulative normal distribution function tends to zero but the
second tends to unity in the limit, hence the leadingorder terms become
µ
c
¯
Ae
(µc−D+
¯
β)(T−t)
=
rKe
−
1
2
¯ ν
2
−
(T−t)
¯ ν
−
_
2π(T −t)
+. . . ,
where we have used the approximations for Φ() introduced earlier. Again it is clear
that there can be no balancing in this regime as T −t →∞, and so there can be no
solution for
¯
β. Finally we have that ¯ ν
±
< 0, in this regime the leadingorder terms
become
µ
c
¯
Ae
(µc−D+
¯
β)(T−t)
¯ ν
+
_
2π(T −t)
e
−
1
2
¯ ν
2
+
(T−t)
=
rK
¯ ν
−
_
2π(T −t)
e
−
1
2
¯ ν
2
−
(T−t)
+. . . ,
which is identical to equation (9.17). Furthermore the large T − t behaviour of the
time derivative of equation (9.31) under the assumption that ¯ ν
±
< 0 is identical to
(9.19) and so it is is clear that a balancing of terms does exist for the call option in
the limit T −t →∞ and moreover that
¯
β will be given by
¯
β =
1
2
σ
2
_
µ
c
+r
µ
c
−r
_
−µ
c
+D,
i.e. the same behaviour as for the British put. Recall that for a British call we must
have that µ
c
< r and all parameters must be positive so we have that
µ
c
+r
µ
c
−r
< −1,
which conﬁrms our original assumption
¯
β < −
1
2
σ
2
−µ
c
+D, i.e. that ¯ ν
±
< 0.
An interesting corollary to this result is that when the dividend yield D is zero, we
have that
¯
β < 0 for all possible values of µ
c
, hence the function S
h
(t) for the British
call must always decay to zero for large times to expiry in this regime. However this
is not helpful from the viewpoint of determining the behaviour of the free boundary
in this limit, since for the British call option the Hfunction is positive below S
h
(t),
and so in this regime we cannot say anything about the freeboundary behaviour for
large times to maturity. On the other hand, if D ,= 0 then the function S
h
(t) will
tend to inﬁnity (and so must the free boundary) provided that the dividend is greater
than some critical value, i.e.
D > D
crit
= µ
c
+
1
2
σ
2
_
r + µ
c
r −µ
c
_
.
CHAPTER 9. THE BRITISH OPTION 198
If the dividend is greater than this value, the function S
h
(t) can show positive expo
nential growth for large times to expiry, more speciﬁcally if the contract drift is less
than the critical value µ
∗
c
given by (9.22).
9.7.2 Analysis close to expiry
The close to expiry analysis for the British call option is identical to that of the
British put option. Using the same transformation we can transform the British call
option to the standard American call option problem with dividends (in the limit
t →T). Again the results of Evans et al. (2002) state that the American call option
with dividends behaves as
b(t) ∼
rK
ˆ
D
_
1 +σα
0
_
2(T −t)
_
,
where α
0
is as before. Hence the British call free boundary behaves as
b(t) ∼
rK
µ
c
_
1 +σα
0
_
2(T −t)
_
e
−(µc−D)(T−t)
. (9.32)
Figure 9.13 once again shows the accurately of the approximation (9.32) when com
pared with the fully numerical free boundary.
So far we have been unable to determine the large time to expiry behaviour of the
British option free boundary directly. The best we have done is to use the behaviour
of the Hfunction in this limit as an analytical proxy of the free boundary and infer
that the free boundary must tend to zero in certain parameter regimes. As a step
to determining the asymptotic behaviour of the free boundary itself, the following
considerations may prove to be useful.
9.8 Integral representations of the free boundary
9.8.1 The American put option
The value of an American option can be written using the socalled earlyexercise
premium representation, due to Kim (1990), Jacka (1991) and Carr et al. (1992)
CHAPTER 9. THE BRITISH OPTION 199
1.25
1.255
1.26
1.265
1.27
1.275
1.28
1.285
0 0.002 0.004 0.006 0.008 0.01
P
S
f
r
a
g
r
e
p
l
a
c
e
m
e
n
t
s
b
(
t
)
t
Figure 9.13: The asymptotic approximation for the British call option free boundary
close to expiry, i.e. (9.32) (dotted line) compared with fully numerical value (solid
line). T = 0.01, K = 1, σ = 0.4, r = 0.1, µ
c
= 0.08 and D = 0.
amongst others; for a full exposition of this representation (including existence and
uniqueness results) see Peskir and Shiryaev (2006). This representation is given by
V (S, t) = E
Q
S,t
_
e
−r(T−t)
G(S
T
, T)
¸
+rK
_
T−t
0
e
−ru
P
Q
S,t
[S
t+u
≤ b(t +u)] du, (9.33)
where G(S
T
, T) is the gain function of the American option, i.e. G(S
t
, t) = (K−S
t
)
+
for a put, and P
Q
S,t
[S
t+u
≤ b(t +u)] is the probability that the process is below the
free boundary at time t + u (conditional on the information available up to time
t). Identifying the ﬁrst term in the equation as the European put value (without
any early exercise), the second term can be seen intuitively as the extra value of the
option due to the ability to exercise early. An explicit expression for the probability
in the integral is well known and furthermore is derived in appendix C and using this
expression (with D = 0 for simplicity) reduces the above representation to
V (S, t) = V
E
(S, t) +rK
_
T−t
0
e
−ru
Φ
_
_
log
_
b(t+u)
S
_
−
_
r −
1
2
σ
2
_
u
σ
√
u
_
_
du. (9.34)
Note that in order to evaluate the integral above and hence determine the option
value we require knowledge of the location of the free boundary b(t). To determine
the location of the free boundary we evaluate equation (9.34) at S = b(t) for which
CHAPTER 9. THE BRITISH OPTION 200
we know the value of the option (by deﬁnition) must be equal to K − b(t). This
leads to the socalled freeboundary equation which completely characterises the free
boundary
K −b(t) = V
E
_
b(t), t
_
+rK
_
T−t
0
e
−ru
Φ
_
_
log
_
b(t+u)
b(t)
_
−
_
r −
1
2
σ
2
_
u
σ
√
u
_
_
du. (9.35)
Note that this is a Volterra integral equation (of the second type) and solving this
(implicit) equation for b(t) will give the location of the free boundary. To illustrate
that this equation does indeed lead to the location of the free boundary, we shall
consider the simple case of a perpetual American put option, hence we are interested
in the behaviour of equation (9.35) in the limit T −t →∞.
Firstly we note that in the limit T −t →∞ the value of a European option trivially
tends to zero (due to discounting). Setting V
E
(b(t), t) = 0 still leaves an implicit
equation for b(t), however it can be reduced to an explicit equation by exploiting the
fact that the American free boundary at large times to expiry tends to a constant
value. Therefore we would expect the ratio b(t + u)/b(t) to be equal to one. for all
values of u ∈ [0, ∞), since if we are at time t then the free boundary at any time in
the future will be the same as it is now. Hence
lim
T−t→∞
log
_
b(t +u)
b(t)
_
= 0
This is eﬀectively the same as making the assumption (or ansatz) that the free bound
ary to be found is a constant, b(t) = b
∞
say. The integral representation thus reduces
to
K −b(t) = rK
_
∞
0
e
−ru
Φ
_
(σ
2
−2r)
√
u
2σ
_
du.
This integral can be solved (with a little work) by ﬁrstly setting s = k
1
√
u where
k
1
=
σ
2
−2r
2σ
leading to
K −b(t) =
2rK
k
2
1
_
∞
0
se
−
rs
2
k
2
1
Φ(s)ds,
and further integrating by parts yields
K −b(t) =
2rK
k
2
1
_
k
2
1
4r
+
k
2
1
2r
√
2π
_
∞
0
e
−
1
2
+
r
k
2
1
s
2
ds
_
.
CHAPTER 9. THE BRITISH OPTION 201
Finally setting k
2
=
1
2
+
r
k
2
1
and using the identity
_
∞
0
e
−k
2
s
2
ds =
1
2
_
π
k
2
,
we arrive at
b(t) =
K
2
_
1 −
1
√
2k
2
_
.
In terms of the original parameters we can see that
k
2
=
1
2
_
σ
2
+ 2r
σ
2
−2r
_
2
,
and so substitution gives the location of the free boundary as
b(t) =
2rK
2r +σ
2
=
αK
α + 1
,
where α =
2r
σ
2
. This agrees exactly with the well known value of the perpetual
American put free boundary, equation (6.3), found in section 6.1.
9.8.2 The British put option
Analogous with the American option, the optimal stopping boundary of the British
option can be characterised as the unique solution of a nonlinear Volterra integral
equation of the second type (cf. Peskir and Samee, 2008a,b). In order to show this
we can apply Itˆ o’s formula to the discounted option value to obtain
e
−rs
V (S
t+s
, t +s) = V (S, t) +
_
s
0
L
St
_
e
−ru
V (S
t+u
, t +u)
_
du +M
t
,
where M
t
is a local martingale, L
St
is the inﬁnitesimal generator of the process deﬁned
by
L
St
=
∂
∂u
+
1
2
σ
2
S
2
∂
2
∂S
2
+ (r −D)S
∂
∂S
,
Applying the operator to the discounted process we can simplify the above expression
to
e
−rs
V (S
t+s
, t +s) = V (S, t) +
_
s
0
e
−ru
(L
St
−r) V (S
t+u
, t +u)du +M
t
= V (S, t) +
_
s
0
e
−ru
L
_
V (S
t+u
, t +u)
_
du +M
t
,
CHAPTER 9. THE BRITISH OPTION 202
where we have used the fact that (L
St
−r) V (S, t) = L
_
V (S, t)
_
where L is the
BlackScholes diﬀerential operator. The next step is to take expectations giving
E
Q
_
e
−rs
V (S
t+s
, t +s)
¸
= V (S, t) +
_
s
0
e
−ru
E
Q
_
L
_
V (S
t+u
, t +u)
_¸
du,
where we have taken the expectation under the integral sign and used the martingale
property that E[M
t
] = 0. Finally letting s = T − t and rearranging yields an
expression for the option value
V (S, t) = E
Q
_
e
−r(T−t)
V (S
T
, T)
¸
−
_
T−t
0
e
−ru
E
Q
_
L
_
V (S
t+u
, t +u)
_¸
du.
Now by deﬁnition V (S
T
, T) = G(S
T
, T) and also we have that L
_
V (S
t+u
, t +u)
_
= 0
in the continuation region and is only nonzero in the stopping region where we have
V (S
t+u
, t +u) = G(S
t+u
, t +u). This modiﬁes the expression to
V (S, t) = E
Q
_
e
−r(T−t)
G(S
T
, T)
¸
−
_
T−t
0
e
−ru
E
Q
_
L
_
G(S
t+u
, t +u)
_
I
_
S
t+u
≤ b(t +u)
_¸
du
= E
Q
_
e
−r(T−t)
G(S
T
, T)
¸
−
_
T−t
0
e
−ru
E
Q
_
H(S
t+u
, t +u)I
_
S
t+u
≤ b(t +u)
_¸
du,
where we have used the deﬁnition of the Hfunction, I() is the indicator function
and b() denotes the location of the free boundary separating the continuation and
stopping regions. We now consider the expectation under the integral sign which can
be expressed as
E
Q
_
H(z, t +u)I
_
z ≤ b(t +u)
_¸
=
_
∞
0
H(z, t +u)I
_
z ≤ b(t +u)
_
f(S, t; z, t +u)dz,
where f(S, t; z, t + u) is the transitional probability density function of the process
started at position S at time t and ﬁnishing at position z at time t +u given by
10
f(S, t; z, t +u) =
1
σz
√
2πu
exp
_
−
_
log
_
z
S
_
−
_
r −D −
1
2
σ
2
_
u
_
2
2σ
2
u
_
.
Hence we have the following representation of the British put option value,
V (S, t) = J(S, t) −
_
T−t
0
e
−ru
L(S, t, b(t +u), t +u)du (9.36)
10
For a derivation see appendix C.
CHAPTER 9. THE BRITISH OPTION 203
where we have deﬁned
J(S, t) = E
Q
S,t
_
e
−r(T−t)
G(S
T
, T)
¸
,
= E
Q
S,t
_
e
−r(T−t)
(K −S)
+
¸
,
= V
P
E
(S, t),
i.e. the corresponding European put option value and
L(S, t, b(t +u), t +u) =
_
∞
0
H(z, t +u)I
_
z ≤ b(t +u)
_
f(S, t; z, t +u)dz,
=
_
b(t+u)
0
H(z, t +u)f(S, t; z, t +u)dz.
with H(S, t) as deﬁned by equation (9.11).
11
Now to determine the free boundary
we can evaluate equation (9.36) at the free boundary S = b(t) where we know that
V (b(t), t) = G(b(t), t), hence
G(b(t), t) = V
P
E
(b(t), t) −
_
T−t
0
e
−ru
L(b(t), t, b(t +u), t +u)du. (9.37)
For the proof of the uniqueness of the above representation see Peskir and Samee
(2008a,b). We can attempt to utilise the nonlinear integral equation (9.37) in order to
determine the large T −t behaviour of the free boundary, much in the same way as we
did for the function S
h
(t). The equations involved are clearly much more complicated
and so it will not be a trivial matter to extract such asymptotic behaviour. As such
this shall be left as the topic of future research.
11
At this stage we can see that using the gain function of the standard American put option, i.e.
G
A
(S, t) = (K − S)
+
will yield H(z, t + u) = −rK and the earlyexercise premium representation
for the American put option, (9.33), immediately follows.
Chapter 10
Conclusions
In this thesis we have investigated a number of models which have been proposed to
incorporate ﬁnite liquidity of the underlying asset into the classical BlackScholes
Merton option pricing framework. Here, the powerful tool of asymptotic analysis
has been used to extract important information about the behaviour of such models
close to expiry. A feature common to a number of these models is that the over
all dispersion term, involving the option gamma, diminishes in magnitude as the
gamma increases in magnitude (as indeed it must as standard payoﬀ conditions are
approached). The upshot of this is that models of this general class cannot exhibit
fully diﬀerentiable solutions at times prior to expiry; instead, we must allow solutions
with discontinuous deltas. This is clearly a somewhat undesirable feature, which is
exacerbated by the possibility of negative option values for puts. Indeed, invariably
these solution features lead to completely spurious solutions if standard numerical
procedures are adopted. However, this analysis also gives guidance on how to tackle
these problems numerically at times away from expiry (the full problem), incorpo
rating the appropriate discontinuities into the numerical scheme. Allied to this, the
vanishing of the denominator in the dispersion term can also be a serious issue. It
is concluded that there is insuﬃcient ﬁnancial modelling to describe the true price
dynamics in such situations.
It is clear that the period close to expiry is the most critical for optionpricing models
204
CHAPTER 10. CONCLUSIONS 205
and any model that successfully treats this regime should also successfully replicate
the option value dynamics for all time. The approach detailed in this thesis should
give guidance for the development of models incorporating ﬁnite liquidity without
the undesirable features observed in a number of the existing models. Several models
in the past have circumvented these diﬃculties close to expiry but generally using
ad hoc, rather than intuitively justiﬁable arguments. The hope is that the analysis
presented in this thesis will help in this respect; in addition, below we discuss brieﬂy
some preliminary ideas about how this may be achieved in future research.
One problem with the modelling framework introduced in chapter 2, from a ﬁnancial
viewpoint, is that the change in price dS becomes unbounded when the ‘forcing’ term
df becomes unbounded, and for the case in which f = ∆ this will happen when d∆
becomes unbounded. Unfortunately, for option contracts with nonsmooth payoﬀ
proﬁles, the unbounded nature of d∆ is unavoidable and so if we are to incorporate
these (common) situations into such modelling frameworks then it is suggested that a
nonlinear response of df to d∆ may well overcome such diﬃculties. We could choose
to incorporate such nonlinearity into our deﬁnition of the forcing term f. i.e. instead
of setting f = ∆ we could set the forcing term to be some function of ∆, i.e.
f = g(∆).
However it is not the function f which ultimately aﬀects the price, but rather its
inﬁnitesimal change df, hence it is the term λdf which we require to remain bounded
(irrespective of the trading strategy ∆). Furthermore, when considering option pric
ing, the only term in df that ﬁlters through into the option price is the
∂f
∂S
dS term,
therefore it is desirable to bound
∂f
∂S
rather than f. This can be done if we model
the derivative of the forcing term (f) with respect to the asset price as a bounded
(nonlinear) function of the derivative of the trading strategy (∆), i.e.
∂f
∂S
= g
_
∂∆
∂S
_
(10.1)
where g() is bounded above. An example of such a function is the ratio of two
CHAPTER 10. CONCLUSIONS 206
n
th
order polynomials such as
g(x) =
α
n
x
n
+α
n−1
x
n−1
+. . . +α
0
β
n
x
n
+β
n−1
x
n−1
+. . . +β
0
,
which in the limit x → ∞ is bounded above by the ratio α
n
/β
n
. If the aim is to
prevent the denominator from vanishing then we require this ratio to be 1/λ or below,
hence the function
g(x) =
x
λx +β
would suﬃce, where β can be chosen to obtain the desired shape of the response curve.
Note also that since the response function g(x) is concave, this model is consistent
with the empirical evidence of (nonlinear) price impact discussed in section 8.2. The
functional form of this dependence, however, has been chosen arbitrarily and so again
this extension to the model can be seen as merely an ‘ad hoc’ ﬁx to the diﬃculties
associated with the vanishing of the denominator. It is possible, however, that the
ideas presented above could be suﬃciently formalised with further research.
Another area of future research could be to exploit the links of the existing models
with the theory of linear and nonlinear elasticity (of solids).
1
Indeed, in some sense,
the models formulated in chapter 2 are analogous to Hooke’s law in linear elasticity,
the more we push the market the more it will move, and in a linear fashion. It may
be that this force/extension relationship can be approximated as linear, however it
is likely that this will only be the case provided the force remains within reasonable
limits. Hence, just as current models are analogous to Hooke’s law for elastic media,
there may also be an analogy to the ‘elastic limit’ of a material, i.e. the elastic limit
of the market, beyond which the market will respond in a nonlinear fashion. This
limit point might correspond to the current market depth, or a point much further
away. However, ultimately the aﬀect on the price must be bounded, since there are
only a ﬁnite number of shares (and hence a ﬁnite force) available. It is hoped that if
the above considerations are incorporated into the current modelling framework then
the problems associated with the vanishing of the denominator may be overcome.
1
This is touched upon in the paper by Sch¨ onbucher and Wilmott (2000), in which it was suggested
that the problems of the vanishing denominator may disappear if some ‘elasticity’ is incorporated
into the response of the market price to large trades.
CHAPTER 10. CONCLUSIONS 207
Finally, it may be possible to preclude the denominator from vanishing if we utilise
the techniques outlined in Soner and Touzi (2007), who extended the ideas origi
nally introduced in Broadie et al. (1998). These works deal with cases in which the
gamma of the replicating portfolio is bounded above (or below) by some trading
constraint. In such situations, not all options can be perfectly replicated due to the
inability of the replicating portfolio to replicate the option value in regions of large
gamma. However, the minimal superreplicating price can be deﬁned as the cheapest
replicating strategy that dominates the option value in the entire domain. For the
classical BlackScholesMerton framework it can be shown that such a minimal super
replicating price corresponds to the perfectreplicating price (i.e. with no constraints)
of the same option, but with a suitably ‘facelifted’ payoﬀ proﬁle. Such a facelifted
payoﬀ proﬁle corresponds ostensibly to a suﬃciently smoothed payoﬀ proﬁle. It is
thought that applying the constraint V
SS
< 1/λ to the nonlinear PDE (4.1) may help
to regularise the solutions. However, it is not obvious that these techniques can be
extended to the illiquid situation, since the results rely on the stochastic representa
tion of the option value (cf. equation (1.4)), a representation that does not exist for
the fully nonlinear equation (4.1). In addition, the superreplicating price is only one
possible paradigm for option pricing in incomplete markets, and so it is not clear cut
that this is the paradigm to use. Furthermore, it is a generally held consensus that
the premium paid to superreplicate (i.e. remain entirely riskfree) is, in practise, too
high.
If it transpires that the problems associated with the vanishing of the denominator
cannot be remedied by the above suggestions, or by some other means, then it is
asserted that, of the three models that were shown to admit wellposed solutions
close to expiry, the Bakstein and Howison (2003) model is one that would be the most
desirable alternative model; the reason for this is twofold. Firstly, whilst remaining
wellposed close to expiry the option price behaviour also remains suﬃciently diﬀerent
from that of the corresponding BlackScholes (liquid) option. Secondly, in the limit
of no price slippage, this model reduces to the model of Cetin et al. (2004) which has
CHAPTER 10. CONCLUSIONS 208
become a popular model for liquidation costs in recent years.
An alternative would be to specify an entirely new framework for incorporating price
impact onto option pricing theory. The aims of such a model would be: (i) to ﬁx
the problem with the denominator vanishing, resulting in wellposed problems for
standard payoﬀ proﬁles, and (if possible) (ii) to incorporate the empirical evidence of
nonlinear price impact. Criteria such as consistency with empirical data, ﬂexibility
in application and also computational aspects (such as the regimes close to expiry
considered in this thesis) would be crucial to the success of such a model.
Also in this thesis, and on a related theme, we have investigated the properties of the
newly introduced British option;
2
a new nonstandard class of early exercise options.
Such options help to mediate the eﬀects of a ﬁnitely liquid market since the contract
does not require the holder to enter the market and hence incur liquidation costs.
Here, once again, the powerful tool of asymptotic analysis, coupled with advanced
numerical methods have been used to shed light on the behaviour of the earlyexercise
boundary for both large and small times from expiry. Furthermore, a pleasingly sim
ple variable transform was found that helped to reduce the associated freeboundary
problem to that of the standard American option (with dividends) in the regime close
to expiry.
Finally, most researchers in quantitative ﬁnance have an opinion on the direction
of future research in the ﬁeld, some more outspoken than others. Wilmott and
Rasmussen (2002) hypothesise that future models will move away from the simplicity
of traditional stochastic models and their assumptions about probabilistic behaviour.
They also suggest that future models will inevitably draw from a wider range of
mathematical tools. Lipton (2001) goes further to suggest that future research needs
to pay much more attention to the issue of determining the spot price and to predict
its short term evolution, in other words, to provide a suﬃciently formal framework in
which to study the market microstructure including supply and demand and liquidity
2
See Peskir and Samee (2008a,b).
CHAPTER 10. CONCLUSIONS 209
eﬀects. In addition, eminent physicist turned ‘quant’ Emanual Derman states in his
blog
3
that hopefully future work will aim to narrow the gap between the invisible
microstructure of markets and the observable macroscopic properties such as market
prices, a gap which at present is particularly large. It is hoped that this thesis is at
least a step in that direction.
3
See http://www.wilmott.com/blogs/eman/.
Bibliography
Abramowitz, M. and Stegun, I. (1968). Handbook of Mathematical Functions. Dover,
New York.
Agliardi, E. and Andergassen, R. (2001). Feedback eﬀects of dynamic hedging strate
gies in the presence of transaction costs. working paper.
Almgren, R. and Chriss, N. (2001). Optimal execution of portfolio transactions.
Journal of Risk, 3:5–40.
Almgren, R., Thum, C., Hauptmann, E., and Li, H. (2005). Equity market impact.
Risk Magazine, July, pages 57–62.
Amihud, Y. and Mendelson, H. (1986). Asset pricing and the bidask spread. Journal
of Financial Economics, 17:223–249.
Avellaneda, M., Kasyan, G., and Lipkin, M. (2007). Powerlaw priceimpact models
and stock pinning near option expiration dates. working paper.
Avellaneda, M., Levy, A., and Paras, A. (1995). Pricing and hedging derivative
securities in markets with uncertain volatilities. Applied Mathematical Finance,
2:73–88.
Avellaneda, M. and Lipkin, M. (2003). A marketinduced mechanism for stock pin
ning. Quantitative Finance, 3:417–425.
Bakstein, D. and Howison, S. (2003). A nonarbitrage liquidity model with observable
parameters. working paper.
210
BIBLIOGRAPHY 211
Bank, P. (2006). Pricing and hedging in illiquid ﬁnancial markets. Presentation at
the 4th world congress of the Bachelier ﬁnance society.
Bank, P. and Baum, D. (2004). Hedging and portfolio optimization in ﬁnancial
markets with a large trader. Mathematical Finance, 14:1–18.
Bank for International Settlements (1999). Market liquidity: Research ﬁndings and
selected policy implications. Committee on the Global Financial System, Basel.
Bank for International Settlements (2001). Structural aspects of market liquidity
from a ﬁnancial stability perspective. A discussion note by the Committee on the
Global Financial System, Basel (March).
Barenblatt, G. I. (1996). Scaling, Selfsimilarity, and Intermediate Asymptotics.
Cambridge University Press, Cambridge. Cambridge Texts in Applied Mathemat
ics, 14.
Barles, G. and Soner, H. M. (1998). Option pricing with transaction costs and a
nonlinear BlackScholes equation. Finance and Stochastics, 2:369–397.
Bergman, Y. Z., Grundy, B. D., and Wiener, Z. (1996). General properties of option
prices. The Journal of Finance, 51:1573–1610.
Bick, A. (1987). On the consistency of the BlackScholes model with a general equi
librium framework. Journal of Financial Quantitative Analysis, 22:259–275.
Bick, A. (1990). On viable diﬀusion price processes of the market portfolio. Journal
of Finance, 45:673–689.
Bj¨ ork, T. (2004). Arbitrage Theory in Continuous Time. Oxford University Press.
Second Edition.
Black, F. and Scholes, M. (1973). The pricing of options and corporate liabilities.
Journal of Political Economy, 81:637–659.
BIBLIOGRAPHY 212
Bordag, L. A. (2007). On the valuation of a fair price in case of an non perfectly
liquid market. Proceedings of the Workshop on Mathematical Control Theory and
Finance, Lisbon, 1014 April, 2007.
Bordag, L. A. and Chmakova, A. Y. (2007). Explicit solutions for a nonlinear model
of ﬁnancial derivatives. International Journal of Theoretical and Applied Finance,
10:1–21.
Bordag, L. A. and Frey, R. (2007). Nonlinear option pricing models for illiquid mar
kets: Scaling properties and explicit solutions. working paper, arXiv:0708.1568v1
[math.AP].
Bouchaud, J. P., M´ezard, M., and Potters, M. (2002). Statistical properties of stock
order books: Empirical results and models. Quantitative Finance, 2:251–256.
Brennan, M. and Schwartz, E. (1989). Portfolio insurance and ﬁnancial market equi
librium. Journal of Business, 62:455–476.
Broadie, M., Cvitani´c, J., and Soner, H. M. (1998). Optimal replication of contingent
claims under portfolio constraints. The Review of Financial Studies, 11:59–79.
Carr, P., Jarrow, R., and Myneni, R. (1992). Alternative characterizations of Amer
ican put options. Mathematical Finance, 2:78–106.
Cetin, U., Jarrow, R. A., and Protter, P. (2004). Liquidity risk and arbitrage pricing
theory. Finance and Stochastics, 8:311–341.
Cetin, U. and Rogers, L. C. G. (2007). Modeling liquidity eﬀects in discrete time.
Mathematical Finance, 17:15–29.
Chalmers, J. M. and Kadlec, G. B. (1998). An empirical examination of the amortized
spread. Journal of Financial Economics, 48:159–188.
Chen, N. and Kan, R. (1995). Expected return and the bidask spread. In Saitou, S.,
Sawaki, K., and Kubota, K., editors, Modern Portfolio Theory and Applications.
Gakujutsu Shuppan Center, Osaka.
BIBLIOGRAPHY 213
Chen, X., Chadam, J., Jiang, L., and Zheng, W. (2008). Convexity of the exercise
boundary of the american put option on a zero dividend asset. Mathematical
Finance, 18:185–197.
Cole, J. D. (1968). Perturbation Methods in Applied Mathematics. Blaisdell Publ.
Co., Walttham, Mass.
Crank, J. (1957). Two methods for the numerical solution of movingboundary prob
lems in diﬀusion and heat ﬂow. Q. J. Mechanics Appl. Math., 10:220–231.
Cuoco, D. and Cvitanic, J. (1998). Optimal consumption choices for a ‘large’ investor.
Journal of Economic Dynamics & Control, 22:401–436.
Cvitanic, J. and Ma, J. (1996). Hedging options for a large investor and forward
backwards SDEs. The Annals of Applied Probability, 6:370–398.
Datar, V., Naik, N., and Radcliﬀe, R. (1998). Liquidity and stock returns: An
alternative test. Journal of Financial Maarkets, 1:203–220.
Davis, M. and Etheridge, A. (2006). Louis Bachelier’s Theory of Speculation: The
Origins of Modern Finance. Princeton University Press.
Diz, F. and Finucane, T. J. (1993). The rationality of early exercise decisions: Ev
idence from the S&P100 index options market. The Review of Financial Studies,
6:765–797.
Dormand, J. R. and Prince, P. J. (1980). A family of embedded RungeKutta formu
lae. Journal of Computational and Applied Mathematics, 6:19–26.
Duck, P. W., Newton, D. P., Widdicks, M., and Yang, C. (2008). Singular pertur
bation techniques applied to multiasset option pricing. Mathematical Finance, to
appear.
Duﬃe, D. (1996). Dynamic Asset Pricing Theory, 2nd Edition. Princeton University
Press.
BIBLIOGRAPHY 214
Duﬀy, D. J. (2004). A critique of the Crank Nicolson scheme strengths and weaknesses
for ﬁnancial instrument pricing. Wilmott Magazine, 4:68–76.
Ekstr¨ om, E., Janson, S., and Tysk, J. (2005). Superreplication of options on several
underlying assets. Journal of Applied Probability, 42:27–38.
El Karoui, N., JeanblancPicqu´e, M., and Shreve, S. (1998). Robustness of the Black
and Scholes formula. Mathematical Finance, 8:93–126.
Eleswarapu, V. and Reinganum, M. (1993). The seasonal behavior of liquidity pre
mium in asset pricing. Journal of Financial Economics, 34:373–386.
Esser, A. and Moench, B. (2003). Modeling feedback eﬀects with stochastic liquidity.
working paper.
Evans, J. D., Kuske, R., and Keller, J. B. (2002). American options on assets with
dividends near expiry. Mathematical Finance, 12:219–237.
Evans, L. C. (1998). Partial Diﬀerential Equations. American Mathematical Society,
Providence, RI.
Fama, E. and French, K. (1996). Multifactor explanations of asset pricing anomalies.
Journal of Finance, 51:55–84.
F¨ ollmer, H. and Schweizer, M. (1993). A microeconomic approach to diﬀusion models
for stock prices. Mathematical Finance, 3:1–23.
Frey, R. (1998). Perfect option replication for a large trader. Finance and Stochastics,
2:115–142.
Frey, R. (2000). Market illiquidity as a source of model risk in dynamic hedging.
Model Risk, pages 125–136.
Frey, R. and Patie, P. (2002). Risk management for derivatives in illiquid markets: A
simulation study. In: K Sandmann and P Sch¨ onbucher (Eds.), Advances in ﬁnance
and stochastics, Springer, pages 137–159.
BIBLIOGRAPHY 215
Frey, R. and Stremme, A. (1997). Market volatility and feedback eﬀects from dynamic
hedging. Mathematical Finance, 7:351–374.
Friedrichs, K. O. (1954). Special Topics in Analysis. New York University.
Gabaix, X., Gopikrishnan, P., Plerou, V., and Stanley, H. E. (2003). A theory of
power law distributions in ﬁnancial market ﬂuctuations. Nature, 423:267–270.
Geske, R. (1977). The valuation of corporate liabilities as compound options. Journal
of Financial and Quantitative Analysis, 12:541–552.
Geske, R. (1979). The valuation of compound options. Journal of Financial Eco
nomics, 7:63–81.
Harrison, J. M. and Kreps, D. M. (1979). Martingales and arbitrage in multiperiod
securities markets. Journal of Economic Theory, 20:381–408.
Harrison, J. M. and Pliska, S. R. (1981). Martingales and stochastic integrals in
the theory of continuous trading. Stochastic Processes and Their Applications,
11:215–260.
Harrison, J. M. and Pliska, S. R. (1983). A stochastic calculus model of continuous
trading: Complete markets. Stochastic Processes and Their Applications, 15:313–
316.
HenryLabord´ere, P. (2004). The feedback eﬀect of hedging in portfolio optimization.
working paper, arXiv:condmat/0404520v1 [condmat.other].
Hirsch, M. W. and Smale, S. (1974). Diﬀerential Equations, Dynamical Systems, and
Linear Algebra. San Diego, CA: Academic.
Hodges, S. D. and Selby, M. J. P. (1987). On the evaluation of compound options.
Management Science, 33:347–355.
Howison, S. (2005). Matched asymptotic expansions in ﬁnancial engineering. Journal
of Engineering Mathematics, 53:385–406.
BIBLIOGRAPHY 216
Jacka, S. D. (1991). Optimal stopping and the American put. Mathematical Finance,
1:1–14.
Jacklin, C., Kleidon, A., and Pﬂeiderer, P. (1992). Underestimation of portfolio
insurance and the crash of october 1987. Review of Financial Studies, 5:35–63.
Jacod, J. and Protter, P. (2003). Probability Essentials. SpringerVerlag Berlin Hei
delberg New York. Second Edition.
Jandaˇcka, M. and
ˇ
Sevˇecoviˇc, D. (2005). On the riskadjusted pricingmethodology
based valuation of vanilla options and explanation of volatility smile. Journal of
Applied Mathematics, 3:252–258.
Janson, S. and Tysk, J. (2003). Volatility time and properties of option. Annals of
Applied Probability, 13:890–913.
Jarrow, R. A. (1992). Market manipulation, bubbles, corners and short squeezes.
Journal of Financial Quantitative Analysis, 27:311–336.
Jarrow, R. A. (1994). Derivative securities markets, market manipulation and option
pricing theory. Journal of Financial Quantitative Analysis, 29:241–261.
Jarrow, R. A. and Protter, P. (2004). A short history of stochastic integration
and mathematical ﬁnance: The early years, 18801970. In The Herman Rubin
Festschrift, IMS Lecture Notes, 45:75–91.
Jeannin, M., Iori, G., and Samuel, D. (2006). Modeling stock pinning. Discussion
paper, City University, London.
Johnson, P. V. (2007). Improved Numerical Techniques for Occupationtime Deriva
tives and Other Complex Financial Instruments. Ph.D. thesis, University of Manch
ester, UK.
Johnson, R. S. (2004). Singular Perturbation Theory: Mathematical and Analytical
Techniques with Applications to Engineering. Springer, New York.
BIBLIOGRAPHY 217
Jonsson, M. and Keppo, J. (2002). Option pricing for large agents. Applied Mathe
matical Finance, 9:262–272.
Jordan, D. W. and Smith, P. (1999). Nonlinear Ordinary Diﬀerential Equations, 3rd
Edition. Oxford University Press.
Kaplun, S. (1967). Fluid Mechanics and Singular Perturbations. Academic Press,
New York.
Kaplun, S. and Lagersrom, P. A. (1957). Asymptotic expansions of the NavierStokes
solutions for small reynolds numbers. J. Math. Mech., pages 585–593.
Karatzas, I. and Shreve, S. (1998). Methods of Mathematical Finance. Springer
Verlag, New York.
Keller, H. (1978). Numerical methods in boundarylayer theory. Ann. Rev. Fluid
Mech., 10:417–433.
Kempf, A. and Korn, O. (1999). Market depth and order size. Journal of Financial
Markets, 2:29–48.
Kim, I. J. (1990). The analytic valuation of American options. The Review of Fi
nancial Studies, 3:547–572.
Kreps, D. (1979). Multiperiod securities and the eﬃecient allocation of risk: A com
ment on the BlackScholes option pricing model. Technical Report 496, Graduate
School of Business, Stanford University.
Krishnan, H. and Nelken, I. (2001). The eﬀect of stock pinning upon option prices.
Risk Magazine, December, pages 17–20.
Kruskal, M. D., Joshi, N., and Halburd, R. (1997). Analytic and asymptotic methods
for nonlinear singularity analysis: A review and extensions of tests for the Painleve
property. Lecture Notes in Physics, 495:171 – 205.
Kuske, R. A. and Keller, J. B. (1998). Optimal exercise boundary for an American
put option. Applied Mathematical Finance, 5:107–116.
BIBLIOGRAPHY 218
Kyle, A. S. (1985). Continuous auctions and insider trading. Econometrica, 53:1315–
1335.
Ladyzenskaja, O. A., Solonnikov, V. A., and Ural’ceva, N. N. (1968). Linear and
Quasilinear Equations of Parabolic Type. American Mathematical Society: Provi
dence, RI.
Landau, H. G. (1950). Heat conduction in a melting solid. Quarterly of Applied
Mathematics, 8:81–94.
LeBaron, B. (2000). Agentbased computational ﬁnance: Suggested reading and early
research. Journal of Economic Dynamics and Control, 24:679–702.
Lillo, F., Farmer, L. D., and Mantegna, R. N. (2003). Single curve collapse of the
price impact function for the New York Stock Exchange. Nature, 421:129–130.
Lipton, A. (2001). Mathematical Methods for Foreign Exchange. World Scientiﬁc.
Liptser, R. S. and Shiryaev, A. N. (2001). Statistics of Random Processes II: Appli
cations, 2nd Edition. Springer Verlag, New York.
Liu, H. and Yong, J. (2005). Option pricing with an illiquid underlying asset market.
Journal of Economic Dynamics & Control, 29:2125–2156.
Liu, W. (2006). A liquidityaugmented capital asset pricing model. Journal of Fi
nancial Economics, 82:631–671.
Lyukov, A. (2004). Option pricing with feedback eﬀects. International Journal of
Theoretical and Applied Finance., 7:757–768.
Mancino, M. E. and Ogawa, S. (2003). Nonlinear feedback eﬀects by hedging strate
gies. In Stochastic Processes and Applications to Mathematical Finance, Proceed
ings of the 2003 Symposium. Ritsumeikan University, World Scientiﬁc: Singapore.
Markowitz, H. (1959). Portfolio Selection: Eﬃcient Diversiﬁcation of Investments.
J Wiley and Sons, New York.
BIBLIOGRAPHY 219
Merton, R. C. (1973). Theory of rational option pricing. Bell J. Econ. Manag. Sci.,
4:141–183.
Ni, S. X., Pearson, N. D., and Poteshman, A. M. (2005). Stock price clustering on
option expiration dates. Journal of Financial Economics, 78:49–87.
Nirenberg, L. (1953). A strong maximum principle for parabolic equations. Commu
nications on Pure and Applied Mathematics, 6:167–177.
O’Hara, M. (1995). Market Microstructure Theory. Blackwell.
Peixoto, M. M. (1997). Generic properties of ordinary diﬀerential equations. MAA
Studies of Mathematics, 14:52–92.
Peskir, G. (2003). Local timespace calculus and extensions of Itˆ o’s formula. Progress
in Probability, Birkh¨ auser, Basel, 55:177–192.
Peskir, G. (2005a). A changeofvariable formula with local time on curves. Journal
of Theoretical Probability, 18:499–535.
Peskir, G. (2005b). On the American option problem. Mathematical Finance, 15:169–
181.
Peskir, G. and Samee, F. (2008a). The British call option. Research Report No. 2,
Probab. Statist. Group Manchester.
Peskir, G. and Samee, F. (2008b). The British put option. Research Report No. 1,
Probab. Statist. Group Manchester.
Peskir, G. and Shiryaev, A. (2006). Optimal Stopping and Freeboundary Problems.
Birkhauser.
Pirrong, C. (2004). Detecting manipulation in futures markets: The Ferruzzi soybean
episode. American Law & Economics Review, 6:28–71.
Platen, E. and Schweizer, M. (1998). On feedback eﬀects from hedging derivatives.
Mathematical Finance, 8:67–84.
BIBLIOGRAPHY 220
Plerou, V., Gopikrishnan, P., Gabaix, X., and Stanley, H. E. (2002). Quantifying
stock price response to demand ﬂuctuations. Physical Review E 027104, 66.
Poincar´e, H. (1886). Sur le int´egrales irr´eguli`eres des ´equations lin´eaires. Acta Math.,
8:259–344.
Potters, M. and Bouchaud, J.P. (2003). More statistical properties of order books
and price impact. Physica A, 324:133–140.
Prandtl, L. (1904).
¨
Uber ﬂ¨ ussigkeitsbewegung bei sehr kleiner reibung. Verh. III.
Intern. Math. Kongr., Heidelberg, pages 484–491. Teubner, Lwipzig, 1905.
Protter, M. and Weinberger, H. (1984). Maximum Principles in Diﬀerential Equa
tions. SpringerVerlag, New York. Corrected reprint of the 1967 original.
Protter, P. (1990). Stochastic integration and diﬀerential equations: A new approach.
Springer, Berlin. Application of Mathematics series.
Protter, P. (2006). Stochastics on the street: Models of liquidity risk. Fi
nancial Engineering News, 49. Available at http://www.fenews.com/fen49/
stochasticsstreet/stochasticsstreet.html.
Rogers, L. C. G. and Singh, S. (2006). Modelling liquidity and its eﬀects on price.
working paper.
Rosenerans, S. I. (1972). Derivation of the HopfCole solution to Burgers’ equation
by stochastic integrals. Proceedings of the American Mathematical Society, 32:147–
149.
Samuelson, P. A. (1965). Rational theory of warrant pricing. Industrial Management
Review, 6:13–32.
Sanfelici, S. (2007). Calibration of a nonlinear feedback option pricing model. Quan
titative Finance, 7:95–110.
Sarr, A. and Lybek, T. (2002). Measuring liquidity in ﬁnancial markets. International
Monetary Fund working paper WP/02/232.
BIBLIOGRAPHY 221
Sch¨ onbucher, P. and Wilmott, P. (2000). The feedback eﬀect of hedging in illiquid
markets. SIAM Journal of Applied Mathematics, 61:232–272.
Sircar, K. and Papanicolaou, G. (1998). General BlackScholes models accounting for
increased market volatility from hedging strategies. Applied Mathematical Finance,
5:45–82.
Smith, G. D. (1978). Numerical Solutions of Partial Diﬀerential Equations: Finite
Diﬀerence Methods. Oxford University Press, Oxford. Second Edition.
Soner, H. M. and Touzi, N. (2007). Hedging under gamma constraints by optimal
stopping and facelifting. Mathematical Finance, 17:59–79.
van Dyke, M. D. (1964). Perturbation Methods in Fluid Dynamics. Academic Press,
New York.
Whaley, R. E. (1982). Valuation of American call options on dividendpaying stocks.
Journal of Financial Economics, 10:29–58.
Whalley, A. E. and Wilmott, P. (1993). Counting the costs. Risk Magazine, 6:59–66.
Widder, D. V. (1975). The Heat Equation. Academic Press, New York.
Widdicks, M. (2002). Examination, Extension and Creation of Methods for Pricing
Options with Early Exercise Features. Ph.D. thesis, University of Manchester, UK.
Widdicks, M., Duck, P. W., Andricopoulos, A. D., and Newton, D. P. (2005). The
BlackScholes equation revisited: Asymptotic expansions and singular perturba
tion. Mathematical Finance, 15:373–391.
Williams, J. C. (1995). Manipulation on Trail: Economic Analysis and the Hunt
Silver Case. Cambridge University Press.
Wilmott, P., Howison, S., and Dewynne, J. (1995). The Mathematics of Financial
Derivatives. Cambridge University Press.
BIBLIOGRAPHY 222
Wilmott, P. and Rasmussen, H. O. (2002). New Directions in Mathematical Finance.
Wiley.
Appendix A
Maximum Principles
Maximum principles are extremely useful tools to investigate the properties of the
solutions to partial diﬀerential equations, such as monotonicity in parameters, unique
ness and convexity. These principles date back to as early as 1839 and for a readable
overview of their history and a more indepth exposition see Protter and Weinberger
(1984).
Nirenberg (1953) states (and proves) the maximum principles applied to equations
that can be written in the form
1
L(V ) +cV = D
where
L(V ) =
n
i,j=1
a
ij
V
S
i
S
j
+
m
α,β=1
b
αβ
V
τατ
β
+
n
i=1
a
i
V
S
i
+
m
α=1
b
α
V
τα
.
If we restrict ourselves to the one dimensional case, i.e. when m = n = 1, this reduces
to
L(V ) = a
11
V
SS
+b
11
V
ττ
+a
1
V
S
+b
1
V
τ
,
and hence we can apply the maximum principles to equations of the form
AV
SS
+BV
ττ
+aV
S
+bV
τ
+cV = D(S, τ). (A.1)
1
Also see, for example, Evans (1998) and Protter and Weinberger (1984).
223
APPENDIX A. MAXIMUM PRINCIPLES 224
In order to do this we require the diﬀerential operator L to be elliptic in the S
variables and parabolic in the τvariables. Therefore we require
n
i,j=1
a
ij
η
i
η
j
> 0,
hence positive deﬁnite and
m
α,β=1
b
αβ
ξ
α
ξ
β
≥ 0
hence positive semideﬁnite for any real vectors η, ξ ,= 0. In the case m = n = 1 this
reduces to
a
11
η
2
1
> 0 ⇒a
11
> 0 ⇒A > 0
and also
b
11
ξ
2
1
≥ 0 ⇒b
11
≥ 0 ⇒B ≥ 0.
Hence we can apply the maximum principle to equations of the form (A.1) provided
A > 0 and B ≥ 0 with no restriction on the sign of a and b. In what follows (for
simplicity) we shall assume that B = 0 and b = −1 to obtain a forward parabolic
(diﬀusion) equation of the form
AV
SS
+aV
S
−V
τ
+cV = D(S, τ), A > 0. (A.2)
Furthermore we make the assumption that c = 0 in the above. This last simpliﬁcation
may seem a little restrictive but it should be noted that it is possible to reduce
equation (A.2) (for any c) to one in which c = 0 by making the transformation
V = e
cτ
u to arrive at an equation of the form
L(u) = Au
SS
+au
S
−u
τ
= e
−cτ
D(S, τ) =
ˆ
D(S, τ), (A.3)
which we now wish to apply the maximum principle to. Before we state the maximum
principle we can formally deﬁne the solution domain as Ω ∈ R
+
and assume that it
is open, connected and bounded. Let
Ω
T
= Ω (0, T],
where T > 0 and also deﬁne
∂
∗
Ω
T
= ∂Ω¸Ω ¦T¦,
APPENDIX A. MAXIMUM PRINCIPLES 225
hence for a onedimensional rectangular domain, ∂
∗
Ω
T
corresponds to the boundaries
τ = 0, S = 0 and S = S
max
. We are now ready to formally state the maximum
principle applied to equations of the form (A.3).
The maximum principle states that if
ˆ
D(S, τ) ≥ 0 and A > 0 everywhere in the
closure of the domain, Ω
T
, and furthermore that the coeﬃcients of equation (A.3)
are bounded in Ω
T
, then the maximum of the solution (which is assumed to be C
2,1
in the interior of Ω
T
) must occur on the boundary ∂
∗
Ω
T
, i.e.
sup
Ω
T
u = max
∂Ω
T
u = max
∂
∗
Ω
T
u.
By analogy we can ﬁnd the minimum principle by making the substitution ˆ u = −u
to obtain
L(−ˆ u) = −Aˆ u
SS
−aˆ u
S
+ ˆ u
τ
=
ˆ
D(S, τ),
⇒Aˆ u
SS
+aˆ u
S
− ˆ u
τ
= L(ˆ u) = −
ˆ
D(S, τ).
Hence the maximum principle states that if −
ˆ
D ≥ 0 or
ˆ
D ≤ 0 then ˆ u has its maximum
on the boundary and hence u = −ˆ u has its minimum on the boundary. To summarise,
applying the maximum (minimum) principle to the equation
L(u) =
ˆ
D(S, τ),
we have that, provided A > 0, then if
ˆ
D ≥ 0 then the maximum occurs on the
boundary ∂
∗
Ω
T
, and alternatively if
ˆ
D ≤ 0 then the minimum must occur on the
boundary. Furthermore if
ˆ
D = 0 then both the maximum and minimum must occur
on the boundary.
An outline of the proof of such maximum principles is easily seen by considering the
heat equation operator
L(u) = u
SS
−u
τ
.
Suppose that u(S, τ) satisﬁes the inequality L(u) > 0 in the domain Ω
T
then u
cannot have a (local) maximum at any interior point, since at such a point u
SS
≤ 0
APPENDIX A. MAXIMUM PRINCIPLES 226
and u
τ
= 0, thereby violating L(u) > 0; for a more formal proof, see Protter and
Weinberger (1984).
Aside
As an interesting aside, it is possible to interpret the maximum principle from a
probabilistic point of view. Let us consider a process started at S and time t and
furthermore let M denote the maximum value of the function on the boundary and
m the minimum value. It follows immediately that
m ≤ G(S
γ
Ω
, γ
Ω
) ≤ M,
where γ
Ω
denotes the ﬁrst exit time of the process from the domain Ω, and G(S, t) is
the value of the function on the boundary of the domain, i.e. ∂Ω¸Ω ¦0¦. Taking
expectations given that the process starts at S, i.e. S
t
= S we have that
m ≤ E[G(S
γ
Ω
, γ
Ω
) [S
t
= S] ≤ M,
⇒m ≤ u(S, t) ≤ M,
where u(S, t) will also satisfy a backwards parabolic PDE via the FeynmanKac rep
resentation theorem outlined in subsection 1.3.3. Hence u(S, τ) will satisfy a forwards
parabolic PDE and the solution to this diﬀerential equation must have it’s maximum
on the boundary ∂Ω¸Ω ¦T¦ = ∂
∗
Ω.
A.1 Nonlinear equations
The maximum principles outlined above also hold for any nonlinear equation that
can be expressed in the form
L(u) =
ˆ
D(S, τ),
where the coeﬃcients of the derivatives in the operator can now be functions of the
solution and its derivatives, in other words
L(u) = A(S, τ, u, u
S
, u
SS
)u
SS
+a(S, τ, u, u
S
)u
S
−u
τ
.
APPENDIX A. MAXIMUM PRINCIPLES 227
As an example, consider a general nonlinear PDE of the form
u
τ
= F(S, τ, u, u
S
, u
SS
).
We can apply the maximum principles provided that the function F is elliptic in all
values of its arguments in Ω
T
. In other words in general form
n
l,q=1
∂F(S
i
, τ, p, p
i
, p
ij
)
∂p
lq
ξ
l
ξ
q
> 0,
i.e. positive deﬁnite. For the one dimensional case (n = m = 1) we have
∂F(S, τ, p, p
1
, p
11
)
∂p
11
ξ
2
1
> 0 ⇒
∂F(S, τ, u, u
S
, u
SS
)
∂u
SS
> 0.
A.2 Uniqueness of PDEs
We can use the maximum principle to prove the uniqueness of the solution to a PDE
of the general form
V
τ
= F(S, τ, V, V
S
, V
SS
),
where we assume that ellipticity has been shown. Let V
1
and V
2
denote two solutions
of the above PDE, hence we have
V
1
τ
−V
2
τ
= F(S, τ, V
1
, V
1
S
, V
1
SS
) −F(S, τ, V
2
, V
2
S
, V
2
SS
). (A.4)
Using the the mean value theorem we can rewrite the righthandside of equation
(A.4) as a linear combination of V
1
− V
2
and its ﬁrst and second derivatives with
respect to S. Doing so we obtain
(V
1
−V
2
)
τ
=
∂F
∂V
¸
¸
¸
¸
V
(V
1
−V
2
) +
∂F
∂V
S
¸
¸
¸
¸
V
S
(V
1
−V
2
)
S
+
∂F
∂V
SS
¸
¸
¸
¸
V
SS
(V
1
−V
2
)
SS
,
where V
2
< V < V
1
, V
2
S
< V
S
< V
1
S
and V
2
SS
< V
SS
< V
1
SS
. Hence the equation
takes the form
L(V
1
−V
2
) +c(V
1
−V
2
) = D(S, τ), (A.5)
APPENDIX A. MAXIMUM PRINCIPLES 228
where
A =
∂F
∂V
SS
¸
¸
¸
¸
V
SS
,
a =
∂F
∂V
S
¸
¸
¸
¸
V
S
,
c =
∂F
∂V
¸
¸
¸
¸
V
.
Finally, recall that this can be reduced to an equation of the form
L(u) = 0
by making the transform u = e
−cτ
(V
1
− V
2
) and so we can now apply the relevent
maximum principle.
We know that on the boundary V
1
−V
2
must equal zero, since both solutions must
satisfy the same boundary conditions, hence we must have that u = 0 on the bound
ary. Therefore provided A > 0 and that the coeﬃcients remain bounded, then the
maximum principle will ensure that
u ≡ 0 ⇒V
1
−V
2
≡ 0 ∈ Ω
T
,
hence V
1
≡ V
2
proving uniqueness.
A.2.1 The linear BlackScholes equation
If we wish to apply the maximum principle to the BlackScholes equation we have the
problem that the coeﬃcient of the diﬀusion term will become degenerate at S = 0.
Fortunately we can make the change of variable x = log S which reduces the equation
to
V
τ
=
1
2
σ
2
V
xx
+rV
x
−rV = F(x, V, V
x
, V
xx
),
hence a constant coeﬃcient linear PDE, which exhibits no such degeneracy. Evalu
ating the derivatives of F, equation (A.5) becomes
1
2
σ
2
(V
1
−V
2
)
xx
+r(V
1
−V
2
)
x
−(V
1
−V
2
)
τ
−r(V
1
−V
2
) = 0.
APPENDIX A. MAXIMUM PRINCIPLES 229
and letting u = e
−rτ
(V
1
−V
2
) leads to
1
2
σ
2
u
xx
+ru
x
−u
τ
= 0.
Hence we can identify A =
1
2
σ
2
> 0 and
ˆ
D(S, τ) = 0 and so the application of the
maximum principle ensures that the maximum of the solution must occur on the
boundary and so we must have that u ≡ 0, hence V
1
≡ V
2
giving uniqueness.
A.2.2 The nonlinear (illiquid) BlackScholes equation
The problem of the degeneracy of the diﬀusion coeﬃcient at S = 0 is also present
in the fully nonlinear equation (4.1) but again making the same log transform as for
the BlackScholes case will remove such a degeneracy. For simplicity we shall make
a further transform to the forward prices for the stock and the option, i.e. we make
the transform S = e
x−rτ
and V = ue
−rτ
, which reduces equation (4.1) to
u
τ
=
σ
2
(u
xx
−u
x
)
2
_
1 −λe
−2x+rτ
(u
xx
−u
x
)
_
2
= F (x, τ, u
x
, u
xx
) , (A.6)
and we are now in a situation where the equation is no longer degenerate. The
next step is to check that the function F in (A.6) is elliptic in all values of its
argument, hence to see under which situations we can apply the maximum principles.
Diﬀerentiating F with respect to the second derivative gives
∂F
∂u
xx
=
σ
2
_
1 +λe
−2x+rτ
(u
xx
−u
x
)
_
2
_
1 −λe
−2x+rτ
(u
xx
−u
x
)
_
3
(A.7)
which can be seen to be strictly positive (and hence elliptic) if and only if we have
[u
xx
−u
x
[ <
e
2x−rτ
λ
∈ Ω
T
.
Transforming the above back to the original variables it is clear that this corresponds
to [V
SS
[ < 1/λ for all (S, τ), in other words the restriction that the denominator in
equation (4.1) cannot vanish (cf. Frey, 1998). The consequence of this is that we are
not able to use maximum principles in the regime where the denominator is allowed
to vanish. This is natural since, if the denominator is allowed to vanish then we
APPENDIX A. MAXIMUM PRINCIPLES 230
have unbounded coeﬃcients of the equation, which also contradicts the requirement
for the applicability of such maximum principles. In addition, smoothness of the
solution can also not be determined a priori if the denominator is allowed to vanish.
However, we shall proceed to prove uniqueness of the solution, and investigate other
properties, of (4.1) in the regime where [V
SS
[ < 1/λ everywhere within the domain.
Next, we wish to provide an alternative uniqueness proof for equation (4.1) to that
proposed by Frey (1998). This can be done simply by using (A.7) in which case
equation (A.5) becomes
σ
2
(1 +λe
−2x+rτ
(u
xx
−u
x
))
2
_
1 −λe
−2x+rτ
(u
xx
−u
x
)
_
3
(u
1
−u
2
)
xx
−(u
1
−u
2
)
τ
= 0,
from which it can be seen that (provided the denominator does not vanish) then we
will have that A > 0. Applying the maximum principle to the above equation yields
the required uniqueness result, V
1
≡ V
2
in Ω
T
, since uniqueness is preserved under
the inverse transforms required to convert equation (A.6) back into equation (4.1).
Note that this result still stands for any functional form of the liquidity parameter
λ(S, τ), provide we do not allow the denominator to vanish.
A.3 Monotonicity in λ
Having proved uniqueness we now wish to determine the dependence of the solution to
equation (4.1) on the liquidity parameter λ. We can do so by diﬀerentiating (directly)
the transformed equation (A.6) with respect to λ. Doing so and setting w =
∂u
∂λ
leads
to the following second order (linear) PDE for w
σ
2
(1 +λe
−2x+rτ
(u
xx
−u
x
))
2
_
1 −λe
−2x+rτ
(u
xx
−u
x
)
_
3
(w
xx
−w
x
) −w
τ
=
σ
2
e
−2x+rτ
(u
xx
−u
x
)
2
_
1 −λe
−2x+rτ
(u
xx
−u
x
)
_
3
. (A.8)
It is advantageous at this point to rewrite the above using the following inverse
transform
u
xx
−u
x
= S
2
u
SS
= e
2x−2rτ
u
SS
= e
2x−rτ
V
SS
= e
2x−rτ
Γ,
APPENDIX A. MAXIMUM PRINCIPLES 231
where we have deﬁned Γ := V
SS
, and thus equation (A.8) becomes
σ
2
(1 +λΓ)
2 (1 −λΓ)
3
w
xx
−
σ
2
(1 +λΓ)
2 (1 −λΓ)
3
w
x
−w
τ
= −
σ
2
e
2x−rτ
Γ
2
(1 −λΓ)
3
. (A.9)
Since the initial condition of any option contract will be independent of the liquidity
parameter λ, we must have
w(x, 0) = 0
and more generally w = 0 on the boundary ∂
∗
Ω
T
. Now applying the maximum
principle, or more speciﬁcally the minimum principle since the righthandside of
(A.9) is negative, gives that the solution w must have its minimum on the boundary,
i.e. w ≥ 0 in the interior of the domain Ω
T
, in other words
∂u
∂λ
≥ 0 ⇒
∂V
∂λ
≥ 0 ∈ Ω
T
,
hence the solution is an increasing function of λ. It should be emphasised that this
result only holds in the regime [V
SS
[ < 1/λ, i.e. when the denominator is not allowed
to vanish.
Appendix B
Nondimensionalisation of the
British Put
The freeboundary formulation of the British put option value (9.3) can be non
dimensionalised by making the following substitution
1
S = Ke
x−(µc−D)(T−t)
,
t = T −
2τ
σ
2
,
V (S, t) = K
_
e
−
2r
σ
2
τ
v(x, τ) + 1 −e
x
_
.
The resulting nondimensional system becomes
v
τ
−v
xx
+ (1 −ρ
1
+ρ
2
)v
x
= e
ρ
1
τ
(ρ
2
e
x
−ρ
1
) ,
v(x, τ) = e
x+ρ
1
τ
Φ
_
x +τ
√
2τ
_
−e
ρ
1
τ
Φ
_
x −τ
√
2τ
_
on x = x
f
,
∂v
∂x
(x, τ) = e
x+ρ
1
τ
Φ
_
x +τ
√
2τ
_
on x = x
f
,
v(x, τ) = (e
x
−1) e
ρ
1
τ
as x →∞,
v(x, 0) = (e
x
−1)
+
,
where ρ
1
=
2r
σ
2
and ρ
2
=
2(µc+D)
σ
2
. In addition we have the condition that
x
f
(0) = log
_
r
µ
c
_
if µ
c
≥ r, which it trivially must satisfy due to the speciﬁes of the option contract.
1
Note that the strike price K is scaled out of the problem (completely) by a simple linear scaling.
232
Appendix C
The Probability Density Function
In order to determine the probability density function under the risk neutral measure
Q we adopt the following procedure. We are given the stochastic process
dS
t
= (r −D)S
t
dt +σS
t
dW
Q
t
which is the process under the measure Q. This process has the closed form solution
S
t+u
= S
t
exp
_
σ(W
t+u
−W
t
) +
_
r −D −
1
2
σ
2
_
u
_
(C.1)
where we have assumed the process was started at position S
t
. Hence we have that
the probability that S
t+u
, i.e. the value of the stock price at time t +u given that it
started at S
t
is under some value z is given by
P
Q
[S
t+u
≤ z] = P
Q
_
S
t
exp
_
σ(W
t+u
−W
t
) +
_
r −D −
1
2
σ
2
_
u
_
≤ z
_
= P
Q
_
exp
_
σ(W
t+u
−W
t
) +
_
r −D −
1
2
σ
2
_
u
_
≤
z
S
t
_
= P
Q
_
σ(W
t+u
−W
t
) +
_
r −D −
1
2
σ
2
_
u ≤ log
_
z
S
t
__
,
⇒P
Q
[S
t+u
≤ z] = P
Q
_
_
W
t+u
−W
t
≤
log
_
z
St
_
−
_
r −D −
1
2
σ
2
_
u
σ
_
_
.
233
APPENDIX C. THE PROBABILITY DENSITY FUNCTION 234
It is known, from the normally distributed independent increment property of the
Wiener process W
t
, that W
t+u
−W
t
follows the same law as
√
uW
1
hence we have
P
Q
[S
t+u
≤ z] = P
Q
_
_
W
1
≤
log
_
z
St
_
−
_
r −D −
1
2
σ
2
_
u
σ
√
u
_
_
,
⇒P
Q
[S
t+u
≤ z] = Φ
_
_
log
_
z
St
_
−
_
r −D −
1
2
σ
2
_
u
σ
√
u
_
_
. (C.2)
where we have used the standard result that P[W
1
≤ y] = Φ(y). The transitional
probability density function is given by the derivative of the above with respect to z,
this can be seen from the deﬁnition,
P
Q
[S
t+u
≤ z[S
t
= S] =
_
z
0
f(S, t; y, t +u)dy,
where f(S, t; y, t +u) is the transitional probability density function of the process at
time t +u and position y, given that it started at S at time t, hence
∂
∂z
_
P
Q
[S
t+u
≤ z[S
t
= S]
_
=
∂
∂z
_
z
0
f(S, t; y, t +u)dy
= f(S, t; z, t +u).
Now directly computing the derivative of (C.2) with respect to z yields
f(S, t; z, t +u) =
1
σz
√
2πu
exp
_
−
_
log
_
z
S
_
−
_
r −D −
1
2
σ
2
_
u
_
2
2σ
2
u
_
under the measure Q.
Contents
Abstract Declaration Copyright Statement Acknowledgements Dedication 1 Introduction 1.1 1.2 1.3 Evidence of increased interest in liquidity . . . . . . . . . . . . . . . . A brief history . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Derivative pricing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.3.1 1.3.2 1.3.3 1.3.4 1.3.5 1.3.6 1.3.7 1.4 1.5 European options . . . . . . . . . . . . . . . . . . . . . . . . . Arbitrage pricing . . . . . . . . . . . . . . . . . . . . . . . . . The FeynmanKac representation theorem . . . . . . . . . . . From FeynmanKac to BlackScholes . . . . . . . . . . . . . . American options . . . . . . . . . . . . . . . . . . . . . . . . . Optimal stopping problems . . . . . . . . . . . . . . . . . . . 11 12 13 14 15 16 17 19 20 21 22 24 25 27 29 30 32 34 35 36 38
Freeboundary problems . . . . . . . . . . . . . . . . . . . . .
Supply and demand economics . . . . . . . . . . . . . . . . . . . . . . Liquidity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.5.1 1.5.2 Deﬁning liquidity . . . . . . . . . . . . . . . . . . . . . . . . . Measuring liquidity . . . . . . . . . . . . . . . . . . . . . . . .
1.6
Price formation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.7 1.8 1.9
Option pricing in illiquid markets: a literature review . . . . . . . . . Introduction to perturbation methods . . . . . . . . . . . . . . . . . . Layout of the thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . .
40 45 46 48 53 53 54 56 56 57 58 58 59 59 60 61 61 64 67 72 77 83 86 87 87 90 92 93 97 99
2 The Modelling Framework 2.1 Technical asides . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.1.1 2.1.2 2.2 Markovian processes . . . . . . . . . . . . . . . . . . . . . . . Applicability of Itˆ’s formula . . . . . . . . . . . . . . . . . . . o
Alternative models . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.2.1 2.2.2 2.2.3 Transactioncost models . . . . . . . . . . . . . . . . . . . . . Reactionfunction (equilibrium) models . . . . . . . . . . . . . Reducedform SDE models . . . . . . . . . . . . . . . . . . . .
2.3
A uniﬁed framework . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.3.1 2.3.2 2.3.3 2.3.4 2.3.5 Cetin et al. (2004) . . . . . . . . . . . . . . . . . . . . . . . . Platen and Schweizer (1998) . . . . . . . . . . . . . . . . . . . Mancino and Ogawa (2003) . . . . . . . . . . . . . . . . . . . Lyukov (2004) . . . . . . . . . . . . . . . . . . . . . . . . . . . Sircar and Papanicolaou (1998) . . . . . . . . . . . . . . . . .
3 Firstorder Feedback Model 3.1 3.2 3.3 Analysis close to expiry: European options . . . . . . . . . . . . . . . Analysis close to expiry: American put options . . . . . . . . . . . . . The vanishing of the denominator . . . . . . . . . . . . . . . . . . . .
4 Fullfeedback Model 4.1 4.2 4.3 4.4 4.5 4.6 4.7 Putcall parity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A solution by inspection . . . . . . . . . . . . . . . . . . . . . . . . . Similarity solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . Perturbation expansions . . . . . . . . . . . . . . . . . . . . . . . . . Numerical solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . Analysis close to expiry . . . . . . . . . . . . . . . . . . . . . . . . . . Numerical results  full problem . . . . . . . . . . . . . . . . . . . . . 4.7.1 A second solution regime . . . . . . . . . . . . . . . . . . . . . 3
5 Smoothed Payoﬀs  Another Breakdown 5.1
102
Local analysis about the singularities . . . . . . . . . . . . . . . . . . 106 5.1.1 5.1.2 5.1.3 5.1.4 5.1.5 5.1.6 5.1.7 Asymptotic matching . . . . . . . . . . . . . . . . . . . . . . . 108 Properties of the inner solution . . . . . . . . . . . . . . . . . 110 Introduction to phaseplane analysis . . . . . . . . . . . . . . 111 Deriving an autonomous system . . . . . . . . . . . . . . . . . 114 Behaviour of the ﬁxed points Structure of the phase portrait . . . . . . . . . . . . . . . . . . 116 . . . . . . . . . . . . . . . . . 120
Other ﬁxed points . . . . . . . . . . . . . . . . . . . . . . . . . 122 127
6 Perpetual Options 6.1
Analytic solutions and perturbation methods . . . . . . . . . . . . . . 131 137
7 Other Models 7.1 7.2 7.3 7.4
Frey (1998, 2000) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137 Frey and Patie (2002) . . . . . . . . . . . . . . . . . . . . . . . . . . . 139 Sircar and Papanicolaou (1998) . . . . . . . . . . . . . . . . . . . . . 140 Bakstein and Howison (2003) . . . . . . . . . . . . . . . . . . . . . . 141 7.4.1 7.4.2 Nonsmooth solutions to the Bakstein and Howison (2003) model146 New nonsmooth solutions to the BlackScholes equation . . . 147
7.5
Liu and Yong (2005) . . . . . . . . . . . . . . . . . . . . . . . . . . . 149 7.5.1 Vanishing of the denominator . . . . . . . . . . . . . . . . . . 150
7.6
Jonsson and Keppo (2002) . . . . . . . . . . . . . . . . . . . . . . . . 152 7.6.1 Connections with the other modelling frameworks . . . . . . . 154 155
8 Explaining the Stock Pinning Phenomenon 8.1 8.2 8.3
Linear price impact . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156 Nonlinear price impact . . . . . . . . . . . . . . . . . . . . . . . . . . 159 A new nonlinear price impact model . . . . . . . . . . . . . . . . . . 161 164
9 The British Option 9.1 9.2
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164 The noarbitrage price . . . . . . . . . . . . . . . . . . . . . . . . . . 167
4
. . . . . . . . . . . . . . . 198 The British put option . .7. . . . . 175 Analysis close to expiry . . . . . . . . .3 9. . . 228 A. . . . .2. . . .1 9. . .1 The linear BlackScholes equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171 Free boundary analysis far from expiry . . . . .2 Analysis far from expiry . . . . . . . . . . . . . . . . . . .6 9. . . . .8 Integral representations of the free boundary . .2. .9. . 196 Analysis close to expiry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 180 Financial analysis of the British put option . . . . . . . . .1 9. . . . . 227 A.7 The gain function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .2 Uniqueness of PDEs . . . . . . .2. . . .3 Monotonicity in λ . . . . . . . . . . . . . . .4 9. 226 A.1 9. . . . . . . . . 198 9. . . . . . . . . . . . . . . . . . . . . . .7. . . . . . 170 Numerical treatment . . . . .2 The American put option . . . . . . . . . . . . . . 186 The British call option . . . . . . . 229 A. . . . . . . . . . . . . . . . . . . . .2 The nonlinear (illiquid) BlackScholes equation . . . . . . . . . . . . . . .5 9. . . . . .8. . . . . . 201 204 223 10 Conclusions A Maximum Principles A. . 198 9. . . . 230 B Nondimensionalisation of the British Put C The Probability Density Function 232 233 Word count 69834 5 . . . 193 9. . . . . . . . . . . .1 Nonlinear equations . . .8. . . . . . . . . .
. . . . . . r = 0. .7. 0. . .4 Asymptotic Matching. . . 79 6 . . σ = 0. . . . K = 1) for λ = 0.2. . location of free boundary (as τ → 0) with λ. .1 Value of European call options with ﬁrstorder feedback (T = 1. . . . K = 1 and for λ = 0. the variation with λ appears to be monotonic. . .7. . σ = 0. . 1. .1. .15.9) (solid line) for τ = 0. .2. . . . 0. K = 1. .04.9 The second derivative (Γ) of the BlackScholes equation (3. . . 79 3. . . .3 3. . r = 0.5 Value of American put options. 0. . . 10. 5.2. . . 2. . .. . 3.04 and σ = 0. σ = 0. . r = 0. . . the variation with λ appears to be monotonic.015. 2. . . . . . . . . Inner solution minus the payoﬀ for put and call options. 3. . . . . . 1. .2. 1. σ = 0.01. . 2. . K = 1) for λ = 0. . . . . . . . T = 1. .4. .2. . . .7 Location of the vanishing of the denominator of (2. .2) (dotted line) and the ﬁrst order feedback PDE (2. . . .1. 10. . σ = 0. .. . .2 Value of European put options with ﬁrstorder feedback (T = 1. 78 76 73 72 66 70 66 The ﬁrst derivative (∆) of the BlackScholes equation (3. .9) with λ = 0. . . . 5. . . .05. . r = 0..04. r = 0. 0. . . 0. .04.2) (dotted line) and the ﬁrst order feedback PDE (2. .04. .2. . 5. . . . K = 1 and for λ = 0. 0. 3. . .9) (solid line) for τ = 0. . . . . . K = 1. . . . 10. the variation with λ appears to be monotonic. r = 0.2. . . . . Compare the location of the vanishing denominator 3. .8 . 3. . . . 0.04. . . .01. . .6 Firstorder feedback put (with early exercise).015. . .05.List of Figures 3. 3. . . Compare the location of the vanishing denominator 3. . . 3. .
3. . . . . . freeboundary location as indicated. . . 4. . .7 Full feedback put. . .95. . . . 0. . . .09 (solid line) and λ = 0. 0. . . . . . . . . . . . . . . λ = 0.2 Perpetual fullfeedback American put. . . .075.1. .2. . 1. . . . .04. . . modiﬁed numerical scheme. . . . v = 0 and the ﬁeld direction lines. . . K = 1. . 0. . . λ = 0. K = 1. . . 4. . 1. . σ = 0. 4. . . 3 cf. . ρ = 0.2.2. . . . . . .0125. . T = 1 and τ = 0. . . K = 1.04.1.2 and λ = 0.10 Firstorder feedback put option value for two diﬀerent values of λ at various times to expiry.04. . . . . . K = 1. . . . . . . .31). . . 0. . . . . . . . .04 and σ = 1. τ = 0. . σ = 0. .2. . 4. . . 0. . . . .24). . . Note the ﬁxed point at u = 243 80 1 3 81 92 93 96 99 99 . .1. . . . .2. . . . . . . . σ = 0. . .04. . .1. . . r = 0. . . . σ = 0.04. . . . 100 4. . . 130 7 . r = 0. σ = 0.04. . .2 Deltas for fullfeedback (European) put. 120 Full feedback American put.1. . . λ = 0. . . . λ = 0) European put option for various time to expiry. . 0. . . . . K = 1 and λ = 0. . . . . . . . K = 1. . . .1 5u . . . . .15. . . r = 0. . . . modiﬁed numerical scheme.25. .1 and τ = 0. . K = 1. . .e. . 1. 0. .5 Full feedback put. .1. . . . . Note that we are in the regime λ < 2ρ and so we should expect no singular behaviour. σ = 0. . .1 and T = 1.01. . τ ) to the BlackScholes (i. r = 0. r = 0. . . σ = 0. . . . 10.1 The leading order correction term V1 (S. . . . .04. K = 1. . . . . . . . .04. . .3 Local (τ → 0) solution of a fullfeedback put. . . . . K = 1. . . . . . . modiﬁed numerical scheme. τ = 0. . .0375. . r = 0. .7. λ = 0. . . . . . σ = 0. . .6 Full feedback call. . . equation (5.04.2. . .1. r = 0. . . λ = 0. . . . .2 and λ = 0. . . . 101 5.2 and λ = 0. .1 (dotted line). . . 4. .15 (smoothed payoﬀ). . .1 Phase portrait of the autonomous system (5. 128 6. . . . . . K = 1. . Compare with ﬁgure 3. .1. 4. r = 0. . . . For r = 0. . . . r = 0. .4 Full feedback put. . .. . . smoothed payoﬀ. The dotted line rep resents an analytic envelope for the phase portrait close to the singular line v = 6. . . .2. . σ = 0. . .
.04. 0.05. .4 5. . . . . . . . . . and λ = −5. . . . . .. . . 139 7.5 Local (τ → 0) put solution of the Bakstein and Howison (2003) model Local (τ → 0) call solution of the Bakstein and Howison (2003) model ˆ K = 1. . 0. r = 0. . . . . . . . . . ˆ model K = 1. . 4. . .04. . . 1. . .04. .1. .3.5. . .5. . and λ = 0 0. 1.2. . .04. . .2 and λ = 0.04. 136 7. α = 1.2. 0. .9) for a put option with λ = 0. σ = 0. . 0. K = 1. .04 and σ = 0.8 Location of the vanishing of the denominator for the Liu and Yong (2005) model for various value of β. .3 The ﬁrst order correction to the BlackScholes perpetual American put option (solid line) compared to the diﬀerence of the fully numerical option value with the BlackScholes (dotted line).1. . . .6 7. . .04. . 5. . . . .4 The ﬁrst order correction to the BlackScholes perpetual American put option (solid line) compared to the diﬀerence of the fully numerical option value with the BlackScholes (dotted line) for various values of λ. . . r = 0.5. .2 and λ = 0. r = 0. . and α = 1. σ = 0. . σ = 0.2. 4. . . . . .2. . . . 141 7. 0. . . 0.01. . and λ = −5. α = 1. 142 7. . . .3 Local (τ → 0) put solution of the Sircar and Papanicolaou (1998) Solution to equation (7. r = 0.. 153 7. . and a = −1. . .9. . K = 1. . . K = 1. 149 7. and λ = 0. . . 0.2. .. 2000) (solid line) and Sch¨nbucher and Wilmott (2000) (dotted line) model o with λ = 0. . . . . . . . .. 1. . . . K = 1. . . .1 Location of the vanishing of the denominator of the Frey (1998. . . λ = 0.1.. r = 0. . σ = 0. 8 . . r = 0. . . . . . . . . σ = 0.75. . . . . . . . . and β = 1 × 105 . r = 0. .1. . . . . . . .9. . . 136 6. . . . . . K = 1. 144 7. .05. K = 1. 5. .2 Local (τ → 0) call solution of the Sircar and Papanicolaou (1998) ˆ model K = 1. . . . . 1. . and a = −1. . . . . . σ = 0. r = 0. . . . . . .2. . . 2 × 105 . . . . . .04. .04. σ = 0. . r = 0. 153 7. . 148 σ = 0. 147 ˆ K = 1. . .75. .7 Nonsmooth solution of the BlackScholes equation. . .2. . . . .2. σ = 0. . σ = 0. . . 151 7. . .9 Local (τ → 0) call solution of the Jonsson and Keppo (2002) model K = 1. . . . r = 0. σ = 0.5.. . . . .5. . . .2. .2. . .10 Local (τ → 0) put solution of the Jonsson and Keppo (2002) model K = 1.04. . 1 × 106 ..6. . . . . . .2.
. T − t = 0. . . . T = 1. . K = 1.05. r = 0. 9. . .. . .27) (dotted line) compared with fully numerical value (solid line). . . . . T = 1.e.01.125. . .7) for p = 0. . . . . . 189 9.1 The pinning probability (8.3 The zero of the Hfunction.1. . and D = 0. r = 0.2. . D = 0. .2.5) (dotted line) and r + 1 σ 2 = 0. . . r = 0.e. . . . . . Sh (t). . . . . . . .8. . 0.8. . . . .125. and σ = 0. .1. 5. T = 1.2 The British put option free boundary for varying volatilities. . . . .4 9. D = 0. . . T − t = 0.102.125. 161 2 1 for T = 0.5. . and σ = 0. . 1. .1. . . K = 1.125. . . . . . . . . 190 9 . . . 0.12. The solid lines denote contours at increments of 10% from 10% to 60%. . σ = 0. . . and µc = 0. . . . .115. .4. . . . The dotted line represents the zero contour. T = 0. µc = 0.11. σ = 0. .2. T − t = 0. . 173 9. K = 1. .1 The British put option free boundary for varying values of the contract drift. . σ = 0. . . . . . σ = 0.4 The asymptotic approximation for the British put option free boundary close to expiry. .4. . . . . . . . . . .1. . . . . .4. . . . and D = 0. . 0. . S0 = 1. . K = 1. 1. . µc = 0. . 163 8. . . . . D = 0. . 0. K = 1. . . σ = 0.4.1.1. . . . . 158 Comparing the pinning probability associated with (8. .1. . . . . and µc = 0. . . . . 186 9. . . . . µc = 0. .1. . . K = 1. . . r = 0. . . i.1. . . D = 0. . .2. . . .6 The diﬀerence in the percentage return of the British put option and the American put option at every possible stopping location.8. . . . . . .8) (solid line) compared to (8. .2 8. . i.9. . .7 and 9. K = 1. . . . . . .5) for values of nE = 0. . .104. and σ = 0. . .. . 173 9. . and r + 2 σ 2 = 0. . 159 2 8. . . . . . µc = 0. . 0. . . .2 Solution to equation (8. . K = 1. 0. . r = 0. . .3 Solution to (8. .1. . . . . . . . . T = 1. . . . . . (9. . . . . . . σ = 0. . . . σ = 0.5.. . . r = 0. .6. . . . T = 50. . . . 1. and r + 1 σ 2 = 0. . . . .16. . . . . K = 1.1. . . .5 Location of the free boundary for the British (solid line) and American (dotted line) put option under investigation in ﬁgures 9.1.4. for varying values of the contract drift. 0.6) (solid line) with the model of Avellaneda and Lipkin (2003) (dotted line) for nE = 0. . 177 9. . .
08. . . T = 1. . . . .1. . . . r = 0. K = 1. r = 0. 0. . . .13 The asymptotic approximation for the British call option free boundary close to expiry. . 0. . . . . . The dotted line represents the zero contour. . . . .9 Schematic representation of the regions in which atthemoney European. . σ = 0.1.09. . r = 0. . . σ = 0. . . . . . . . . . K = 1. .1. . and D = 0.4. σ = 0. . .01. . . . . . S0 = 1.4. 190 9. . . S0 = 1. and µc = 0.11 The British call option free boundary for varying values of the contract drift. . . . . . . for increasing moneyness. . . σ = 0. .. . and µc = 0. . American and British put option would provide the greatest return on an investment. Note the change of orientation.1. .4.9. . . r = 0. . . .05. . . . . . . . . K = 1. 0. 195 9. .1. .32) (dotted line) compared with fully numerical value (solid line). . D = 0. . . . . σ = 0. . K = 1. 192 9.10 Figures representing the region in which American put options would provide a greater expected return that its British option counterpart. . . T = 1.1 and D = 0.. Again the solid lines denote contours at increments of 10% from 0% to 70%. .4. The dotted lines represent the free boundaries of the American and British put option for reference. T = 0. . 0. . T = 1. r = 0. 9. . . (9. 0. .4. . . . . .5. . . µc = 0. . . . i. . . . . . . . . .06. T = 1. The solid lines denote contours at increments of 10% from 70% to 30%. . . . . . . . . σ = 0. . . . . 191 9. . . . .1. . . µc = 0. .7 The diﬀerence in the percentage return of the British put option and the European put option. r = 0. D = 0. .125. . . .05. . K = 1.e. .12 The British call option free boundary for varying volatilities. . . . . . . The dotted line represents the zero contour. T = 1. . . . . 195 9. . . . T = 1. . . . . and σ = 0. . . . . K = 1. . 199 194 10 . .4. .1 and D = 0.055.8 The diﬀerence in the percentage return of the American put option and the European put option. D = 0. . K = 1. .08 and D = 0. r = 0.
In the ﬁrst class.The University of Manchester Kristoﬀer John Glover Doctor of Philosophy The Analysis of PDEs Arising in Nonlinear and Nonstandard Option Pricing October 23. 2008a. which can help to mediate the eﬀects of a ﬁnitely liquid market. and perhaps even more disturbingly.b). Detailed analysis reveals that the form of the nonlinearities introduced can lead to serious solution diﬃculties for standard (put and call) payoﬀ conditions. speciﬁcally for times close to. 2008 This thesis examines two distinct classes of problem in which nonlinearities arise in option pricing theory. a new nonstandard class of early exercise option. detailed asymptotic analysis. we investigate the properties of the recently introduced British option (Peskir and Samee. In both classes. since the contract does not require the holder to enter the market and hence incur liquidation costs. which for the most part result in highly nonlinear partial diﬀerential equations (PDEs). negative option values. In particular. One is associated with the inﬁnite gamma and in such regimes it is necessary to admit solutions with discontinuous deltas. we consider the eﬀects of the inclusion of ﬁnite liquidity into the BlackScholesMerton option pricing model. It is concluded in this case is that the model irretrievably breaks down and there is insuﬃcient ‘ﬁnancial modelling’ in the pricing equation. The repercussions for American options are also considered. and far from. we investigate a model studied by Sch¨nbucher and Wilmott (2000) and furthermore. o show how many of the proposed existing models in the literature can be placed into a uniﬁed analytical framework. expiry. 11 . In the second class of problem. coupled with advanced numerical techniques (informed by the asymptotics) are exploited to extract the relevant dynamics. Here we choose to focus on the interesting nonlinear behaviour of the earlyexercise boundary. A second failure (applicable to smoothed payoﬀ functions) is caused by a singularity in the coeﬃcient of the diﬀusion term in the optionpricing equation.
12 .Declaration No portion of the work referred to in this thesis has been submitted in support of an application for another degree or qualiﬁcation of this or any other university or other institute of learning.
the Copyright and any Intellectual Property Rights and/or Reproductions described in it may take place is available from the Head of the School of Mathematics. for example graphs and tables (“Reproductions”). publication and exploitation of this thesis. trade marks and any and all other intellectual property rights except for the Copyright (the “Intellectual Property Rights”) and any reproductions of copyright works. Such Intellectual Property Rights and Reproductions cannot and must not be made available for use without the prior written permission of the owner(s) of the relevant Intellectual Property Rights and/or Reproductions. 13 . This page must form part of any such copies made. educational and/or teaching purposes. either in full or in extracts.Copyright Statement i. Copies of this thesis. promotional. The author of this thesis (including any appendices and/or schedules to this thesis) owns any copyright in it (the “Copyright”) and s/he has given The University of Manchester the right to use such Copyright for any administrative. The ownership of any patents. Details of these regulations may be obtained from the Librarian. ii. iii. designs. may be made only in accordance with the regulations of the John Rylands University Library of Manchester. iv. Further information on the conditions under which disclosure. may not be owned by the author and may be owned by third parties. which may be described in this thesis.
Duck and David P. Philip Haines. EPSRC funding is gratefully acknowledged. In particular. and to Erik Ekstr¨m for o his insight and friendship.Acknowledgements I am extremely grateful to my supervisors Professor Peter W. I thank my parents for their love and unwavering support for which these mere expressions of gratitude do not suﬃce. enthusiasm and eﬃciency. Thank you to my colleagues and friends for their invaluable advice and numerous enlightening discussions. I hope. In addition. John Heap. I thank Hannah for everything. Helen Burnip. Sebastian Law and Vicky Thompson. To my close friends. I hope we both ﬁnd what we’re looking for. In particular. I thank Peter for his boundless knowledge. both old and new.D. despite the distances between us. 14 . our friendships can continue to ﬂourish. I thank you for creating the good times and for being there through the bad. Newton for their expert guidance and continued support throughout the course of this Ph. and in particular to Jonathan Causey. to Goran Peskir for his time and enthusiasm for the subject. Finally. and David for his caring supervision and his conﬁdence in my abilities.
15 . in loving memory.Dedication To Gran.
One of the more subtle was that the market in the underlying asset1 was inﬁnitely (or perfectly) elastic. However. much of the work undertaken in mathematical ﬁnance has been aimed at relaxing a number of the modelling assumptions. . Since the deﬁnitive papers of Black and Scholes (1973) and Merton (1973). this does not preclude us from trying to quantify the ﬁnancial markets and to utilise the powerful tools of mathematics in order to better understand such markets. 16 .Chapter 1 Introduction Nowadays people know the price of everything and the value of nothing. If we relax this assumption. but the nature of man and his markets. this is partly due to the fact that any model incorporating such a feature will inevitably lead to nonlinear behaviour (feedback). then we see some rather interesting and possibly counterintuitive behaviours. In particular. As we shall show later. we shall be concerned for the most part with nonlinear partial diﬀerential equations (PDEs) arising from the study 1 Termed underlying in the sequel. such that trading had no impact on the price of the underlying. one is not trying to model Mother Nature and her laws.Oscar Wilde (18541900) The Picture of Dorian Gray (1891) Mathematical ﬁnance is not a branch of the physical sciences. There are no laws of nature just waiting to be discovered.
geometric Brownian motion. One approach to incorporate these dynamics is to ﬁnd a stochastic process that ﬁts most closely the distribution of returns of the underlying. and as such provides little insight into which of the many factors aﬀecting the price dynamics are actually the most important. This provides much greater insight into how prices are actually formed in the market. and as a result are also highly nonlinear. 1. The aim of modelling the behaviour of the underlying is to capture the dynamics of the observed market prices as faithfully as possible. These models involve optimisation over all possible values of volatility. there is the so called BlackScholesBarrenblatt equation introduced by Avellaneda et al. In this chapter we introduce the basic ideas and concepts and review the results of the classical BlackScholesMerton option pricing theory used in later chapters. In addition. An alternative approach (and e that to be followed in this thesis) is to retain one the simplest stochastic process. and has the advantage of being consistent with the bulk of the literature over the past thirtyﬁve years. but to provide an endogenous mechanism by which the dynamics diﬀer from this standard geometric Brownian motion. It is by no means a complete treatment of the relevant theories. for example L´vy processes. which was one of the ﬁrst nonlinear PDEs to arise in the ﬁeld of mathematical ﬁnance. (1995) in the study of uncertain volatility models.CHAPTER 1. the exogenous processes required tend to be diﬃcult to handle mathematically. Work that has led to this class of PDEs in ﬁnance to date includes Whalley and Wilmott (1993) in relation to transaction costs. In addition. This is an exogenous strategy. just enough for the unfamiliar reader to understand the contributions of the following chapters. INTRODUCTION 17 of ﬁnitely elastic markets. This has resulted in trading volumes .1 Evidence of increased interest in liquidity Recent worries about the health of the modern ﬁnancial system have deterred people from getting involved in the derivatives markets.
In a recent blog entry regarding the subprime induced liquidity crisis Paul Wilmott states that5 . This stems from another key theme now haunting the markets: namely that liquidity is evaporating from numerous corners of the ﬁnancial world.. INTRODUCTION 18 decreasing and hence increased liquidity problems.. January 29 2008. 3 2 .. August 1 2007.as market turmoil rises ﬁnancial problems are no longer simply conﬁned to a risky corner of the US mortgage market. FT. motivating further investigation into the eﬀects of reduced liquidity on all aspects of the ﬁnancial markets.com. November 23 2007... 4 See Liquidity alarm bells sound. February 27 2008.. 5 Quoted in Science in Finance IV: The feedback eﬀect Paul Wilmott. blog entry at http://www.. Paul J Davies.this should spur on the implementation of mathematical models for feedback. Joanna Chung and StacyMarie Ishmael. The current liquidity crisis can be traced back to the collapse of the US subprime mortgage market.. Chris Giles. Gillian Tett. which may in turn help banks and regulators to ensure that See Derivative liquidity crisis ‘to continue’. David Oakley of the Financial Times2 warns that .com/blogs/paul/. In August 2007 the Financial Times is quoted as saying that4 . FT. David Oakley. FT.. Clearly. Quoted in Bank deputy downbeat on economy. Rachel Lomax.com.the largest ever peacetime liquidity crisis.. the Bank of England’s Deputy Governor goes on to describe the recent ﬁnancial turmoil in the wake of the American subprime mortgage problems3 as .the sharp slowdown in these [derivative] markets is a serious warning sign of the growing problems in the ﬁnancial world as they are usually highly liquid. as both investors in hedge funds and the banks that lend to them try to cut and run from recent losses. Further. turning over vast amounts of trade every day. liquidity becomes an ever important issue.CHAPTER 1.com. in times of crisis. wilmott.
For a translated version with commentary and a foreword by Paul Samuelson see Davis and Etheridge (2006). which rests ultimately on the Central Limit Theorem. INTRODUCTION 19 the press that derivatives are currently getting is not as bad as it could be.CHAPTER 1.6 The origins of much of ﬁnancial mathematics trace back to a dissertation (entitled Th´orie de la sp´culation 7 ) published in 1900 by Louis Bachelier (18701946). which gives a stochastic See for example Jacod and Protter (2003). 1. In fact. The work of Richard P. and it was not until the 1920s that the rigorous mathematical underpinnings of the theory of Brownian motion was provided by Norbert Wiener (18941964). wherever we look we see a random world and therefore Brownian motion (named in honour of Robert Brown) is an invaluable tool for describing this randomness. Feynman (19181988) in the late 1940s on quantum mechanics using path integrals.2 A brief history In 1828 Robert Brown (17731858). the ubiquitous nature of Brownian motion can be seen as the dynamic counterpart of the ubiquitous nature of the normal distribution. a Scottish botanist. In e e it he proposed to model the movement of stock prices with a diﬀusion process or Brownian motion. as quantum mechanics emerged in the 1920s it began to become clear that the quantum picture is both inescapable at the subatomic level and intrinsically probabilistic. whose aggregate eﬀect was (apparently) random. Meanwhile. Feynman’s work was made mathematically rigorous by Mark Kac (19141984) and the socalled FeynmanKac formula. Note that this was ﬁve years before Einstein’s seminal paper outlining the theory of Brownian motion. observed the apparently random motion of pollen particles suspended in water and subsequently during the 19th century it became clear that the pollen particles were being bombarded by a multitude of molecules of the surrounding water. introduced the Wiener measure into the heart of quantum theory. 7 6 . In addition.
3). 1. was introduced (see section 1. the machinery o needed in order to use Brownian motion to model stock prices successfully. Scholes and Merton wrote down their famous equation for the price of a European call and put option in 1969. INTRODUCTION 20 representation for the solution to certain classes of PDEs.3. work for which the surviving members (Scholes and Merton) received the Nobel Prize for economics in 1997. Alternatively. However.3. the socalled martingale approach and the PDE approach. and which would later become an essential tool of modern ﬁnance. a stochastic process for the underlying is speciﬁed and an equivalent probability measure is found that turns the discounted underlying into a martingale. it was not until 1965 that economist Paul Samuelson (1915) resurrected Bachelier’s work and advocated Itˆ’s geometric Brownian motion model as a suitable model for stock price movements. this thesis deals solely with continuous time models.3) and it should be noted that both approaches can be used for complete . In the former. In 1944 Kiyoshi Itˆ (1915) went on to develop stochastic calculus. A more comprehensive overview of the early years of mathematical ﬁnance can be found in Jarrow and Protter (2004). in the PDE approach. In continuoustime modelling there are two main approaches to calculating the price of a given derivative security.CHAPTER 1.3 Derivative pricing When we discretise a problem it becomes easier to deﬁne or understand but much harder to solve without the use of continuous time calculus. The price of the derivative is then deﬁned as the conditional expectation of its discounted payoﬀ under this new (riskneutral) measure. a stochastic process for the underlying is likewise speciﬁed and then Itˆ’s formula for a function of the underlying stochastic o process is used to derive a PDE involving the coeﬃcients of the underlying process. The two approaches are deeply linked via the famous FeynmanKac formula (outlined in section 1. o After this it was not long until Black.
The holder of a call option written on a certain underlying asset (usually a stock) has the right. T ) = (ST − K)+ := max{ST − K. The derivative securities studied in this thesis.e. are options contracts. then the holder would not exercise. without exception.CHAPTER 1. is worth more then K then the (rational) holder would exercise the option and make a proﬁt ST −K. If the underlying at time t = T . The fair price of a derivative security (and all other ﬁnancial instruments) is determined by the expected discounted value of some future payoﬀ. arriving at a unique price for a derivative requires additional assumptions. whereas using the PDE approach the choice of martingale measure is analogous to specifying the socalled market price of risk of the nontraded variable. Thus. all sources of risk are traded. denoted K. but not the obligation. 1.1) Under suitable restrictions. Since the models introduced in this thesis result in complete markets. which is itself dependent on the future value of the underlying asset. both the martingale approach and the PDE approach should arrive at the same price. the future value of the underlying is not known a priori. then this arbitrariness is reﬂected in the choice of equivalent martingale measure. A brief overview of the types of contracts referred to in the main body of the thesis will be considered next. and price processes are often modelled by stochastic processes.1 European options European options are the simplest type of options contract and within this class the most common are call options and put options. resulting in the option expiring worthless. 8 (1. 0}.3. . see chapter 2. If one is using the martingale approach. denoted T . if ST is less than K. Therefore. INTRODUCTION 21 and incomplete markets. this section attempts to provide such an understanding. an understanding of the behaviour of such stochastic processes is a valuable prerequisite for the study of derivative pricing. and at some predetermined price. to buy the underlying at some predetermined date. the value of the call option at expiry (T ) is given by V C (ST . In the latter case.8 i. Alternatively. Of course. ST .
T ) = (K − ST )+ := max{K − ST .2 Arbitrage pricing An arbitrage opportunity corresponds to a riskfree proﬁt. For an option to be described as European.e.1) and (1. resulting in the value of the put option: V P (ST .2) are called payoﬀ proﬁles and will be referred to as such throughout this thesis.3) .3. h(ST ) say. To state their results. Options that allow exercise at times prior to expiry are said to have an earlyexercise feature. we have a market consisting of a bank account which grows according to the (deterministic) dynamics dB = rBdt. of course. (1.3. More speciﬁcally if the option allows exercise at any time prior to expiry such an option is referred to as an American option. and one risky asset. this results in tractability in many situations. In an eﬃcient market there should be no such arbitrage opportunities and indeed the seminal work by Black and Scholes (1973) and Merton (1973) used the noarbitrage principle to arrive at a unique price for the fair value of an option contract. with stochastic price dynamics dSt = µSt dt + σSt dWtP . it is the opportunity to construct a trading strategy (i.5. Indeed we shall return to them shortly in section 1. ST . Note that these contracts dependent only on the price of the underlying at expiry.2) The functions (1.CHAPTER 1. More formally. 0}. and not on the path of the price prior to maturity. (1. many diﬀerent options contracts with more general payoﬀ proﬁles. T . INTRODUCTION 22 Similarly. There are. its contract must specify that exercise is only possible at a single maturity time. and will play an important role in much of this thesis. buying and selling ﬁnancial instruments) in such a way that the initial investment (at t = 0) is zero and the wealth at time T is nonnegative with a nonzero probability of a strictly positive wealth. These options are very popular in practise. 1. the holder of a put option has the right to sell the underlying for the exercise price K.
WtP denotes a standard Brownian motion under the probability measure P .01 it would take ∼ 10. 10 In fact. deﬁned by the process (1.4) in words. any security whose payoﬀ h(ST ) is known at time T (where h(ST ) is any FT measurable This subtlety was the main innovation of option pricing research in the 1970s.3). Q.3) can then be described in terms of a standard QBrownian motion WtQ as dSt = rSt dt + σSt dWtQ . except that the drift of St under Q is equal to the interest rate r instead of µ. S. the stock price process (1. µ the drift and σ the volatility of the underlying price process. In this model. The indices indicate that the process for St is started at S at time t and also that the expectation is calculated under the socalled riskneutral probability measure.10 The model analysed above is an example of a complete market model.CHAPTER 1. P . Consequently the drift parameter µ does not appear anywhere in the pricing formula for European options. t) = EQ e−r(T −t) h(ST ) . 9 . (1. Prior to this.t (1. The fair value or price of a European option contract V (S. since in practise the drift parameter is notoriously diﬃcult to measure from past time series of the underlying process. the expected discounted future payoﬀ. The simplest deﬁnition of a complete market is one in which every derivative security can be replicated by a selfﬁnancing trading strategy in the stock and bond. INTRODUCTION 23 where r is the positive (constant) interest rate. as opposed to the real world measure. this fact undoubtedly contributed to the widespread application of the BlackScholesMerton pricing methodology in the years subsequent to its publication. 000 years to obtain such an estimate. For example if = 0. expectations had been taken under the real world measure P .5) Note that the dynamics of St under the riskneutral measure Q are the same as the dynamics under the realworld measure P .9 The riskneutral measure is deﬁned as the unique measure equivalent to P under which the discounted price process is a martingale. Liptser and Shiryaev (2001) show that the expected waiting time to obtain an estimate of the drift (via the naive approximation St /t) that is within of the true drift is proportional to −2 . Consequently. t) with payoﬀ proﬁle h(ST ) can be shown to be given by V (S.
7) The indices on the expectation indicates that the process St is started at S at time t and in addition the superscript indicates that the expectation is taken under the probability measure P .3. such as the pricing equation (1.4). In the following section we describe the FeynmanKac representation theorem. this thesis focuses primarily on the latter. t)dWtP . t) 2 + µ(S. T ) = h(S). t) (1. t). t)dt + σ(St . are linked to the solution of (linear) parabolic partial diﬀerential equations (PDEs) via the famous FeynmanKac representation theorem. ∂t 2 ∂S ∂S Suppose we are given the PDE for the unknown function u(S. 1983). a characterisation of the arbitragefree principle is that there exists a unique equivalent martingale measure Q. σ(S. u(S. INTRODUCTION 24 random variable with E [h2 (ST )] < ∞) can be replicated by some unique selfﬁnancing trading strategy. t) = EP [h(ST )] S.7). Thus. The FeynmanKac formula tells us that the solution can be written as an expectation. This equation is sometimes called the Kolmogorov backward equation. Finally. t) = 0. we note that in a complete market. with a P Brownian . the price of a European option can be studied using both stochastic methods and parabolic PDE methods. (1.6) subject to the ﬁnal condition u(S.3 The FeynmanKac representation theorem ∂2u ∂u ∂u 1 2 + σ (S. where µ(S. corresponding to the stochastic process (1. For more on this characterisation see the original works of Harrison and Kreps (1979) and Harrison and Pliska (1981.t where St is a stochastic process given by the equation dSt = µ(St . Expected values of solutions to stochastic diﬀerential equations (SDEs). t) and h(S) are known functions and T a parameter.CHAPTER 1. 1. under which the discounted prices of traded securities are martingales.
T ) − u(St . t) t ∂u dWtP . 4.3.4) involves discounting and so it is useful to make the transformation V (S.t 1. S. T )] = EP [h(ST )] . t) is assumed to be the solution of the PDE (1. The proof of the FeynmanKac representation is fairly straightforward and so we shall outline the basic idea here. t) = EP S.t [u(ST . t) ∂u dWtP . t) 2 ∂t ∂S 2 ∂S dt + σ(S. t) . Proof. 2004) resulting in the o required result u(S.CHAPTER 1. Consider an unknown function u(S. by assumption the O(dt) terms above are zero if u(S.t S.6). t). t) ∂u dWtP . prop. Integrating the above equation we obtain T t T du = u(ST . as the solution to a secondorder linear parabolic PDE. t) + σ (S. taking expectations and reorganising a little we arrive at T u(S. conversely. ∂S Finally. for example. representing the price of a European option. Applying Itˆ’s formula we have o du = ∂u 1 2 ∂2u ∂u + µ(S. ∂S Next.4 From FeynmanKac to BlackScholes Having satisﬁed ourselves of the validity of the FeynmanKac representation theorem. expectations of functions of stochastic processes via deterministic PDEs. t) = e−r(T −t) u(S. The ﬁrst point to note is that (1. t) = σ(S. T )] − EP S. it can be shown that the expectation of an Itˆ integral with respect to a o Brownian motion is zero (see.t t σ(S.4 of Bj¨rk.4). This useful representation allows us to solve deterministic PDEs via stochastic methods and. ∂S Now. INTRODUCTION 25 motion WtP . we can now use it to represent the expectation given in (1. t) = EP [u(ST .
8) where we have used the riskneutral process (1.4) is taken under the riskneutral measure Q and so the corresponding PDE representation of (1. this problem is often glossed over or simply not mentioned in the literature.6) to obtain the PDE 1 ∂2V ∂V ∂V + σ 2 (S. T ) = h(S).5). INTRODUCTION 26 in equation (1.9b) (1.4) is given by ∂V 1 ∂2V ∂V + σ 2 S 2 2 + rS − rV = 0.9). However. if we assume a stochastic process of the much more general form (1. S. .e. Remarkably. assumptions that are often not satisﬁed by many models used in practise.CHAPTER 1.9c) (1. t) 2 + µ(S. t). What follows is a brief overview of the some of these analytic conditions. ∂t 2 ∂S ∂S (1. ∂t 2 ∂S ∂S with the following condition V (S. Moreover. t) = EP e−r(T −t) h(ST ) .10) with the same boundary conditions as previously. (1. V (S.7). t) → h(S)e−r(T −t) as S → ∞. V (0. t) − rV = 0. ∂t 2 ∂S ∂S which we have shown can be represented as the conditional expectation V (S. In some sense the 11 Again note the independence of the realworld drift µ(S.t However. i.9a) (1. it can be shown that standard FeynmanKac type results only hold under (quite restrictive) analytic conditions on the coeﬃcients of the SDE and PDE. t) 2 + rS − rV = 0. then the corresponding (generalised) BlackScholes equation obtained via the FeynmanKac formula is given by11 LBS (V ) = ∂V 1 ∂2V ∂V + σ 2 (S. (1. note that the expectation in (1. t) = h(0)e−r(T −t) . Note that in what follows this shall be referred to as in the BlackScholes equation (which should also be credited to Merton).
t) > 0 ∀(S.9) then the diﬀusion coeﬃcient σ(S. i. The time γ at which The stochastic process derived in chapter 2 can be seen to exhibit singular behaviour and. where Ω is the domain of the process St .4) to be the unique classical solution to the BlackScholes equation (1. t) cannot degenerate be zero. 12 . Hence.CHAPTER 1. In order for the conditional expectation (1. t) ≤ D(1 + S) for some constant D > 0. In other words. i. t) ∈ Ω × [0. t) = σS. as such. these conditions are no longer satisﬁed. Note that even in the simplest cases. T ]. the volatility term degenerates in certain regions of state space. and also satisfy a linear growth condition in S. jumps may be seen at the location where the diﬀusion coeﬃcient becomes singular.5 American options Unlike European options discussed in section 1. t) = 0. T . such as geometric Brownian motion where σ(S. t) = σ which is no longer degenerate. for example Ω = {S > 0} for geometric Brownian motion. INTRODUCTION 27 behaviour of the models presented in this thesis can be attributed to the failure of the coeﬃcients of the relevant equations to satisfy the conditions outline below. t) must be strictly positive at every point in the solution domain (S. t) − σ(S2 . meaning (in this onedimensional situation) that the coeﬃcient σ(S.e. 1. we have the restriction that σ 2 (S. σ(S1 .3. we are no longer in a regime where standard results from SDE and PDE theory can be applied. the diﬀusion coeﬃcient σ 2 (S. We can avoid this diﬃculty here (and also in many other more general situations) by making the change of variable x = log S giving σ(x.3.12 Another condition is that the operator LBS must be uniformly elliptic. i.10) with the conditions (1. t). t) must be suﬃciently regular. More precisely. American options have the extra feature that they can be exercised at any time prior to expiry. Speciﬁcally limS→0 σ(S.e.1. σ(S. t) ≤ CS1 − S2  for some C > 0. In addition. it must be locally Lipschitz.e. here the nonLipschitz nature of the coeﬃcients means that the solutions to the corresponding SDE need no longer remain continuous.
2006). the information contained in the ﬁltration Ft ). Peskir and Shiryaev. t) = sup EQ e−r(γ−t) h(Sγ ) . Samuelson visited many practitioners on Wall Street prior to writing his paper. t) ≥ h(S) (1.CHAPTER 1.11) we have the inequality V (S. he used the European and American preﬁxes but reversed the ordering. Immediately from the deﬁnition (1. The practitioner commented that only the more sophisticated European mind (as opposed to the American mind) could understand the former. when Samuelson (an American) wrote the paper. the holder of the option needs to decide whether to exercise at each point in time based only on the information up to time t ≤ T (i. the supremum of the expected value of the discounted payoﬀ over all random times γ that are stopping times with respect to the ﬁltration generated by the Brownian motion used to specify the dynamics of the underlying process for St . S.e. In response.e. This is a natural condition since if V (S. t) < h(S) then there would be an obvious instant arbitrage at time t. If the payoﬀ proﬁle is given by h. According to a private communication with Robert C. Merton. one more complex (that could be exercised early) and one much simpler (that could only be exercised at expiry). . the unique noarbitrage price of an American option can be shown to be given intuitively by V (S. Using the theory of optimal stopping (cf. INTRODUCTION 28 the option is exercised is called the exercise time and because the market cannot be anticipated. and the holder of the American option decides to exercise early then she receives the amount h(Sγ ) at time γ. This is a rather intuitive deﬁnition of the American option price.t t≤γ≤T (1.12) since the stopping time γ = t is included in the supremum. The terms European and American were ﬁrst coined in Samuelson (1965) and the story behind their naming is noteworthy. One of his industry contacts explained to him that there were two types of options available.11) i.
t) is the corresponding European option price. in general. Thus. the price of an American call reduces to the price of a European call. Another point to note is that when pricing American options we cannot. In o other words. INTRODUCTION 29 In addition. for nonnegative interest rates and no dividends. where V E (S. since an American option gives its holder more rights than the corresponding European option with the same payoﬀ function and expiration date.11) is attained for the stopping time γ = T when considering the payoﬀ function of a call option. which can be done for their European counterparts. set the interest rate to zero. When faced with an optimal . this is intuitive. 2004). Hence the value of the American contract cannot coincide with that of its European counterpart. the supremum in expression (1.12). which does have an explicit formula. violating the condition (1. choosing γ = T gives the further inequality V (S. t) ≥ V E (S.CHAPTER 1. ﬁrst derived by Black and Scholes (1973).2) for suﬃciently small S. Pricing American derivatives is mathematically more involved than the European case and closedform expressions for American option prices are rarely obtained. it can be shown by noarbitrage arguments that. the price of an American call option is the same as its corresponding European option (see. Indeed it can be seen (directly from its wellknown analytic expression) that the European put option price crosses below the payoﬀ function (1. strictly higher than the price of the corresponding European put option.14 of Bj¨rk. for example.6 Optimal stopping problems The observant reader may have noticed already that there is a strong link between pricing American options and optimal stopping problems. It can also be shown that the price of an American put option is. However.3. prop. Again. t). without loss of generality. We therefore use a put option as our canonical example of an American option throughout the remainder of this thesis. 1. 7.
If the function h is continuous. the ﬁrst time that the price of the American option drops down to the value of its payoﬀ. t) : V (S. in other words to determine the stopping time that realises the supremum in (1. 1. in addition to some other technical conditions. the optimal stopping time γ ∗ can be formulated as the ﬁrst exit time from the continuation region deﬁned by C := {(S. INTRODUCTION 30 stopping problem.3). Peskir and Shiryaev (2006) . Alternatively.e.CHAPTER 1. u) = h(Su )}. / The continuation region is so named due to the fact that in this region it is not optimal to exercise the option. u) ∈ C}. if the value V (S. For a nonnegative payoﬀ function h. there are two facets of the solution that we are most interested in. the price of American options can be shown to satisfy partial diﬀerential inequalities. the price of an American 13 See. as γ ∗ := inf{u ≥ t : (Su . for example. i.7 Freeboundary problems Analogous to the FeynmanKac representation theorem for European options (outlined in section 1.13 then the supremum is attained for the stopping time γ ∗ := inf{u ≥ t : V (Su . then it is not optimal to exercise the option. but ﬁrst we state a key result from the theory of optimal stopping. t) at some time t is strictly larger than the payoﬀ proﬁle h(S).3.e.11). Clearly.3. and more practically. Determining the value function will be discussed shortly. t) > h(S)}. i. The ﬁrst is to determine the price of the option V (called the value function in optimal stopping terminology) and the second to determine the optimal strategy for the option holder.
13a) (1. h(S. t) : 0 ≤ t ≤ T. (1. the free boundary. ∂t 2 ∂S 2 ∂S LBS (V ).14e) to be solved in the domain {(S.1 diﬀerentiable. V (S. (1. it can be shown that the BlackScholes equation holds at all points in the continuation region and that at the boundary of the continuation region. Further to this.CHAPTER 1. t) − V (S. t) = −1. separating the continuation region from the stopping region.14 This principle states that the value function V (S. but also over the boundary of the continuation regions. ∂t 2 ∂S ∂S V (Sf . INTRODUCTION 31 option as deﬁned in (1.13) can be formulated as the freeboundary problem 16 1 ∂2V ∂V ∂V + σ 2 S 2 2 + rS − rV = 0. t) = 0. VS (Sf . S > Sf (t)}. t). 15 At least for points at which the payoﬀ proﬁle h(S. t) is C 1.1 diﬀerentiable.14a) (1. t) > (K − S)+ . t) → 0 as S → ∞.13b) (1. This is alluded to in Peskir (2005b). t) = K − Sf . It also transpires that for a standard American put option there is an increasing function Sf (t). T ) = (K − S)+ . we must apply the smooth pasting or smooth ﬁt principle.14d) (1.15 not only in the continuation regions. in other words the boundary of the domain is to be solved as part of the problem. This implies that for S > Sf (t) the value V (S. T ) = h(S). t) must be at least C 1. t) ≥ h(S.14b) (1. 0 ≤ t ≤ T } with the ﬁnal condition V (S. to be solved in the entire domain {(S. t) : S > 0. and for S ≤ Sf (t) the In fact the principle of smooth ﬁt in probability. the principle of no arbitrage in ﬁnance and the conservation of energy law in the physical sciences can be seen as diﬀerent formulations of the same principle. denoted by ∂C.14c) (1.11) is given by the solution to the following linear complementarity problem: V (S. compare Jacka (1991). 16 See for example Karatzas and Shreve (1998) 14 . t) must satisfy V (S. As such the linear complementarity problem (1.13c) ∂V 1 2 2 ∂2V ∂V LBS (V ) = + σ S + rS − rV ≤ 0. V (S.
the price at which supply matches demand is often called the equilibrium price or market clearing price.4 Supply and demand economics Many of the models presented in this thesis make assumptions about the structure of the markets and the intentions of the participants of these idealised markets. Chen et al. these relationships are often called the law of demand and the law of supply. The asymptotic behaviour of Sf (t) for times close to expiry can also be determined and indeed this shall be expounded upon in further detail in chapter 9. the existence and uniqueness of the freeboundary problem (1. nothing more. then the forces of demand and supply will . we should expect an inverse relationship between price and quantity demanded. This motivates a brief discussion of how prices are actually formed in these markets. in short a discussion of supply and demand. 1. 1991) that the American put option free boundary Sf (t) is a monotonically increasing function and that it approaches K as t approaches T . so called because it is at this price that all the surpluses are cleared from the market and the forces of demand and supply are not acting to change this equilibrium. If disequilibrium exists. Conversely. the backbone of a market economy. a market is a place where buyers (providing demand) and sellers (providing supply) meet. Explicit solutions to parabolic freeboundary problems are rare. however it can be shown (cf. (2008) have recently proved the convexity of the resulting free boundary. hence. hence we should expect a positive relationship between price and supply. In addition. In the economics literature.14) can be proved. Jacka. Starting with the basics. INTRODUCTION 32 value satisﬁes V (S. In a free market. In a market. all being equal. prices are determined solely by the interaction of demand and supply. an increase in price will usually lead to an increase in the number of people wishing to sell at that price. there will be more demand for an asset at a lower price than at a higher price and.CHAPTER 1. In addition. Furthermore. nothing less. for put options without dividends. t) = (K − S)+ .
though there is a tendency to confuse the two. x2 . The limiting cases PED = 0 and PED = ∞ imply that the asset is perfectly price inelastic and elastic respectively. and with excess supply. . prices will be forced downwards.e. INTRODUCTION 33 automatically adjust the market to equilibrium. At its heart this concept is a purely mathematical one which aims to measure the responsiveness of one variable to a change in another variable. . With excess demand. whereas liquidity is concerned with the availability to trade the underlying asset at a given price. More speciﬁcally given any functional relationship y = f (x) the point elasticity. However (unlike elasticity) liquidity is not a welldeﬁned concept. is deﬁned as = dy x d(log y) dy/y = = . . The next section explores the concepts of liquidity in much more detail. given a function of more than one variable y = f (x1 . prices will be forced upwards due to the shortage that exists. PED > 1 implies that the good is price elastic. PED < 1 implies that the good is price inelastic and when PED = 1 we have unit elasticity. xn ) the partial point elasticities are given by i = ∂(log y) ∂y xi = . . The price elasticity of supply (PES) is deﬁned similarly.CHAPTER 1. . dp/p dp q where q is the quantity demanded of an asset and p is the price per unit of that asset. hence there is much ambiguity in the connection between the two concepts. Similarly. dx/x dx y d(log x) i. . ∂xi y ∂(log xi ) Applied to the economics of supply and demand the price elasticity of demand (PED) is deﬁned as PED = dq p dq/q = . The PED measures the responsiveness of the quantity demanded to the change in price. due to the surplus that exists. An important point to note at this stage is that elasticity and liquidity are not the same. the ratio of percentage changes. An important concept crucial to the models discussed in this thesis is that of elasticity. Elasticity deﬁnes a relationship between price and the quantity demanded (as deﬁned above).
17 See Protter (2006).CHAPTER 1. including Bank for International Settlements (1999) and Bank for International Settlements (2001). Liquidity risk arises in situations where a party interested in trading an asset cannot do so because she cannot ﬁnd a willing counterparty to that trade. Of course. . implicitly assuming a level of liquidity that is without limits. these standard models assume that the trader will not experience any liquidity risk. In fact one of the most important attributes of ﬁnancial markets is to provide immediate liquidity to investors. More relevant to this thesis. • Credit risk. • Operational risk. INTRODUCTION 34 1. • Model risk. • Liquidity risk. Recent crises in the ﬁnancial markets have triggered studies on the subject of market liquidity.5 Liquidity Risk can be classiﬁed into the following categories17 • Market risk. the stock market crises in October 1987 and 1989. the Asian crisis in 1997 and the problems at LongTerm Capital Management Fund (LTCM) led the Committee on the Global Financial System to conduct several studies discussing the importance of liquid ﬁnancial markets. and the liquidity of a given market varies over time and in addition can dramatically dry up in times of crisis. some markets are more liquid than others. since it aﬀects their ability to trade. The standard models implicitly assume that the only risk experienced by a trader is that due to the uncertain nature of the market. Liquidity risk becomes particularly important to parties who are about to hold or currently hold an asset. For example.
i. In fact.e.e. i. is that the action of a large trade may itself impact the price. immediacy. One particular interpretation of liquidity in the literature ﬁts nicely with the philosophy of this thesis. . Liu (2006) identiﬁes four dimensions to liquidity. this is often termed the liquidation cost. However. even for a highly liquid market. INTRODUCTION 35 1. trading speed. the socalled bidask spread. Many researchers have attempted to do so but the best that can be done is to classify its many dimensions. reﬂecting the eﬃciency of the trading. First. Clearly. trading cost. breadth. orders are both numerous and large in volume with minimal impact on prices. liquidity is a tricky concept to deﬁne (let alone measure). and due to this multidimensional nature comparing individual assets liquidities is also problematic. having low transaction costs. i. Third. Kyle (1985) describes market liquidity in terms of three attributes. of the trade. i. namely.e. the existence of abundant orders both above and below the price at which an asset currently trades. new orders ﬂow quickly to correct order imbalances. independent of all the other factors aﬀecting the price dynamics.CHAPTER 1. trading quantity. clearing and settlement systems. and most relevant to this thesis. Second. this is termed price impact. Alternatively.1 Deﬁning liquidity Market liquidity is often associated with the ability to quickly buy or sell a particular item without causing a signiﬁcant movement in the price. the speed with which orders can be executed. since one asset could be more liquid along one dimension of liquidity while the other is more liquid in a diﬀerent dimension. depth. and ﬁnally resiliency. Howison (2005) states that market liquidity can manifest itself in three possible forms. there is a diﬀerence between the prices for buying and selling the asset. and price impact. namely the tightness. i. trading beyond the quoted depth of the market usually results in a higher purchase price (or a lower selling price) for part. depth and resilience of the market.e. the concept of liquidity is multifaceted and illdeﬁned. due to limited availability of a stock at the quoted price. Sarr and Lybek (2002) state that liquid markets exhibit ﬁve characteristics: tightness.5. the price paid for trading the asset depends on the quantity traded. if not all.e. such as a small bidask spread as well as other implicit costs.
selection bias and simultaneity bias are involved when using this measure. there is no single method for measuring it. in this case the spread is called the realised spread. or daily changes regressed on daily volume. such as measurement error. 18 The ﬁrst is transaction cost measures that capture the costs of trading ﬁnancial assets and trading frictions in secondary markets. Measures which are often used in the empirical literature on liquidity and asset pricing include the bidask spreads. this is primarily used to measure the breadth and depth of the market. and various measures of order ﬂow. However.CHAPTER 1. Sarr and Lybek (2002) classify the existing liquidity measures into four categories.2 Measuring liquidity Because there are many dimensions of liquidity. One particularly intuitive measure of transaction costs is the percentage bidask spread. . deﬁned as BAS = 2 PA − P B PA + P B . or absolute price changes regressed on absolute volume. Of all these measures. the latter being a better estimate of the actual transaction costs since trades may not take place at the actual quoted prices. Trading volume is traditionally used to measure the existence of numerous market participants and transactions and is deﬁned as Vol = i=1 18 n Pi Q i (1. where the ask price PA and bid price PB can be calculated from the quotes on the market or using a weighted average of actual executed trades over a period of time. Measures of the price impact of order ﬂow include price changes regressed on signed volume. Measures of volume include numbers of trades and daily volume measured in dollars. tricky econometric issues. various measures of the price impact of order ﬂow. INTRODUCTION 36 1. despite the advantages of using the price impact of order ﬂow as a measure of liquidity. the price impact of order ﬂow is perhaps the most widely used. the advantage of this measure being that it is based on the actual observed price changes associated with trades.15) See Sarr and Lybek (2002) for a good review of many examples of each class of liquidity measure. In the second category are volumebased measures that attempt to distinguish liquid markets by the volume of transactions compared to the price variability.5.
Most of the existing literature attempting to measure liquidity has focused on the diﬀerent dimensions of liquidity individually. and the most relevant to the focus of this thesis.15). However. More recently Liu (2006) introduced a new measure of liquidity (called the standardised turnoveradjusted number of zero trading volumes over the prior 12 months) that . One such question is whether liquidity is priced in asset returns. clearly no single measure can manage to fully capture the multifaceted nature of liquidity. such as Chalmers and Kadlec (1998). in the main these attempt to measure resiliency of the market. (1998) (who instead use the turnover rate) argue that liquidity is priced. Chen and Kan (1995) and Eleswarapu and Reinganum (1993) suggest that it is not.15). In fact this problem of no universal liquidity measure has resulted in many unanswered questions in market microstructure theory. Pi and Qi are prices and quantities of the ith trade during a speciﬁed period. N is the outstanding stock of the asset and P is the average price of the n trades in (1.CHAPTER 1. There are many other volumebased measures. The resulting turnover rate gives an indication of the number of times the outstanding volume of the asset changes hands. and as such there is no universally accepted measure of liquidity. The fourth and ﬁnal category. This can be given more meaning by relating it to the outstanding volume of the asset. whereas others. which focuses on determining the processes by which information is incorporated into prices. The third category of liquidity measures are equilibrium pricebased measures that try to capture orderly movements towards equilibrium prices. such as general market conditions or arrival of new information. The turnover can thus be deﬁned as TO = Vol NP where Vol is the trading volume deﬁned in (1. these attempt to measure both elements of resiliency and speed of price discovery. For example Amihud and Mendelson (1986) (who simply use the bidask spread) and Datar et al. are marketimpact measures that attempt to diﬀerentiate between price movements due to the degree of liquidity from other factors. INTRODUCTION 37 where Vol is the dollar volume traded.
the new twofactor CAPM model is able to account for the booktomarket eﬀect. liquidity risk adjusted capital asset pricing model (CAPM) that well explains the crosssection of stock returns. Such models are not referred to speciﬁcally in this thesis and so it suﬃces to describe brieﬂy the role of some of the more important market participants. known as specialists. These are individuals or ﬁrms that will take both long and short positions in a given security in order to facilitate trading.CHAPTER 1. (possibly) answering the question whether liquidity is priced. for an overview of this topic see O’Hara (1995). where many models of price formation have been proposed. However knowledge about how these prices are actually formed in the market are of great interest. From a market microstructure perspective. 1. INTRODUCTION 38 aims to capture multiple dimensions of liquidity. The marketmaker accepts a certain level of risk in holding the ﬁnancial instrument or commodity but hopes to be compensated by making a proﬁt on the bidask spread. In addition. Their main function being to provide the other side of trades when there are shortterm buyandsellside imbalances in customers orders. The dynamics by which this information is incorporated into the current price is addressed in the market microstructure literature. In the United States. which the Fama and French (1996) threefactor model fails to explain. Using this measure Liu (2006) outlines a twofactor. since we wish to see exactly whereabouts in the price formation process liquidity issues become important. One of the most important members of any ﬁnancial market are the socalled market makers. price movements are caused primarily through the arrival of information. many markets have oﬃcial market makers for each given security. and thus add to the liquidity and depth of the market.6 Price formation We have alluded to the fact that the price of ﬁnancial instruments may be considered as entirely dependent on supply and demand. .
which instead use an automated system called SETS). however. we are not guaranteed execution but we are guaranteed price. On the London Stock Exchange (LSE) there are oﬃcial market makers for many securities (except for the largest and most heavily traded companies. in these orderdriven markets liquidity now becomes selforganised. and as such gives no guarantee on the price but is guaranteed to be executed immediately. most ﬁnancial markets have become fully electronic and operate on what is called a matched bargain or order driven basis. This may be to exploit arbitrage opportunities such as index arbitrage (the misalignment of the price of an index and the sum of its constituent stocks) or to perform portfolio insurance. the automated execution of a deterministic . In these markets. each stock always has at least two market makers and they are obliged to deal. in the sense that any agent can choose. INTRODUCTION 39 In return. Here. the specialist is granted various informational and trade execution advantages. providing liquidity by posting limit orders or consuming liquidity by issuing a market order. On the LSE one can always buy and sell stock. when a buyer’s bid price meets a seller’s oﬀer price the stock exchange’s matching system will decide that a deal has been executed. The two main type of orders are the market order. It should. In an orderdriven market there are numerous types of orders that can be placed. The introduction of electronic markets has seen a sharp increase in another type of market participant. at any instant of time. Nowadays. which is an order to buy or sell immediately at the best available price.CHAPTER 1. This is in contrast with much smaller order driven markets in which it can be extremely diﬃcult to determine at what price one would be able to buy or sell any of the many illiquid stocks. Alternatively we have limit orders which are not to be executed unless the speciﬁed price is met (or bettered) by current bids or asks. A program trader is one who uses a computer to automate his trades. either to provide or to consume liquidity. the program trader. however. be noted that limit orders often incur higher commission fees. In traditional exchange ﬂoor markets the burden of providing liquidity is given to market makers or specialists. Further. each catering to the diﬀerent needs of diﬀerent market participants.
(1992) argue that one of the causes was actually information about the extent of portfolio insurancemotivated trading suddenly becoming known to the rest of the market. Sircar and Papanicolaou (1998) and Sch¨nbucher and Wilmott (2000) were inspired by the temporary equio librium approach of F¨llmer and Schweizer (1993). 1990) showed how geometric Brownian motion. Moreover Bick (1987. These large orders can contribute to the existing momentum of the market. 19 . They derived equilibrium diﬀusion models for the asset price based on interaction between the two.7 Option pricing in illiquid markets: a literature review Authors such as Kreps (1979) and Bick (1987. This shall be seen in a more mathematical framework in chapter 3.CHAPTER 1. Furthermore F¨llmer and Schweizer (1993) were the ﬁrst to use a microeconomic o approach to construct diﬀusion models for asset price movements. 1. INTRODUCTION 40 hedging strategy. thereby increasing market volatility. 1987 market crash19 and to be responsible for an increased stock market volatility. They deﬁne information traders who believe in a fundamental value of the asset. one of the fundamental assumptions of the BlackScholesMerton model. since they quickly dump large orders on the market at critical times. 1990) have placed the classical BlackScholesMerton formulation into the framework of a consistent model for market equilibrium with interacting agents having very speciﬁc investment characteristics (see section 1. can be derived in a general equilibrium model with pricetaking agents. Consequently. This prompted the realisation that assets had been overvalued because the information content of trades induced by hedging concerns had been misinterpreted. and noise traders whose demands are from hedging requirements.6). Program traders are thought to have been a contributing factor of the October 19. Jacklin et al. Starting from a microeconomic o equilibrium and deriving a diﬀusion model for stock prices which endogenously incorporates the demand due to hedgers and in particular delta hedgers. general price levels fell sharply. Many of the models discussed in this thesis such as Platen and Schweizer (1998).
However in markets that allow large traders to impact the price of the asset there is the possibility of price manipulation and so called market corners and market squeezes. In such models the large trader can move the price by his actions. They concluded . there have been many examples of such activities. Frey and Stremme (1997) studied the perturbation of volatility induced by a delta hedging strategy for a European option whose price is given by a classical BlackScholes formula with constant volatility. it is not clear that an option is still perfectly replicable. and only dependent on the current position of the trades. INTRODUCTION 41 The literature on liquidity falls broadly into two approaches. Moreover. i. An example of a market squeeze is the (alleged) soybean manipulation of 1989 for which more details can be found in Pirrong (2004). especially in (less regulated) developing markets. 1994) provided a discretetime model which allows the large trader to impact the market via some reaction function. He showed that the price of a derivative in this framework must be equal to the hedge cost. The ﬁrst involves the price impact due to a large trade. is dependent on the large trader’s position in the underlying and the derivative asset.e. the trader achieves control by disruption in the supply of the cash commodity. but this cost. 1994) it was showen that to prevent any such manipulation the price impact mechanism must not exhibit any delay. hence it is no longer straightforward how to derive option prices from the prices of the underlying. A market corner is a successful eﬀort of a trader to manipulate the price of a futures contract by gaining eﬀective control over trading in the futures and the supply of the deliverable goods. a detailed and readable account of which can be found in Williams (1995). In addition a suﬃcient condition to exclude proﬁtable market manipulation (in discretetime) was given. that the price mechanism must be independent of the history of the trades. Although price manipulation violates the Commodity Exchange Act. Jarrow (1992. However in the theoretical framework proposed by Jarrow (1992. An example of a market corner is the Hunt silver manipulation of 19791980. leading to nonlinearity. and hence the price. in the presence of price impact.CHAPTER 1. In a market squeeze. Bank and Baum (2004) later extended Jarrow’s results to continuous time.
They did show.20 Frey and Patie (2002) extended the work of Frey (2000) with an asset dependent liquidity parameter which attempts to incorporate so called liquidity drops. whereby market liquidity drops if the stock price drops. this can be characterised more succinctly as the solution of a nonlinear PDE. . the volatility smile. more eﬀectively. Other continuous time models similar to Frey (1998) include Sch¨nbucher and Wilmott o (2000). place a heavy restriction on the amount of market illiquidity that the model allows and rely on the terminal payoﬀ being suﬃciently smooth. and hence there is still risk associated with hedging in illiquid markets. then it is possible to replicate an option perfectly (provided certain conditions on market liquidity and the nonlinearity of the payoﬀ condition are satisﬁed). These results. however. the question as to whether options could be perfectly replicated in a ﬁnitely elastic market reduces to solving (recursively) a ﬁnite number of equations. who used a market microstructure equilibrium model to derive a modiﬁed stochastic process under the inﬂuence of price impact. 2000) then showed that if feedback is taken into account in a more general hedging strategy (which we term full feedback ). In the continuous time framework of Frey (1998). In the discretetime framework of Jarrow (1994). the aim being to reproduce. Platen and Schweizer (1998) proposed a model using an approach that attempted to explain the volatility smile and its skewness endogenously and 20 For further discussion on these restrictions see chapter 4. that increasing heterogeneity of the distribution of hedged contracts reduces both the level and price sensitivity of this unhedged risk. then it is not possible to replicate perfectly an option. The PDEs derived by these latter authors correspond to those derived in chapter 2 of the present study. Sircar and Papanicolaou (1998) derived a slightly diﬀerent nonlinear PDE that depends on the exogenous income process of the reference traders and the relative size of the program traders. Frey (1998. for which Frey (1998) gave existence and uniqueness results.CHAPTER 1. however. INTRODUCTION 42 that if a hedging strategy is used which does not take into account the feedback eﬀect (which we term ﬁrstorder feedback ). both of which can be seen as undesirable restrictions.
amongst others.e. The approach is relevant if an agent wishes to trade a large amount in a short time. Other notable work in the area of liquidity and price impact includes Agliardi and Andergassen (2001) who extended the work of Sch¨nbucher and Wilmott (2000) and o Frey (1998) to include sink transaction costs as an additional source of illiquidity. as such. i. Lyukov (2004) then extended the model of Platen and Schweizer (1998) with more realistic assumptions about market equilibrium conditions (taking into account the presence of a market maker) and also obtained a very similar nonlinear PDE to that derived in chapter 2. These models have been considered by Rogers and Singh (2006) and Cetin and Rogers (2007). without longterm eﬀects.CHAPTER 1. who propose a series of independent auctions. Another model in this category is the work of Cetin et al. The majority of these models will be considered in more detail in chapter 7. which the latter study was trying to avoid. Esser and Moench (2003) extended the work of Frey (2000) to incorporate stochastic liquidity into the price impact framework. Another ‘tweak’ of these models was made by Liu and Yong (2005) who attempted to regularise the PDE close to expiry. INTRODUCTION 43 Mancino and Ogawa (2003) proposed a very similar model in the same vein. Bakstein and Howison (2003) adopted a similar approach to Rogers and Singh (2006) but the former study leads to feedback eﬀects. the actions of the traders do not inﬂuence the underlying stochastic process. Motivated by empirical evidence. (2004). who modelled the liquidation cost as dependent on the quadratic variation of the trading strategy which again leads to a nonlinear PDE when considered in continuous time. they are concerned more with the liquidation cost than permanent price impact. HenryLabord´re (2004) incorporated the feedback model of Sch¨nbucher e o and Wilmott (2000) into the portfolio optimisation of Markowitz (1959) to ﬁnd that . The main diﬀerence with the ﬁrst class of models is that these are now local in time models. More recently. These models eliminate the feedback eﬀects discussed above and. The second approach to liquidity seen in the literature involves the price impact due to the immediacy provisions of market makers. In these models. supply and demand are equalised by the market maker in the shortterm market.
such as that by Bank (2006) who attempted to price options on illiquid underlyings using the utility indiﬀerence of the market maker. The ﬁrst is the socalled free round trip phenomenon. Finally Jonsson and Keppo (2002) derived a somewhat diﬀerent nonlinear PDE. then we must allow the actions of all large agents to aﬀect the price and furthermore the eﬀect of hedging a multitude of options on the underlying. 1994) gave suﬃcient conditions to exclude proﬁtable market manipulation. Clearly this would result in an impossibly cumbersome problem. for a detailed discussion of how things can go wrong in continuous time see Sch¨nbucher o and Wilmott (2000). but only in the discretetime framework. the price impact models of a large trader have numerous shortcomings. The second. Cvitanic and Ma (1996) and Cuoco and Cvitanic (1998) studied a diﬀusion model for price dynamics when the drift and volatility coeﬃcient are functions of the large traders’ trading strategy. using a model with an exogenously deﬁned exponential price eﬀect function. Recall that Jarrow (1992. They showed an increase in BlackScholes implied market volatility between 1% and 7% for values for the fraction of the market subject to portfolio insurance varying between 1% and 20%. According to Rogers and Singh (2006) and Cetin and Rogers (2007). This approach leads to a stochastic optimisation problem and the HamiltonJacobiBellman equation. and by far the most important problem with these models is that if we allow the action of one large agent to aﬀect the price. which is also inherently nonlinear. One of the few attempts to analyse the aforementioned models from an empirical standpoint was carried out recently by Sanfelici (2007) who considered the model of . which can result in the possibility of market manipulation by a large agent. INTRODUCTION 44 portfolio optimisation has the eﬀect of reducing market volatility and thus the price of options in that market. There has also been some recent work using alternative pricing paradigms in place of portfolio replication.CHAPTER 1. Brennan and Schwartz (1989) also analysed the transformation of market volatility under the impact of portfolio insurance and under the assumption of CRRA utility.
Two phases in the history of asymptotics may be identiﬁed. However. Any other relevant literature will be discussed in the body of the thesis where it is appropriate. as x → ∞. Sanfelici concluded that the nonlinear models above can contribute to the explanation of the implied volatility smile but not as well as the other possible explanations. a theory in which the governing equations are highly nonlinear and as such have . including their rich history. both of which we exploit in this thesis.e.8 Introduction to perturbation methods Perturbation methods. due in the most part to the models’ limited capability to reproduce skewed probability distributions. Taylor expansions are the most well known of this class of expansions. i. These ideas have been developed alongside the theory of ﬂuid dynamics (especially. x say. They are used extensively throughout this thesis and so a brief introduction.CHAPTER 1. INTRODUCTION 45 Mancino and Ogawa (2003) (amongst others) and attempted to calibrate the model to market data.e. comparing the results with other popular models. also known as asymptotic methods. i. These techniques are primarily concerned with coordinate expansions. More recent developments in asymptotic theory have been associated with so called parametric expansions. but not exclusively). asymptotic expansions in which the independent variable. such as jumpdiﬀusion or stochastic volatility. the study did ﬁnd that the model of Mancino and Ogawa (2003) was more stable through time and consistent with market data. often called classical asymptotics dates back to the work of Poincar´ (1886) on the farﬁeld bee haviour of linear ordinary diﬀerential equations. The ﬁrst. is outlined below. are a collection of mainly analytical techniques that can be used to solve (or simplify) mathematical problems involving a small or large parameter. 1. Coordinate expansions can also be used to investigate the behaviour of diﬀerential equations near any singular point x0 as x − x0 → 0 or for large values of the independent variable. plays the role of the large or small parameter.
The relative simplicity of the asymptotic solution does. which we now refer to as singular perturbations. however. An early attempt to tackle this problem was due to Prandtl (1904) in his paper on boundarylayer theory. For a recent overview of how these techniques have been exploited in ﬁnance to date see Howison (2005).9 Layout of the thesis The remainder of this thesis is organised as follows. Chapter 3 investigates the socalled ﬁrstorder feedback model (an exceptional case of a linear PDE) and . under certain circumstances. Prandtl’s idea was to subdivide the solution domain into separate regions where diﬀerent asymptotic forms apply. Chapter 2 provides an heuristic derivation of one of the more intuitive models attempting to incorporate liquidity into option pricing theory and in addition shows how the majority of models introduced in the current literature can be formulated in this framework. Nevertheless for many practical purposes the results obtained can be made suﬃciently accurate. During the ﬁrst half of the twentieth century it was demonstrated that a number of problems involving small or large parameters developed a pathological behaviour. Over the next half century these ideas were developed by many (Friedrichs (1954). INTRODUCTION 46 tractability only in the simplest of situations. and indeed can often provide invaluable insight into the qualitative behaviour of the problem under investigation. Kaplun and Lagersrom (1957). Cole (1968) and van Dyke (1964) to name but a few) and so the method of matched asymptotic expansions emerged. Exploiting the smallness of certain parameters and seeking a power series solution in the smallness parameter can often reduce the original system of equations to a much simpler asymptotic set of equations. 1. whose solutions (both analytical or numerical) will often be much easier to ﬁnd. Kaplun (1967). where the power series expansion solution sought no longer becomes an asymptotic series. come at the expense of the approximate nature of the results. This behaviour stemmed from the fact that in certain regions of the solution domain we have a so called asymptotic breakdown.CHAPTER 1.
Chapter 8 considers brieﬂy the socalled stock pinning eﬀect which appears to be well explained by the models outlined in this thesis. Chapter 7 takes a look at the existing models in the literature from a viewpoint of the results found in chapters 26. INTRODUCTION 47 furthermore highlights interesting diﬀerences of both the option value and American option earlyexercise boundary from the classical BlackScholesMerton case. here we concern ourselves mainly with the behaviour of the free boundary which exhibits some interesting qualitative diﬀerences with the standard American option free boundary. Chapter 5 takes a closer look at this model in a regime in which it appears no classical solutions exist. here socalled phase plane analysis is found to be useful in determining such behaviour. In chapter 6 the perpetual options of the full feedback model are considered.CHAPTER 1. Chapter 9 concerns itself with the related topic of the British option. which can provide an investor with protection against the liquidity issues discussed in the rest of the thesis. Finally chapter 10 provides some concluding remarks and ideas for future research. Chapter 4 investigates the fully nonlinear full feedback model using both analytical and numerical techniques. .
This has not previously been done in the literature.1) where S is the price of the underlying. µ and σ are the (constant) drift and volatility respectively and Wt is a standardised Brownian motion. . It is possible to add a forcing 48 . we present the following arguments. We shall assume the underlying process to be a geometric Brownian motion (but this can be generalised to any stochastic process) dS = µSdt + σSdWt . (2. which are similar to those employed in Lipton (2001) and Liu and Yong (2005). In order to provide a derivation of the primary governing PDEs considered in this thesis.Fischer Black (19381995) in 1986 In this chapter we present an heuristic derivation of one particular class of model for incorporating liquidity into option pricing theory.Chapter 2 The Modelling Framework In the end. but because researchers persuade one another that the theory is correct and relevant. a theory is accepted not because it is conﬁrmed by conventional empirical tests. We also attempt to highlight the links between the existing models and furthermore we transpose these models into a single intuitive analytical framework.
Note that at this stage no assumptions need be made regarding the form of the functions λ(S. even at this early stage in the derivation. gives to leading order1 1−λ ∂f ∂S dS = µS + λ ∂f ∂t dt + λ ∂2f (dS)2 + σSdWt . f (S.2) where λ(S.e. t) is a function of S and t only.4) where we have used the condition that (dWt )2 → dt as dt → 0.5) . t) is an arbitrary function. Since f (S.2). t) = ˆ ∂f 1 − λ ∂S µ(S. to the process. We shall return to this issue numerous times throughout the remainder of the thesis.6a) (2. i. (2. as dt → 0 (dS)2 = σ 2 S 2 dt ∂f 1 − λ ∂S 2 + o(dt). .1): dS = µ(S. we begin to observe possible singular behaviour.CHAPTER 2. t). which can be obtained by simply squaring equation (2. t). dS = µSdt + σSdWt + λ(S. t)df. σ (S. (2. . (2. . and with a little rearranging.3). . . ˆ ˆ where ∂2f 1 ∂f 1 µS + λ ˆ + 2 σ2 2 ∂f ∂t ∂S 1 − λ ∂S σS .6b) Note the term on the lefthandside and how. t)dWt .4) into (2.3) In order to proceed further we require an expression for (dS)2 . t) and f (S. analogous to (2. and particular ﬁnancial interpretations can conveniently be postponed until certain manipulations are complete.3) to yield. it is possible to incorporate the additional contribution to the price dynamics into the drift and volatility coeﬃcients µ and σ. 2 ∂S 2 (2. t)dt + σ (S. Substituting (2. We commence by using Itˆ’s formula on the function f (S. we arrive at the following stochastic process. to obtain o df = ∂f 1 ∂2f ∂f dt + dS + (dS)2 + . ∂t ∂S 2 ∂S 2 which substituting into (2. which is dependent on the stock price and time. t) = ˆ 1 (2. t). THE MODELLING FRAMEWORK 49 term.
CHAPTER 2.1) being modiﬁed to the process (2. t) will specify the number of shares needed to be bought or sold at time t and price S due to such a strategy. The ﬁnancial interpretation of this forcing term will be considered in full detail shortly. i. However the volatility of the process also becomes singular at the same location and so it may be possible to move beyond the ‘boundary’. t) as a forcing mechanism on an underlying stochastic process which results in the process (2. In the context of markets with ﬁnite elasticity. The interested reader is referred to Sch¨nbucher and Wilmott (2000) o who provide an analysis of the behaviour of the modiﬁed stochastic process at such singular locations. t) to be the number of extra shares that should be held due to some deterministic hedging/trading strategy and hence df (S. Markets are not complete to traders who do not have the opportunity to trade continuously. the drift of the modiﬁed process µ(S. Also λ(S.5). t) does not appear in the option pricing PDE. consistent with standard BlackScholes arguments.e. when ∂f ∂S = 1/λ. which leads to the following pricing PDE for the modiﬁed stochastic process incorporating the aforementioned forcing term σ2S 2 1 ∂V + ∂t 2 1 − λ ∂f ∂S ∂2V ∂V + rS − rV = 0. 2 2 ∂S ∂S (2. t) can be interpreted as some function dependent on how we choose to model the form of price impact and liquidity.7) Note that.5) will exhibit some boundary behaviour at the location of the singularity in the drift. We now turn our attention to option pricing under the modiﬁed stochastic process. It is tempting to expect that the stochastic process (2. since at this location the drift will become inﬁnite and consequently the process might be contained within a ﬁxed domain.1). Only large institutions can trade close to continuously and so. ˆ Thus far we have been deliberately vague about the ﬁnancial interpretation of the forcing term in (2. To do this we will use the wellknown Generalised BlackScholes equation (for a more detailed derivation see for example Duﬃe. in a complete . we shall return to this issue in section 2. 1996). we can deﬁne f (S. THE MODELLING FRAMEWORK 50 We can interpret the function f (S.3.
This leads to the linear PDE 1 ∂V + ∂t 2 σ2S 2 1− 2 V BS λ ∂ ∂S 2 ∂2V ∂V + rS − rV = 0.7) should be set to an option delta. options provide no extra trading opportunities to them. therefore.8) This leads to an interesting question about which strategy the hedgers are assumed to follow.CHAPTER 2.9) This case we call ﬁrstorder feedback. ∂V 1 σ2S 2 + ∂t 2 1 − λ ∂2V ∂S 2 ∂2V ∂V + rS − rV = 0. i. in equation (2.10) 1 using the simple substitution V (S. the trading strategy adopted has to be found as part of the problem.7). Another (more interesting and challenging) case is when the hedger is assumed to be aware of the feedback eﬀect and so would change the hedging strategy accordingly. Note that although λ can be scaled out of (2.10) which is dealt with in chapter 4. based on some form of option V ∗ . In this case the trading strategy f . which is analysed in detail in chapter 3. and we therefore chose not to do this. t).e.2 f = ∆∗ = ∂V ∗ . This case is considered in the model of stock pinning by Avellaneda and Lipkin (2003) and will be discussed further in chapter 8. options open up new trading possibilities. We shall call this case full feedback. we have assumed that European contingent Note that this is for a net long position in the market. which corresponds to the case when V ∗ ≡ V .e. ∂S (2. resulting in many large institutions selling the options to the small traders and then hedging the risk by replicating the option. similar to Liu and Yong (2005). 2 . i. This results in a net long position (of the large institution) for stocks in the market. t) = λ w(S. A naive strategy would be if V ∗ were the BlackScholes value V BS and thus distinct from the solution V of equation (2. Note too that for simplicity. this would then introduce λ into the standard payoﬀ conditions. 2 2 ∂S ∂S (2. In such a market there is a high demand for these replicating strategies and it is not. This leads to the fully nonlinear PDE. 2 2 ∂S ∂S (2. THE MODELLING FRAMEWORK 51 market. For small traders however. unreasonable to assume that a trading strategy that could impact the price signiﬁcantly is that of delta hedging. if there was a net short position then we would set f = −∆∗ .
Finally. the question as to whether or not the model described above leads to a complete market is of interest. Frey (1998) has the similar form λ(S. as noted by Frey (1998). the precise manner in which this is achieved is discussed in section 7. and it appears to be introduced for numerical expediency to avoid diﬃculties associated with the growing option gamma as expiration approaches. Equation (2. t) is a constant. t) = λS ˆ where λ ∈ R is again some measure of the liquidity of the market. The simplest case occurs in Sch¨nbucher and Wilmott (2000). THE MODELLING FRAMEWORK 52 claims are settled by physical delivery of the underlying asset at maturity.9) and (2. T − t is time to expiry and β a decay ˆ for λ(S. in Almgren and Chriss (2001). for example. analogous with the work of Sch¨nbucher and o Wilmott (2000).CHAPTER 2. hence that there exists a unique solution to equations (2. This is primarily due to the fact that the exact liquidation value is diﬃcult to determine. since it depends on the liquidation strategy chosen by the investor. such market traders who know that some other market participant has to dissolve a large hedged portfolio in the near future.5. optimal liquidation strategies are discussed. t) = λ(1 − e−β(T −t) ) coeﬃcient. In this context the answer to this question reduces to showing whether or not an option can be perfectly hedged in such a market.10) has appeared in the literature several times with diﬀering forms of the function λ(S. As far as we can ascertain there is little ﬁnancial justiﬁcation for this. Indeed. Another form ˆ where λ is a constant price impact coeﬃcient. t) according to the modelling assumptions. although a good deal of the analysis close to expiry presented here is quite widely applicable to other models.10) with the appropriate . In what follows it is assumed for the most part that λ(S. How we choose to close out the contracts is not just an academic exercise since. and this highlights that this factor fundamentally changes the optionprice dynamics. who have λ(S. In this case we do not need to introduce liquidation costs at maturity into the replicating portfolio. can try to proﬁt from this information by front running the anticipated trades. t) constant and a dimensionless o ˆ measure of the liquidity of the market. in relatively illiquid markets. t) can be found in Liu and Yong (2005) in which λ(S.
For example. and so the solution to the PDE no longer corresponds to V (S. Introducing the nonlinearity in the above manner results in a breakdown of the FeynmanKac representation.3) to formulate the solution to a nonlinear PDE as an expectation. Q. 3 (2. It is well understood that the solution of a linear parabolic second order PDE can be expressed as an expectation of a Markov process.t where the expectation is taken under the riskneutral measure. see for example Rosenerans (1972).11) . however. although under some fairly restrictive assumptions. Note that this equation can be obtained from an appropriate transformation of Burgers’ equation.7.3. In the ﬁrstorder feedback case the linear nature of the equation makes it easy to show this. of the modiﬁed stochastic process (2. S. Recall.1 Technical asides In this section we describe some of the more technical details regarding the application of standard analytic methods to the problem as outline above.1 Markovian processes In the full feedback case we eﬀectively break the link between the solution of the PDE and the solution of the SDE. x u(x.5). t) = EQ e−r(T −t) (K − ST )+ . This is not to say that one cannot use the FeynmanKac representation theorem (see section 1. 2. 2. It is possible to do so if the nonlinear PDE can be linearised by an appropriate transformation.1. the nonlinear equation3 ut + auxx + bu2 = 0. For the full feedback problem it is not immediately clear. THE MODELLING FRAMEWORK 53 payoﬀ proﬁle. T ) = h(x).CHAPTER 2. for more information see section 1. that Frey (1998) showed such existence and uniqueness results.
1 in spacetime in two separate regions . t). i. In the above example the functional is given by the ColeHopf transform. Markovian). This full nonlinearity makes it extremely unlikely that such a linearising transform can be found and indeed the author could ﬁnd no such transform. compared to a function which is just dependent on the value of the process at the current time (i. for example the integral process or the maximum process. t) a 54 . THE MODELLING FRAMEWORK b u(x. A functional is any function on the sample path of a process. w(x.e.1 o diﬀerentiable. t) = exp resulting linear system wt + awxx = 0. a can thus be written as an expectation using the standard FeynmanKac representation theorem for Markovian processes. This is a speciﬁc example of the fact that nonlinear equations can be expressed as a functional of a Markov process. and so a link is reestablished between the nonlinear PDE and a Markovian stochastic process. 2adWtP . t) is C 2. however.5 of Protter (1990) or Peskir (2003). that equation (2.2 Applicability of Itˆ’s formula o In the application of Itˆ’s formula it is assumed that the function f (S. The linearising transformation can then be applied to the above expectation to recover the original (nonlinear) value function u(x.t where xt follows the dynamics dxt = √ b h(xT ) a . w(x. The can be linearised using a ColeHopf transformation w(x. For a detailed deﬁnition of local time see section IV. Consider for example a function that is C 2. For more on the ColeHopf transform solution to the nonlinear Burgers’ equation see Rosenerans (1972).1.e.11) is a quasilinear PDE as opposed to equation (2. T ) = exp b h(x) . 2.10) which is fully nonlinear. if it is not then we have to include local time contributions. t) = EP exp x. Note.CHAPTER 2.
Evaluating df of (2. K − ). K + )} .4 The ﬁrst point to note is that this function does not have a welldeﬁned second derivative at S = K and so df is illdeﬁned at this point. t) = dS + 2  (dS)2 . To illustrate this we brieﬂy describe (a slight modiﬁcation of) the socalled ItˆTanaka o formula. K + )} . THE MODELLING FRAMEWORK 55 separated by a kink along a curve in spacetime. S ∈ (K + . S ∈ (K − . sgn(x) = 0.14) 4 This would correspond to a trading strategy in which we would always hold stock and the size of our holding would be equal to the distance the current stock price was from the strike price K. df = sgn(S − K)dS + σ 2 K 2 lim ↓0 1 {dtSt ∈ (K − . x > 0. where we wish to apply Itˆ’s formula to the function f (S. S ∈ (K + . 2 σ 2 S 2 dt (2. x < 0. 5 The sgn function is deﬁned as −1. We can. 1.  dS. K − ). .12) gives −dS. however. x = 0.CHAPTER 2. (S−K)2 f (S. Note that here we have a welldeﬁned second derivative in the entire domain. 2 S ∈ (0. ∞). If we are to apply Itˆ’s formula o across the curve.   . t) = + 2  S − K. S ∈ (0. 2005a).13) Next we can calculate (dS)2 = and thus we obtain 1 − λ sgn(S − K) 2. (S−K) 1 df (S. then we have to take into account the local time spent on that curve (see Peskir. ∞).12) where > 0 is a small parameter. 2 (2. overcome this by approximating the nonsmooth function by a smooth function such as K − S. S ∈ (K − . t) = S −K where o K can be thought of as the strike price. (2. K + ). K + ). Taking the limit ↓ 0 leads to5 df = sgn(S − K)dS + lim ↓0 1 {(dS)2 S ∈ (K − .
t]Su ∈ (K − . In this framework the transaction price St is given by the formula ˆ St (α) = eρα St . Note that for the BlackScholes model the value function of a standard put option can be shown to be C 2.1 diﬀerentiable in the interior of the domain then we will have to apply a local time correction to our application of Itˆ’s formula in a similar manner to the o above. K + )} .13) will become df = sgn(S − K)dS + σ 2 K 2 dLt . 2. in particular the latter two. 2 and hence equation (2.2 Alternative models Bordag and Frey (2007) identify three distinct frameworks that attempt to model illiquid markets: transactioncost models. Since Itˆ’s formula is not being applied across the kink in the payoﬀ proﬁle then Itˆ’s o o formula. without local time.2. (2004).CHAPTER 2. However smoothness of the solution to the full feedback PDE has not been determined a priori and if it transpires that it is not C 2. is all that is required. Hence our application of Itˆ’s o formula is valid for the ﬁrstorder feedback case. THE MODELLING FRAMEWORK 56 We can deﬁne the local time spent on the curve St = K as Lt given by Lt = lim ↓0 1 {u ∈ [0.1 Transactioncost models The main model is this class is due to Cetin et al.1 diﬀerentiable everywhere in the interior of the solution domain with nonsmoothness only on the domain’s boundary (due to the payoﬀ proﬁle). 2. reactionfunction or equilibrium models. A brief overview of each framework is provided below. and reducedform SDE models. The derivation of the governing PDEs provided in this chapter aims to illustrate the links between these frameworks.
The normalised stock demand of the ordinary investors at time t is modelled as a function D(St . (where supply meets demand) D(St . It can be shown that under suitable assumptions on the function D. The normalised stock demand of the large investor is written in the form λΨt . Sircar and Papanicolaou (1998). Frey and Coauthors (19962001). when α > 0) and receive less when selling (α < 0). Frey and Stremme (1997). Furthermore assuming that this strategy Ψt is Markovian. We shall return to this framework after we have considered the somewhat more intuitive ﬁnal framework. where λ ≥ 0 is a parameter that measures the size of the trader’s position relative to the total supply of the stock. i. the above admits a unique solution and hence it can be inverted to solve for St . ρ > 0 is a liquidity parameter and α is the number of shares being traded. namely ordinary investors and a large investor.e. t) + λf (St . Ut . Obviously the trader will pay more than the fundamental price for the stock when buying (i. t) for a smooth function f .2 Reactionfunction (equilibrium) models The models in this class include Jarrow (1994). Platen and Schweizer (1998). Lyukov (2004) and Bank and Baum (2004). THE MODELLING FRAMEWORK 57 ˆ where St is some fundamental stock price (usually given by geometric Brownian ˆ ˆ ˆ motion dSt = µSt dt + σ St dWt ).CHAPTER 2. In such o models there are two types of traders in the market. (2004) show that this transaction cost is proportional to the quadratic variation of the stock trading strategy. then the market clearing condition becomes D(St . The overall supply of the stock is normalised to one. t) + λΨt = 1. that Ψt = f (St . see Bank and Baum (2004). Ut . Cetin et al. . Ut . The equilibrium price St is then determined by the market clearing condition. Sch¨nbucher and Wilmott (2000). Diﬀerent models propose diﬀerent functional forms for D which lead to slightly diﬀerent prices.2.e. t). For a very general analysis of the dynamics of selfﬁnancing strategies in reactionfunction models. where St is the price of the stock. 2. t) = 1.
the transaction cost model described in section 2. t)dt + σ(St . i.15) where as before Ψt is a semimartingale representing the trading strategy. . in turn. Jandaˇka and c ˇ e Sevˇcoviˇ (2005) and Liu and Yong (2005). This framework provides a very intuitive means of obtaining the stock price process in the presence of a large trader. t)dΨt . t)dWt + λ(St .15). Here investors are assumed to be large c traders in the sense that their actions aﬀect the equilibrium price and the liquidity adjusted price process in the presence of a large trader is given directly by dSt = µ(St . Frey and Patie (2002).2.2.e. In the next section we aim to illustrate how all three frameworks may be ‘transposed’ into a reducedform SDE (2.15) and hence all the modelling can be encapsulated into the function λ(S. THE MODELLING FRAMEWORK 58 2. We show that any of the equilibrium models existing in the literature can be recast in terms of a reducedform SDE and so (at least for the purposes of analysis) the reducedform SDE models can be seen as the more general framework. Note that the model outlined in the ﬁrst half of this chapter falls under the reducedform SDE class. since the ﬁnal term in the SDE can be considered as incorporating price impact. that Ψt = f (St . We shall consider. whose properties we aim to analyse in this thesis.3 Reducedform SDE models The models in this class are due to Frey (2000).1 can also be rewritten as a reduced form SDE and we shall discuss this next. t). 2. (2.3 A uniﬁed framework In this section we attempt to unify the equilibrium and reducedform SDE models.CHAPTER 2. In addition. Again we can make the assumption that the strategy is Markovian. t) for a smooth function f . each of the equilibrium models and show that they can be rewritten as an SDE of the form (2.
However.3.2 Platen and Schweizer (1998) The model of Platen and Schweizer (1998) deﬁnes the market clearing condition D(St . Ut . (2004) as ˆ St = eλf (S. t) + O(dt) where we have not included any quadratic variation terms in the calculation as this would aﬀect only the drift of the process and for the purposes of option pricing we are only really interested in the volatility and price impact terms. Hence6 dD(St .CHAPTER 2. t). (2004) Rewriting the model of Cetin et al. 2.3. t) + O(dt) = 0. t) = constant where the demand is modelled as D(St . Ut .1 Cetin et al. The function f (S. t) = dUt + γ 6 dSt + df (St . THE MODELLING FRAMEWORK 59 2. t) to be the trading strategy as before. where S0 and γ constant. St Again. t) in the above model is identiﬁed as the number of shares traded rather than held and so it is not obvious that we should thus identify f (S. in the interest of brevity we will not investigate this model any further. note that we are neglecting the quadratic variation terms since these would only inﬂuence the drift and clutter the algebra. Ut .t) St we can take diﬀerentials to get ˆ ˆ dSt = d eλf St + eλf dSt ˆ ˆ ˆ = λdf eλf St + eλf µSt dt + σ St dWt + O(dt) ˆ = eλf St (λdf + µdt + σdWt ) + O(dt) = St (λdf + µdt + σdWt ) + O(dt) ⇒ dSt = µSt dt + σSt dWt + λSt df (St . t) = Ut + γ (log St − log S0 ) + f (St . .
t) = −νγ −1 . t) . t) = −γ −1 dH . for any function H. ν ∈ R and so we have that dSt = −γ −1 mdt + νdWt + df (St . df . but in this formulation it is clear that it must be negative in order to produce a price process that has a positive drift and volatility. where the demand is modelled as D(St .CHAPTER 2. There is discussion about the sign of the parameter γ in the paper of Platen and Schweizer (1998). t) = −γ −1 . Again dUt = mdt + νdWt so we have that dSt dH = −γ −1 mdt + νdWt + df (St . St hence we can identify σ(St . 2. t) = −νγ −1 . Ut . λ(St . Also the authors assume that dUt = mdt + νdWt for m. t) = Ut + γ (log St − log S0 ) + H f (St . t) + O(dt).3. t) + O(dt). t) = constant. λ(St .3 Mancino and Ogawa (2003) The model of Mancino and Ogawa (2003) extends the work of Platen and Schweizer (1998) and so they also deﬁne the market clearing condition as D(St . Ut . THE MODELLING FRAMEWORK 60 where St denotes the equilibrium price. St df hence we can identify σ(St .
3.2.5 Sircar and Papanicolaou (1998) The model of Sircar and Papanicolaou (1998) deﬁne the (normalised) market clearing condition to be (cf. t) = (LN )−1 .2) D(St . . where the right hand side is the supply from the market maker. ν ∈ R.16) where f (St . t) + ρf (St . t). Ut . section 2. He also deﬁnes liquidity L to be L= and dUt = N mdt + N νdWt .3. (2. where N is the total number of shares and m. THE MODELLING FRAMEWORK 61 2. t) = D(St . t) = 1. t) = νL−1 . Ut . σ(St . Ut ). dMt /N dSt /St 2. St L L LN hence we can identify µ(St . t) = mL−1 . λ(St . Putting these together we have dSt m ν 1 = dt + dWt + df (St .4 Lyukov (2004) The model of Lyukov (2004) deﬁnes the market clearing condition to be dUt + df (St . t) = dMt . The authors also note that if we restrict the demand function to the form D(St .CHAPTER 2. t) is the normalised aggregate demand per security traded and ρ is the ratio of options being hedging to total supply.
St and so we have dUt ρdf dSt =γ + + O(dt).18) + ρdf + O(dt) = 0. hence βUtγ . where dUt = µ1 dt + σ1 dWt . linear. t) + ρdf = 0. t) = St Therefore the market clearing condition becomes7 dD(St .18) that βUtγ = D(St . St Ut βUtγ Now it is clear from (2. StBS hence γ is the ratio of the volatility of the reference process Ut to the volatility of the BlackScholes process StBS . Ut . σ1 (2. Ut .17) where σ0 is the volatility of the BlackScholes process given by dStBS = µ0 dt + σ0 dWt . t) = 1 − ρf. there is no explicit time dependence.e. St (2. Ut ) = φ(Utγ /St ).CHAPTER 2. ⇒ dSt dUt ρSt =γ + df + O(dt). Further the authors assume that φ(z) = βz.e. D(St . i.16) and (2. Ut and γ is the ratio γ= σ0 . then in order for the model to reduce to the BlackScholes model in the limit of ρ → 0 the demand must take the form D(St . St Ut 1 − ρf 7 Ignored the quadratic variation terms. ⇒d ⇒ βUtγ St γ dUt dSt − Ut St βUtγ + ρdf = 0. Ut . . THE MODELLING FRAMEWORK 62 i.
THE MODELLING FRAMEWORK 63 Finally substitution for Ut from (2. has a signiﬁcant eﬀect on the option replication price. St σ1 1 − ρf hence we can identify σ(St . even in this case. t) = σ0 . t) = ρ .17) gives σ0 µ 1 ρ dSt = dt + σ0 dWt + df + O(dt). especially as we approach expiry. 1 − ρf With this survey complete we now proceed in the next chapter to investigate the case of ﬁrstorder feedback (with λ constant for simplicity) and show how the illiquidity.CHAPTER 2. . λ(St .
+ rS 2 2 ∂S ∂S (2.10). but under a diﬀerent guise.Chapter 3 Firstorder Feedback Model As a starting point to investigating how liquidity can aﬀect the option value. highlighting the diﬀerences with the classical BlackScholes model. They call o the solution to the PDE (2. This leads to the linear PDE (assuming λ constant) ∂V 1 + ∂t 2 σ2S 2 V 1 − λ ∂ ∂S 2 2 BS ∂2V ∂V − rV = 0. on a small scale.9) models the replicating 64 . rather than the delta from the modiﬁed option price. This idea of ﬁrstorder feedback leading to a modiﬁed.9) the price taker’s price. a small trader can trade any number of shares.9) which is somewhat easier to solve than the fullfeedback problem PDE (2. We also consider American options in this framework which has not previously been attempted. we assume that a hedger holds the number of stocks dictated by the analytical BlackScholes delta. but still linear PDE also appears in Sch¨nbucher and Wilmott (2000). without aﬀecting the price. in particularly in the region close to expiry. In an illiquid market inﬂuenced by a large trader (or by an equivalent large group of small traders) following the BlackScholes hedging strategy. This chapter investigates the analytical properties of equation (2. but still has important and interesting diﬀerences from the classical BlackScholes PDE. Hence equation (2.9).
1) on page 21. and exercise price. a more detailed analysis (applicable for times close to expiry) follows. The techniques used are more often than not standard and have well understood error estimates. for these traders only. in such situations ﬁnite diﬀerence methods are much easier to implement.9) (obtained using a CrankNicolson procedure1 ) for European call options (all with time to maturity. σ = 0. 1 . who are aware of the large traders inﬂuence on the market. 5. and these too strongly point to a monotonic asymptote on to the payoﬀ function (for ﬁxed T ) as the liquidity parameter λ increases. As λ is increased.CHAPTER 3. zero (some of the analysis below will conﬁrm this). but important diﬀerences. The result with λ = 0 is. Here. 2. of course. the market appears liquid. 10. FIRSTORDER FEEDBACK MODEL 65 cost of an option for such small traders.1). K = 1) for λ = 0. T = 1 year. Corresponding results for put options (using the same parameters as for ﬁgure 3. and this highlights some subtle. Although the illiquid results appear to be rather qualitatively similar to the liquid (λ = 0) results. the classic BlackScholes result.2) are presented in ﬁgure 3. Figure 3.04.e. i. the standard call expiry payoﬀ at t = T has been implemented. Finite diﬀerence methods have the advantage over the alternative ﬁnite element methods because we almost exclusively work in rectangular domains (asset price and time). often with errors within the line width of the graphs presented. Note that ﬁnite diﬀerence methods will be the preferred method of solution throughout this thesis.2. riskfree rate. For more on ﬁnite diﬀerence methods see Smith (1978).1 shows numerical results from the solution of equation (2. the option value is apparently eroded monotonically towards the amount by which the contract is currently in the money or. r = 0.2. if out of the money. It should be noted at this stage that the size of errors and related implementation details will be omitted from the presentation of numerical results. 1. In all calculations the error levels were more than acceptable. with the standard put payoﬀ condition (1. volatility. condition (1.
6 λ=0 λ = 10 PSfrag replacements S 0.8 Put value 0. FIRSTORDER FEEDBACK MODEL 66 0.1 0 0 0. 1. 10. 10. r = 0. the variation with λ appears to be monotonic.7 0.1 0 0 0.8 1 1.3 0.8 1 1.04. σ = 0.5 Call value 0.6 0.6 S 0.9 0. 2.2 0.4 Figure 3.2 0.2. 2.4 0. .2 1.1: Value of European call options with ﬁrstorder feedback (T = 1.2: Value of European put options with ﬁrstorder feedback (T = 1.4 0. r = 0. 5.2. σ = 0.4 0.2 1.2 λ=0 λ = 10 PSfrag replacements 0.4 0.2 0. K = 1) for λ = 0.6 0. 1.5 0.CHAPTER 3. 5. 1 0. K = 1) for λ = 0.04.4 Figure 3.3 0. the variation with λ appears to be monotonic.
5) For a survey of these methods applied to the equations arising in continuum mechanics see Barenblatt (1996). This is generally the most critical and intricate period for option pricing models and oﬀers us some insight into the valuation dynamics.2 It can be shown (see for example Wilmott et al. Naturally the domain and solution outside of this region are given the preﬁx outer.1 Analysis close to expiry: European options In this section we consider the behaviour of the option value close to expiry.9) into Vτ − σ 2 S 2 VSS BS 2 (1 − λVSS ) 2 − rSVS + rV = 0. we wish to zoom in on the solution domain close to strike and expiry.2) We next investigate the small τ behaviour of (3. hence as τ → 0 the solution takes the form V BS = τ 2 f (η) + O(τ ).e. This zoomed domain we refer to as the inner region and the solution in this domain. 2 (3. for which we also need to know the behaviour of the BlackScholes equation (3. 2 (3.4) which is often called the inner variable and whose form can be veriﬁed a posteriori. FIRSTORDER FEEDBACK MODEL 67 3.. The standard substitution. setting τ = T − t (representing time to expiry). (3. 1995) that the solution f when η = O(1).1) (i. where η= S−K τ2 1 1 (3. such a solution is sometimes called a selfsimilar solution. shedding more light on the value of options as the parameter λ is increased. is given by the solution to the ordinary diﬀerential equation (ODE) σ 2 K 2 fηη + ηfη − f = 0.3) . (3. . Mathematically this zooming is done via the following well known transformation. transforms (2.1) where subscripts now denote derivatives and V BS is the solution to the corresponding BlackScholes equation 1 BS BS VτBS − σ 2 S 2 VSS − rSVS + rV BS = 0. for times close to expiry). the inner solution.CHAPTER 3.2) in this limit. To obtain this.
rewriting in terms of the original variables. 1 2 (3.1). where ξ= S−K . Λ 2π 1 η 2 Λ η − 1 + √ e− 2 ( Λ ) Λ 2π (3.CHAPTER 3. First we seek a solution of the form V = τ α g(ξ). To investigate the small τ behaviour of this equation. To illustrate this powerful technique the analysis will be explained in detail. the local BlackScholes gamma is given by BS VSS = σK 2πτ 1 √ e− 2τ σ2 K 2 . consequently. we can perform a local similarity analysis similar to that just performed for the liquid BlackScholes equation. 1995) that equation (3. Λ 2π x −∞ e− 2 y dy. FIRSTORDER FEEDBACK MODEL 68 For a put the appropriate boundary conditions become f → 0 as η → ∞.8) We can now proceed to incorporate this into the ﬁrstorder feedback illiquid problem (3.9) . (S−K)2 (3. f → 0 as η → −∞.7) and.. f → −η as η → −∞ Using these boundary conditions it is straightforward and well known to show (cf. and for a call f → η as η → ∞.6) where Λ = σK and Φ(·) is the standard normal cumulative distribution function deﬁned as 1 Φ(x) = √ 2π Considering the second derivative gives P C fηη = fηη = 1 η 2 1 √ e− 2 ( Λ ) .10) (3. Wilmott et al.5) leads to the solution for a put f P (η) = η Φ and likewise for a call f C (η) = ηΦ 1 η 2 Λ η + √ e− 2 ( Λ ) . τβ (3.
In this (outer) region the second derivative of the solution is eﬀectively zero since the payoﬀ proﬁle has no curvature here.13b) for puts and calls respectively.e.e. To ﬁx α we exploit the fact that the inner solution τ α g(ξ) must match with the outer solution.8) in terms of the new inner variable ξ gives τ α−1 (αg − βξgξ ) − σ 2 K 2 τ α−2β gξξ 2 1− λ √ e σK 2πτ 2 2β−1 − ξ τ2 2 2σ K 2 + rKτ α−β gξ + rτ α g = 0. Consequently the second derivative term is O(τ α−2β+1 ) as τ → 0. Thus we require τ α g(ξ) = S − Ke−rτ for a call option in the limit ξ → ∞. FIRSTORDER FEEDBACK MODEL 69 and is assumed to be O(1).9) without the diﬀusion term. i.11) It is clear that the denominator of the second term is O(τ −1 ) in the limit τ → 0 (provided that 2β − 1 > 0. (3.9) and (3. the solution that is valid outside of our inner region close to strike and expiry. Hence the outer solution is given by the solution to equation (2.1) and BS rewriting the approximation of VSS close to expiry (3. Substituting (3. which can veriﬁed a posteriori ). The ﬁrst point to note is that if the BlackScholes local solution described above is used. . Vτ − rSVS + rV = 0. i. To obtain an appropriate scaling we balance this with the time derivative term which is O(τ α−1 ). to obtain V P (S. τ ) = Ke−rτ − S V C (S. (3. leading to the conclusion that α − 1 = α − 2β + 1 ⇒ β = 1. Therefore it would appear that close to expiry the dynamics of the ﬁrstorder feedback model are signiﬁcantly diﬀerent from the standard BlackScholes model. 1 α = β = 2 . in conjunction with the appropriate initial condition. then the second derivative term in (3. τ ) = S − Ke−rτ + + .10) into (3.CHAPTER 3. α and β are constants to be determined. (3.13a) (3. .1) becomes much smaller than the other terms in the equation in the limit τ → 0.12) This can be solved using the method of characteristics. We shall discuss this matching procedure in more detail shortly.
λ2 (3. (3. that is somewhat smaller than the classical BlackScholes model.3). . τ ) = S − Ke−rτ ξ→∞ K O(τ ) V P (S. We conclude that3 α = β = 1.CHAPTER 3. which is O(τ 2 ) as τ → 0 (see (3. by construction ξ = O(1) in the inner region. (3. τ ) = τ g C (ξ) V P (S.10) suggests we must also have S − K = O(τ β ). S PSfrag replacements V P (S. and taking the O(1) terms leads to πσ 4 K 4 gξξ + (ξ + rK) gξ − g = 0. Note that the scaling here implies a region O(τ ) in asset space S. τ ) = τ g P (ξ) V C (S. close to the strike price. τ ) = 0 ξ → −∞ τ Figure 3. hence (3. we are eﬀectively replacing (3.14) Note that on the relatively short ξ = O(1) scale.4)). this is clearly an important diﬀerence from the standard BlackScholes model behaviour close to expiry. Substituting for α and β into (3. τ ) = 0 V C (S.15) Note that this satisﬁes our a priori assumption that 2β − 1 > 0.3: Asymptotic Matching. FIRSTORDER FEEDBACK MODEL 70 however the preceding arguments suggest that for the matching to work we require S − K = O(τ α ).11) gives g − ξgξ − σ 2 K 2 τ −1 gξξ 2 1− λ √ e− 2σ2 K 2 σK 2πτ ξ2 τ 1 2 − rKgξ + rτ g = 0. In addition. τ ) = Ke−rτ − S V C (S.8) with BS VSS = 3 σK 2πτ 1 √ .
14) can be solved analytically.13). taking on the payoﬀ form away from this point.17) and for a call ξ + rK κ 1 ξ+rK 2 κ + √ e− 2 ( κ ) . g → 0 as ξ → −∞.2. A schematic of this is given in ﬁgure 3. on the inner scale at least. the solution is g P (ξ) = (ξ + rK) Φ where κ = √ 2 2 πσ K . for a put.17) and (3. FIRSTORDER FEEDBACK MODEL 71 The appropriate boundary conditions to equation (3.18) with respect to λ directly gives ∂g C σ 2 K 2 − 1 ( ξ+rK )2 ∂g P = = −√ e 2 κ < 0. This procedure leads to the boundary conditions for a put: g → 0 as ξ → ∞. λ ξ + rK κ 1 ξ+rK 2 κ − 1 + √ e− 2 ( κ ) . Furthermore diﬀerentiating (3. . the option value is monotonic decreasing in the liquidity parameter λ. τ g(ξ) in the limit ξ → ∞ to match with the outer solution (3. τ ) = lim τ g(ξ).18) where Φ(·) is the cumulative normal distribution function deﬁned by (3.18) become increasingly focused about ξ = −rK. van Dyke. i. 2π (3. For asymptotic matching we require the inner solution. S = K(1 − rτ ).17) and (3. 1964). consistent with our observations above regarding ﬁgures 3.14) for standard puts and calls are obtained from an asymptotic matching procedure (cf.e.τ →0 (3. ∂λ ∂λ 2λ2 conﬁrming that. ξ→∞ S→K. Equation (3.7). g → −(ξ + rK) as ξ → −∞. Note that increasing illiquidity (λ → ∞) implies κ → 0 and this in turn indicates that (3.CHAPTER 3.1 and 3.16) Practically we rewrite the outer solution in terms of the inner variable ξ and then equate the result to the inner solution τ g(ξ) to give the appropriate boundary conditions. 2π g C (ξ) = (ξ + rK)Φ (3.3. and for a call: g → ξ + rK as ξ → ∞. and formally the matching condition is deﬁned as lim V OUTER (S.
In the case of the puts. FIRSTORDER FEEDBACK MODEL 72 Figure 3.2 0.15 0. note that the free boundary (optimal exercise price) of the illiquid put option.5 0 0. 3.1 0. 0.04. which opens up the potential for the optimal early exercise (on the ξ scale).05 1 1. the most consistent model has the delta in (2.2.3 0. In the context of ﬁrstorder feedback.05 2 1. obtained via a standard Projected Successive . 0. V − payoﬀ < 0) for g P (ξ) − [−ξ]+ PSfrag replacements PSfrag replacements certain ranges of ξ.2 Analysis close to expiry: American put options The remarks above naturally beg the question as to the value of a put option on a ﬁnitely liquid underlying if early exercise is permitted.4 shows the diﬀerence between the inner solution and the payoﬀ function for both put and call options for various values of the liquidity parameter λ. BS need not necessarily be the same as the free boundary of the liquid option VAM . in the present case it is never optimal to early exercise calls.5 ξ (a) Put 0 λ↓0 0 λ↓0 2 1. Figure 3. . r = 0.5 shows results for the American put with the same ﬁnancial parameters as for the earlier European options. 0. which does permit early exercise.25 0.e.15.1.5 2 Figure 3.4: Inner solution minus the payoﬀ for put and call options.5 1 0.05 g C (ξ) − [ξ]+ 0. 0. it is possible for g P − [−ξ]+ < 0 (i. . .1 0.CHAPTER 3.15 0.5 1 0.e.25 0.8) on page 51 computed BS using the liquid (λ = 0) American put value VAM .. In addition to conﬁrming the monotonic behaviour in λ. g C −[ξ]+ > 0 for all ξ). V .5 1 1.4.3 g C (ξ) − [ξ]+ g P (ξ) − [−ξ]+ 0.05 0 0. it can be seen that call values always lie above the payoﬀ curve (i. As with the classical BlackScholes result for European calls.2 0.2. σ = 0.5 ξ (b) Call 0.5 2 0. and so a consideration of this possibility is considered next. K = 1 and for λ = 0.
CHAPTER 3. FIRSTORDER FEEDBACK MODEL
73
Over Relaxation (PSOR) iterative procedure. The convergence criteria chosen for the algorithm was that maximum error between subsequent iterations was less than 1 × 10−6 . At each node the BlackScholes American value was computed using a PSOR algorithm and then this value was used in the PSOR algorithm for the ﬁrstorder feedback PDE (2.9) subject to (1.2). Again we see the ‘collapse’ of the option value on to the payoﬀ as the liquidity parameter λ is increased (which implies the location of the free boundary always moves towards the exercise price as λ increases). However, although the results appear to be qualitatively similar to the λ = 0 case, there are subtle diﬀerences, as we shall now show.
0.25
0.2
Put value
0.15
λ=0
0.1
PSfrag replacements
0.05
λ = 10
0 0.8 0.9 1 1.1
S
1.2
1.3
1.4
1.5
Figure 3.5: Value of American put options, T = 1, r = 0.04, σ = 0.2, K = 1 and for λ = 0, 1, 2, 5, 10; the variation with λ appears to be monotonic. Analysis of the liquid (λ = 0) American put option close to expiry leads to a somewhat complicated structure, as detailed by Kuske and Keller (1998). Here, the η scale deﬁned in (3.4) can be shown to fail to capture the free (exercise) boundary. Instead, √ the free boundary is located at a somewhat larger distance (O( −τ log τ )) from the exercise price (with a signiﬁcant price variation in a region O( −τ / log τ ) of this exercise boundary). It was shown by Widdicks (2002) that as τ → 0, on the η = O(1) scale, the solution of the liquid (λ = 0) American option takes the same form as that
CHAPTER 3. FIRSTORDER FEEDBACK MODEL
74
of its European counterpart, i.e. (3.6). Therefore, it is entirely self consistent to use this form and, indeed, the European gamma (3.8), for the American case when η = O(1) or smaller, which is relatively distant from the free boundary. However, recall that for the case when λ = 0, with ξ = O(1) we have clear indications of the possibility of early exercise (on the ξ scale) for the illiquid put. In other words the ξ scaling for the ﬁrstorder feedback equation encompasses the free boundary (unlike the η scaling for the liquid (λ = 0) case). Therefore, the American problem in this case reduces to the solution of (3.14), subject to the conditions4 g → 0 as ξ → ∞, g = −ξ and gξ = −1 on ξ = ξf , (3.19a) (3.19b)
where we have used the usual smooth pasting conditions (continuity of the option value and its derivative) and ξf denotes the location of the free boundary (on the ξ scale). It is straightforward to solve the system (3.14), (3.19) fully numerically, however it is also possible to reduce the above problem to a transcendental equation for ξf , a procedure which has not previously been done in the literature and shall be outlined below. It is convenient to make the shift z = ξ + rK which transforms the system (3.14), (3.19) to κ2 gzz + zgz − g = 0, g → 0 as z → ∞, g = −z + rK, gz = −1 on z = zf , (3.20a) (3.20b) (3.20c)
where κ is as before and zf the free boundary on the zscale. To proceed we seek a solution of the form, g(z) = zˆ(z) g
Note that here the outer solution is given by V P (S, τ ) = (K − S)+ rather than V P (S, τ ) = (Ke − S)+ since we are dealing with the American option and so the outer solution will be trivially the payoﬀ proﬁle.
−rτ 4
CHAPTER 3. FIRSTORDER FEEDBACK MODEL
75
which upon substitution leads to 2 z gzz ˆ = − − 2. gz ˆ z κ Integrating once gives gz = ˆ A − 1 ( z )2 e 2 κ , z2
where A is a constant of integration. Integrating once more gives A − 1 ( y )2 e 2 κ dy, y2 z ∞ A ∞ − 1 ( y )2 A 1 y 2 − 2 e 2 κ dy, ⇒ g (∞) − g (z) = − e− 2 ( κ ) ˆ ˆ y κ z z 1 y A 1 z 2 A ∞ − 2 ( κ )2 e ⇒ g (z) = − e− 2 ( κ ) + 2 ˆ dy, z κ z [ˆ]∞ = gz where we have applied the boundary condition (3.20b). Returning to the original function g(z) gives g(z) = zˆ(z) = −Ae− 2 ( κ ) + g
1 z 2
∞
Az κ2
∞ z
e− 2 ( κ ) dy.
1 y
2
(3.21)
It is now required to determine the value of A using the remaining boundary conditions on the free boundary. To do this we diﬀerentiate equation (3.21) to obtain A gz (z) = 2 κ
∞ z
e− 2 ( κ ) dy.
1 y
2
We can now apply the two boundary conditions at zf , i.e. g(zf ) = −zf + rK = −Ae− 2 ( κ ) +
1 zf 2
Azf κ2
2
∞ zf
e− 2 ( κ ) dy,
1 y
2
(3.22) (3.23)
gz (zf ) = −1 = From (3.23) we have
A κ2
∞ zf
e− 2 ( κ ) dy.
1 y
A= hence substituting into (3.22) gives
zf
−κ2 , 2 ∞ −1(y ) 2 κ e dy
−zf + rK = ⇒ rK
∞ zf
y 2 κ
zf
κ2 e − 2 ( κ ) − zf , 2 ∞ −1(y) e 2 κ dy
1 z 2
zf
2
f 1 1 e− 2 ( ) dy = κ2 e− 2 ( κ ) ,
CHAPTER 3. FIRSTORDER FEEDBACK MODEL
76
which can be written in the form of the standard cumulative normal distribution function as rK 2π Transforming back to the original ξ variable yields Φ or further Φ
8
Φ
zf κ
= 1−
κ √
e− 2 ( κ ) .
1
zf
2
ξf + rK κ λ(ξf + rK) √ 2 2 πσ K
=1−
κ −1 √ e 2 rK 2π
ξf +rK κ
2
,
σ 2 K − λ2 =1− √ e 2π 2rλ
ξf +rK σ2 K 2
2
.
(3.24)
6
4
log(−ξf ) PSfrag replacements
2
0
2
4
6 0 1 2
λ
3
4
5
Figure 3.6: Firstorder feedback put (with early exercise), location of free boundary (as τ → 0) with λ, K = 1, r = 0.04, σ = 0.2. Figure 3.6 shows the variation of the local free boundary ξf (more particularly log(−ξf )) with λ for the ﬁnancial parameters considered earlier, i.e. r = 0.04, σ = 0.2, K = 1. The key point to note is that solutions of the system do exist, i.e. the short S − K = O(τ ), ξ = O(1) scale captures the location of the free boundary with ﬁrstorder feedback, whilst as noted above, the liquid (λ = 0) case evolves on a relatively √ longer scale of S − K = O( −τ log τ ); consistent with this as λ → 0, ξf → −∞. Further asymptotic analysis can describe this behaviour, but is omitted in the interests of brevity. Note that transforming back to the original (S, τ ) variables using equation (3.10) we can see that Sf (τ ) = K + ξf τ + . . .
FIRSTORDER FEEDBACK MODEL 77 where ξf is determined by equation (3.CHAPTER 3.9) vanishes is given by the solution to the equation λe− 2 d1 (S . τ ) = log S K + r + 1 σ2 τ 2 √ . This shall be even more important in our consideration of the full feedback problem discussed in chapter 4.3 The vanishing of the denominator One further interesting property of the ﬁrstorder feedback model is that we have the possibility of the denominator of the volatility term vanishing. Using the analytic solution for the BlackScholes equation we ﬁnd that BS VSS e− 2 d1 (S.7 shows the results for the same set of ﬁnancial parameters as used throughout this chapter. Hence the free boundary approaches the strike price at expiry linearly for small times to expiry. 3. It can be seen that equation (3.9) one might naively think that a singularity occurs when 1 − λVSS = 0. . τ0 ) for some ﬁnite timetoexpiry τ0 . determined by the solution of the transcendental equation (r + σ 2 )τ0 = log 5 λ √ σK 2πτ0 .25) This equation can be solved explicitly by setting x = ln(S/K) and doing so we arrive σK 2πτ λ √ 1 2 . σ τ Hence the location where the denominator of equation (2.24). hence a possible breakdown in the diﬀusion term of the ﬁrstorder feedback PDE.5 Considering equaBS tion (2.25) has two distinct solutions for τ ∈ (0. 1− σS ∗ 2πτ at √ 3 S (τ ) = K exp − r + σ 2 τ ± 2σ 2 τ (r + σ 2 )τ − log 2 ∗ 1 ∗ 2 (3.τ ) √ = σS 2πτ 1 2 for both puts and calls where d1 (S.τ ) √ = 0. Figure 3.
2. See.025 0.98 0.005 0. V. FIRSTORDER FEEDBACK MODEL rag replacements 78 1. Hence possible singularities of the solution occur at singularities of the function F (·). . Even though equation (2.03 0.01 S∗ 1 0.9) with λ = 0.04 0.02 1.e. for example.9) in standard form to give VSS = 6 2 BS (Vτ − rSVS + rV ) 1 − λVSS σ2S 2 2 .99 τ0 0. S) . for a second order ODE in the form VSS = F (VS . (1997) . Note that when considering the Frey (2000) model an explicit expression for τ0 can be obtained (see section 7. r = 0.97 0 0. It is a well known property of ODEs6 that singularities of the solution occur only at singularities of the equation when it is written in the standard form V (n) = F V (n−1) .04 and σ = 0. K = 1.02 τ 0.035 0. V. .015 0. it is informative to rewrite (2. . For τ > τ0 the equation has no solutions. and thus the above result cannot be applied directly.01 0.e. i. i.9) vanishes it is now interest to investigate the behaviour of the solution at these points. V (n−2) .1. Also note that the location of the singularity at τ = 0 is K as one might expect.CHAPTER 3. Having determined the location in which the denominator of equation (2.9) is a PDE.7: Location of the vanishing of the denominator of (2. S .1). the denominator does not vanish in this region. Kruskal et al. .045 Figure 3.
0.7.CHAPTER 3. 0). . Hence it is expected that no singularities will appear in the solution for τ > 0.2) (dotted line) and the ﬁrst order feedback PDE (2.9: The second derivative (Γ) of the BlackScholes equation (3. Compare the location of the vanishing denominator 3.85 0. 45 40 35 ∂2V ∂S 2 30 τ = 0.2 ∂V ∂S 0.2) (dotted line) and the ﬁrst order feedback PDE (2.8 and 3.05 1.9 0.05.05.6 τ = 0.015.95 S 1 1.7.. which only occurs at (S.15 Figure 3.05 0..4 ∆= 0.85 0. .95 S 1 1.1 1. Compare the location of the vanishing denominator 3.8: The ﬁrst derivative (∆) of the BlackScholes equation (3.05 5 0 0. .8 τ = 0. .9 0.01. 0 0.1 1. 0. Figures 3. 0.01 25 20 Γ= S∗ 15 10 τ = 0.015. . . instead it is clear that at the locations of the vanishing of the denominator the second derivative will be forced to zero.01 PSfrag replacements 1 0.9) (solid line) for τ = 0. FIRSTORDER FEEDBACK MODEL PSfrag replacements 79 BS from which we can see that the right hand side has no singularities except at VSS  → ∞.15 Figure 3. τ ) = (K.01. 0.05 1.9) (solid line) for τ = 0.9 show the ﬁrst and (more importantly) the second derivative .
9) in the vicinity of S ∗ respectively. It can be seen that the solution at S ∗ appears to be at least C 2. This is true only in certain models and is not necessarily true for certain models when in higher dimensions.e. however this in not the whole story for τ = O(1).1) for two slightly diﬀerent values of λ (the dotted line representing the higher value) and also at three diﬀerent times to maturity. τ ) = ˆ σ2S 2 BS (1 − λVSS ) 2. however as we increase the time to expiry. Figure 3.1 diﬀerentiable and further that VSS remains positive. We can begin to see what may be happening if we take a closer look at the modiﬁed volatility term of the ﬁrstorder feedback case. σ 2 (S. FIRSTORDER FEEDBACK MODEL 80 of the solution of (2. will be convexity preserving. (3.e. τ ). Convexity preservation means that given convexity at t the option price remains convex for all times prior to t. Indeed convexity preservation is a wellknown property of option prices under the assumption of geometric Brownian motion. It can be seen that for the closest time to maturity a monotonic decreasing relationship is shown. just ‘skimming’ zero as S ∗ is approached. Ekstr¨m et al. despite the vanishing of the denominator. i.10 shows the solution to equation (3. a model for the underlying with an arbitrary volatility function σ(S. i. see for example Bergman et al. More importantly they show that (in one dimension) a local volatility model. this relationship appears to be increasing for suﬃciently large times away from maturity. We have shown that for small times to expiry the solution is indeed monotonically decreasing in the parameter λ. The numerical investigations of the ﬁrstorder feedback equation described above indicate that the solution proﬁle does indeed remain convex. o (2005) show that geometric Brownian motion is the only convexity preserving model in higher dimensions.e VSS ≥ 0 for all (S.CHAPTER 3. τ ) as we have here.26) . (1996) or El Karoui et al. i. (1998). One ﬁnal subtlety of the behaviour of the ﬁrstorder feedback equation comes when we take a closer look at the hypothesised monotonic decreasing behaviour of the option value with the liquidity parameter λ. in agreement with the stated results. thus indicating that the solution remains convex as τ increases.
dependent on the size of VSS .1 1. (3. 0.27) indicates that the solution dependency on the liquidity parameter λ is BS monotonic increasing in regions in which VSS < 1/λ and monotonic decreasing for BS BS VSS > 1/λ. τ = 0.1) (for convex payoﬀ proﬁles) remains convex for all (S.0125 0. K = 1 and λ = 0.075. El Karoui et al.09 (solid line) and λ = 0.95 S 1 1.1 τ = 0.06 τ = 0. Compare with ﬁgure 3. τ ).1) does not vanish for τ BS 0.7 hence to determine the option price dependence on λ we need only determine the dependence of the modiﬁed volatility on λ. As such we have the relationship that an increase in volatility will result in an increase in option price.27) BS and it is clear that the sign of the above may change.16 0. 0.12 0. (1996). (1998) or Janson and Tysk (2003) .04. The above agrees well with the results shown in ﬁgure 3. ﬁgure 3.08 0. which it appears we have.26) with respect to λ yields BS σ 2 S 2 VSS ∂σ2 ˆ = BS 3 ∂λ 2 (1 − λVSS ) (3.10. FIRSTORDER FEEDBACK MODEL rag replacements 81 0. Diﬀerentiating (3.02 0 0.14 0. 7 See for example Bergman et al.05 1. (2005) suggest that the solution to the PDE o (3. In fact.0125. Recall that the results of Ekstr¨m et al. σ = 0.075 0.7.85 0. For r = 0.1 (dotted line). However in the limit as λ → ∞ the region VSS > 1/λ expands to ﬁll the whole domain and so this will become the dominant behaviour for a ﬁxed time to expiry.2.0375.04 τ = 0. since for the parameters used.9 0.CHAPTER 3.10: Firstorder feedback put option value for two diﬀerent values of λ at various times to expiry.15 Figure 3.0375 V 0.7 indicates that the denominator of equation (3.04 and so we are in the region in which VSS < 1/λ and we should thus expect monotonic increasing behaviour in λ.
FIRSTORDER FEEDBACK MODEL 82 With this necessary background complete.CHAPTER 3. in the next chapter we may proceed to investigate the more challenging fullfeedback model. .
are not applicable to standard put or call options (nor any nonsmooth payoﬀ proﬁle) and so here we investigate the solutions to the PDE in such regimes.1) where the trading strategy assumed to aﬀect the price is not simply the BlackScholes delta hedging strategy as discussed in the previous chapter.Chapter 4 Fullfeedback Model We now turn our attention to the full feedback case. 2 (1 − λ(S. The aim being to illustrate exactly how things go wrong. In this case the trading strategy has to be determined as part of the problem. This corresponds to a situation where all market participants performing such hedging strategies are aware of the eﬀect that their strategies have on the price. These results. Existence and uniqueness was given by Frey (1998) which shows that options in such a market can be perfectly replicated. namely the equation Vτ − σ 2 S 2 VSS − rSVS + rV = 0. however. 83 . hence the market is complete. t). τ )VSS )2 (4. The full feedback equation has been studied extensively in the literature with various functional forms of the liquidity parameter λ(S. with the view to informing modellers on how to incorporate nonsmooth payoﬀs into the intuitive framework outlined in chapter 2. resulting in nonlinearity. but rather is based on the actual delta of the modiﬁed price. and as a consequence the price impact is fully considered in the trading strategy.
As such the aim of this chapter is investigate thoroughly the properties of the nonlinear PDE (4. As such there is a signiﬁcant diﬀerence in the value of the option to someone holding a long position as opposed to a short position. and due to the negative slope of the demand curve the realised value would be less than the paper value. numerous diﬃculties arise when liquidation strategies are incorporated into such dynamic hedging strategies. for example in many transaction cost models such as Barles and Soner (1998). hence we must take care to specify the option . FULLFEEDBACK MODEL 84 The fullfeedback model described here is what Sch¨nbucher and Wilmott (2000) call o the paper value replication for the large trader.1). The nonlinearity of equation (4.1). VSS )S 2 VSS − rSVS + rV = 0.CHAPTER 4. Equations of this form were ﬁrst studied by Barenblatt and coworkers in the context of hydrodynamics and more recently have arisen in models occurring in quantitative ﬁnance. with the standard put and call payoﬀ proﬁles (1. Liquidating the portfolio would change the price. using various analytical and numerical techniques.1) has the form 1 Vτ − F (S. We also give a brief overview of some of the techniques used to solve general nonlinear PDEs and appraise their appropriateness for equation (4. If we have a nonlinear PDE then one of the most striking diﬀerences with linear equations is that the sum of two or more solutions is no longer necessarily a solution itself. However.1). socalled because the value satisfying the PDE is just the paper value of the option. It is for this reason that the majority of models in the literature (and the present study) consider only the paper value or make the assumption that the option is closed out using physical delivery to bypass any diﬃculties with the liquidation value.1) has many important consequences.2) and (1. τ. 2 and as such is fully nonlinear. Note that equation (4. Nonlinear diﬀusion equations are a frequent occurrence in the physical sciences and the work done in these disciplines can provide much insight into the nonlinear behaviour of the models arising in mathematical ﬁnance.
Many PDEs do not have classical solutions but are nonetheless wellposed if we allow for properly deﬁned generalised or weak solutions. well known to exhibit possible multiple solutions. let us consider any nonlinear transaction cost model.CHAPTER 4. transaction costs will be incurred on both options and such costs will aggregate causing a disparity between the value of the portfolio and the sum of its constituent parts. The problem has a solution. FULLFEEDBACK MODEL 85 position. For linear (diﬀusion) equations. 3. there are many questions that one must ask. The solution depends continuously on the data given in the problem. existence and uniqueness results are well known. This solution is unique. 1 See for example Evans (1998) . and if so is this solution unique? The latter question becomes even more important when dealing with nonlinear equations. a problem is wellposed if:1 1.6. Stating some results from general PDE theory. see section 4. When faced with a partial diﬀerential equation. and in fact the BlackScholes equation can be shown to be a suﬃciently wellposed problem. then we call a solution with this much smoothness a classical solution of the PDE. A portfolio consisting of the same option held both long and short would be priced at zero since this portfolio has a zero payoﬀ and is thus worthless. However when pricing each option separately. to the concept of the bidask spread. It is quite remarkable therefore that the inherent nonlinearity of the pricing equations leads. 2. the diﬀerence being that of the total transaction costs incurred. In addition if we require the solution of a PDE of order k to be at least k times continuously diﬀerentiable. quite naturally. As it turns out this idea of a weak solution is exactly what is needed when tackling the solution to equation (4. The most fundamental being does there exist a solution. The eﬀect of transaction costs and liquidity costs are always a sink of money for hedgers. As an example of this.1) with nonsmooth payoﬀ proﬁles.
Frey (1998) showed that equation (4. τ. 0) = (S − K)+ . τ ). τ ) for which existence and uniqueness was shown using arguments similar to Ladyzenskaja et al. since the second derivative of the put and call option coincide. VSS ).2 In the sequel we will be considering the more general scenario of nonsmooth payoﬀ proﬁles.e. hence we can recover the call option value from the put option value.1) diﬀers from the BlackScholes equation by a function of the second derivative of the option value only. Doing so we obtain VτC with − C σ 2 S 2 VSS 2 (1 − C 2 λVSS ) C − rSVS + rV C = 0. τ ) = λS and for suﬃciently smooth payoﬀ proﬁles. σ = σ(S.e. even in this highly nonlinear situation. V C (S. i. This can be shown by simply substituting the put call parity relationship V P = V C − S + Ke−rτ into the nonlinear equation for the put value V P (S. 2 . V P (S. (1968). An alternative and novel approach to providing existence and uniqueness results. 0) = (K − S)+ .CHAPTER 4. This was done by diﬀerentiating the fully nonlinear PDE in V to obtain a quasilinear PDE in the hedging strategy ∆(S. Note that the nonlinear equation (4. 4. exploiting the maximum principle for parabolic equations.2. FULLFEEDBACK MODEL 86 As previously mentioned. where existence an uniqueness results have not been established. i.1) was wellposed for ˆ λ(S.2. can be found in section A. hence it is clear that the parity relationship will still hold.1 Putcall parity Firstly we note that putcall parity can be shown to still hold. VτP − and the payoﬀ proﬁle P σ 2 S 2 VSS 2 (1 − P 2 λVSS ) P − rSVS + rV P = 0.
and furthermore is only valid for payoﬀ proﬁles depending quadratically on S. 1 . These similarity solutions. It should be . In addition. It should be emphasised here that we have been restricted to a constant λ(S. FULLFEEDBACK MODEL 87 4. if we assume that λ(S.1) gives hτ − σ2h − rh = 0.1) is a function of τ only. where A is a constant determined by the payoﬀ proﬁle.1).1) we can employ the powerful technique of similarity solutions. Note too that regularity issues arise if h = which are not unrelated to the analysis of chapter 5. Such solutions can be obtained by exploiting symmetries of the governing equation (and boundary conditions). 2λ 4. As an example of this. λ = λ(τ ).2 A solution by inspection It should be noted that it is possible to ﬁnd exact analytic solutions to the PDE (4. i.CHAPTER 4. τ ) in equation (4. however the application of such methods tends to be limited by the fact that many nonlinear PDEs do not have such symmetries. Substitution into (4. (1 − 2λh)2 which is now an ODE and so open to standard solution techniques. t) in order to obtain this solution (without any real ﬁnancial justiﬁcation). τ ) = S 2 h(τ ). based on the theory of Lie groups.3 Similarity solutions In order to gain more analytical insight into the behaviour of the highly nonlinear PDE (4. the solutions obtained in this way are generally only valid for very restrictive boundary conditions. Making the further assumption that λ is constant we can arrive at the implicit solution for h: σ2 σ2 ln(h) + ln (1 − 2λh)2 + 2r r σ + √ arctan r √ r (1 − 2λh) σ = (r + σ 2 )τ + A. can be useful in investigating nonlinear problems.e. then we can seek a solution of the following form V (S.
In addition the case k = 0 corresponds to the model developed by Frey and his coworkers (see section 7.CHAPTER 4. (4. for example the initial behaviour of the American option freeboundary problem and the value of an atthemoney option shortly before exercise. FULLFEEDBACK MODEL 88 noted. Under this transformations an equation is invariant and so S √ τ is the only combination of S and τ that is independent of ν and S √ τ hence the solution must be a function of only. because it requires special symmetries in the equation and the initial/boundary conditions. which are diﬃcult to resolve numerically.1) gives the following (highly nonlinear) ODE (a − r)uz + rku − σ 2 uzz + (1 − 2k)uz − k(1 − k)u ˆ 2 1 − λ uzz + (1 − 2k)uz − k(1 − k)u = 0.e. (4. Hence for equation (4. τ ) = S 1−k u(z) where z = log S + aτ.3) 2 If k = 1 then equation (4. In that paper. however.1) can be found in Bordag (2007).4) 2 .1) where λ(S. under the assumption that λ = λ(S).2) into equation (4.3) becomes the much simpler σ 2 uzz − uz (a − r)uz + ru − ˆ 2 1 − λ(uzz − uz ) = 0. Note that k = −1 corresponds to the model of Sch¨nbucher and Wilmott (2000). On the other hand.2) ˆ of invariant solutions is also found.1). it can be extremely useful when considering a local analysis in space or in time. it is proved that λ ∼ S k+1 where k ∈ R is the only case with a nontrivial symmetry group and a complete set there exists a global similarity transform of the form V (S. (4. τ → ν 2 τ for any real number ν. Such a scaling is called a oneparameter group of transformations. that the similarity solution technique is rarely successful in solving a complete boundary value or Cauchy problems. τ ) = λS k+1 . o with constant liquidity parameter λ. The idea behind similarity solutions is to exploit the fact that the equations and the initial (ﬁnal) and boundary conditions are invariant under a certain scaling. i. The global similarities (of Lie type) to equation (4. Substitution of (4. for example S → νS. a = 0.
all the solutions found in this way correspond to diﬀerentiable payoﬀ proﬁles.5) under the assumption of zero interest rates (r = 0). these global solutions may be useful to test the accuracy of numerical techniques applied to such highly nonlinear systems.1).5) by setting a = r. however it does not satisfy any practical boundary and payoﬀ conditions. which diﬀer from the majority of the payoﬀ proﬁles considered in the present thesis (which are more ﬁnancially relevant). τ ) = AS − Be−rτ which is indeed a valid solution of equation (4. For the present model being studied. After transforming back to the ﬁnancial variables this gives V (S. which appears to have no analytic solution and so we are forced to turn to numerical techniques. k = −1. for the k = 1 case see Bordag (2007) and for k = 0 see Bordag and Chmakova (2007) and Bordag and Frey (2007). This leads to the ODE uzz + uz = 0 which has the most general solution u(z) = A − Be−z where A and B are constants to be determined by the boundary and payoﬀ conditions.e.4) and (4. (4. In addition Bordag and her coworkers showed families of explicit solutions to equations (4.3) reduces to (a − r)uz − ru − σ 2 uzz + 3uz + 2u ˆ 2 1 − λ uzz + 3uz + 2u 2 = 0.5) Under the assumption of nonzero interest rates we can quite easily obtain a solution to (4. i. FULLFEEDBACK MODEL 89 and when k = 0 (corresponding to the Frey model) we have the (even simpler) equation (a − r)uz − σ 2 uzz + uz ˆ 2 1 − λ(uzz + uz ) 2 = 0. However. equation (4.CHAPTER 4. We can start to see already that singular behaviour is inherent in this . However.
e. First we rewrite equation (4. we are forced to turn to numerical solutions. τ ) = V0 (S. 2 Now we expand V (S.6) and collecting together terms of the same order in λ gives3 O(λ0 ) : O(λ1 ) : O(λ2 ) : 1 V0τ − σ 2 S 2 V0SS − rSV0S + rV0 = 0. 4.4 Perturbation expansions Since it appears that no analytical solution can be found with the required boundary conditions. τ ) as follows V (S. τ ) + λ2 V2 (S. τ ) are functions to be found. This can be done by exploiting the techniques of asymptotic expansions. i. 2 1 2 2 V1τ − σ S V1SS − rSV1S + rV1 = 2V0SS (V0τ − rSV0S + rV0 ) . However it is possible to ﬁnd an approximate solution for small values of the parameter λ.6. (4. τ ) + λV1 (S. 2 1 2 2 V2τ − σ S V2SS − rSV2S + rV2 = 2V0SS (V1τ − rSV1S + rV1 ) 2 2 + 2V1SS − V0SS (V0τ − rSV0S + rV0 ) .1) in the more convenient form 1 (1 − λVSS )2 (Vτ − rSVS + rV ) − σ 2 S 2 VSS = 0. where Vn (S. We shall return to this in section 4. . ˆ λ 2 ψ u2 + zz σ2 ˆ ˆ − 2λψ + 2λ2 φψ uzz + 2 σ2φ ˆ ˆ + ψ − 2λφψ + λ2 φ2 2 =0 where φ = 3uz + 2u and ψ = (r − a)uz − ru. FULLFEEDBACK MODEL 90 highly nonlinear system by rearranging the above equation into a quadratic equation in uzz . 3 . If we solve this using the quadratic formula we obtain uzz = σ 1 −φ− 1± ˆ ˆ λ 4 λ2 ψ 2 ˆ 8λψ 1− 2 σ 1 2 from which it is clear that diﬃculties will occur if the square root were to become negative. .6) Note that a similar perturbation analysis is outlined in the appendix of Sch¨nbucher and o Wilmott (2000). Substituting this expansion into (4.CHAPTER 4. τ ) + . .
. . this is in agreement with the result shown in ﬁgure 4. V1 ). These results indicate that (similar to ﬁrstorder feedback in the regime VSS  < 1/λ) the inclusion of market illiquidity increases the put option price. V1 .1.e. τ ) to the price of a European put option. Moreover.CHAPTER 4. For more on singular perturbations see Johnson (2004).4 This cannot be guaranteed a priori and the unbounded second derivative of the payoﬀ proﬁle suggests that we may not be able to apply such a regular expansion in the region around any singular points. it can be shown (see appendix A) that if we restrict ourselves to the regime in which VSS  < 1/λ in the entire solution domain (corresponding to suﬃciently smooth payoﬀ proﬁles) then the solution to equation (4. This recursive process is continued until the desired level of accuracy is required (although in practise solving past the second correction term V2 becomes too analytically cumbersome or numerically expensive). i. LBS V2 = f2 (V0 . LBS V0 = 0. LBS V1 = f1 (V0 ). it should be noted that in order to permit a regular asymptotic expansion of the kind outlined above we must assume suﬃcient regularity in the function V (S. Vn−1 ). the lefthandside is merely the BlackScholes operator LBS acting on the nth approximation and the right hand side is a function of the previous approximations (which have been found). However.1) is monotonic increasing in the liquidity parameter λ. . FULLFEEDBACK MODEL 91 This reveals some structure in the successive approximations. . .1 shows the ﬁrstorder correction term V1 (S. . LBS Vn = fn (V0 . Figure 4. τ ). τ ) to be bounded. . 4 See for example Johnson (2004). . speciﬁcally we need the derivatives of V (S.
this is considered below. but of course. FULLFEEDBACK MODEL rag replacements 92 0.5 Numerical solutions Consider next a numerical treatment of (4.25 V1 (S.2. the output was found to be highly dependent on the choice of grid.2.2 1. . which even a cursory inspection of (4.6 0.05 0 0. r = 0.15 0. σ = 0. incorporating iteration in order to treat properly the inherent nonlinearity in the problem. K = 1. . The . λ = 0) European put option for various time to expiry. The ﬁrst is linked to the inevitable inﬁnite behaviour of the gamma with standard payoﬀ conditions. Figure 4. τ ) 0.04.e. even though they were obtained with a relatively ﬁne grid (timestep δτ of 10−3 . T = 1 and τ = 0. gridsize δS of 5 × 10−4 ). Note that this erroneous behaviour is not simply due to the well documented ‘ringing’ behaviour associated with the CrankNicolson ﬁnitediﬀerence scheme (see Duﬀy.4 Figure 4. .1) suggests will be problematic. which are not unconnected. subject to the put payoﬀ condition (1. The results (for the delta) are clearly erroneous. . in addition.1: The leading order correction term V1 (S.1) with λ constant for simplicity.3 0. 0.2). 2004). τ ) to the BlackScholes (i. This sort of diﬃculty is understandably sidestepped in published works (for speciﬁc details see chapter 7) but a study of its causes will surely be helpful for the next phase of modelling in the ﬁeld.2 0.1 τ increasing 0.1.2 shows results obtained using a similar CrankNicolson scheme to that successfully employed on the ﬁrstorder feedback model. 4. 1. In fact there are two problematic issues with regard to these diﬃculties.8 S 1 1.CHAPTER 4.
04.CHAPTER 4. A discussion of this issue. we have a diﬀerent form for the valuation equation in two regions. τ ) = −τ 2 ηH(−η) + τ φ(η) + . close to expiry.1)) is the likelihood of diﬃculties if there is a zero in the denominator of the volatility term. clearly we are allowing for a discontinuous delta close to expiry at the . Thus. Consequently. σ = 0.6 Analysis close to expiry As noted earlier.7) where η is deﬁned in (3. λ = 0.5 PSfrag replacements 1 1.2.2: Deltas for fullfeedback (European) put. H(·) denotes the Heaviside function.1 and T = 1. . a thorough asymptotic analysis of the option valuation close to maturity (τ → 0) can yield signiﬁcant insight into the dynamics of the problem. For this we seek a local solution for the put value of the form (which can be justiﬁed a posteriori ) V (S. K = 1. r = 0.5 1 ∆= ∂V ∂S 0.02 1. FULLFEEDBACK MODEL 93 second diﬃculty (again revealed by a cursory inspection of (4. will be deferred until chapter 5.1 τ =1 0.4).04 Figure 4. 4. 1 (4.98 S 1 1. 1. although the option value is assumed to be continuous. and consequently this limit is studied next. which is necessary to ‘mimic’ the behaviour of the payoﬀ. one in S > K (above the strike) and the other in S < K (below the strike).96 0.5 0.5 0 τ = . . which is associated with smoothed payoﬀ functions.
we sought solutions with continuous deltas. η < 0) the appropriate equation is (4. which is rather broader than the scale appropriate for the ﬁrstorder feedback options (where S − K = O(τ )). η > 0) the following equation describes φ: η σ 2 K 2 φηη φ − φη − = 0. τ ) − (K − S)+ ≡ τ φ(η). It should be noted that the above indicates that the crucial regime is within a distance O(τ 2 ) of the exercise price as τ → 0 (a result determined through asymptotic analysis). φ − φη − 2 2 (1 − λφηη )2 (4.9) in these limits.4) it is evident that V (S. 2 2 (1 − λφηη )2 σ 2 K 2 φηη η + rK = 0. 2 2 + rK [2H(−η) − 1] = 0. FULLFEEDBACK MODEL 94 exercise price. Note also the slower decay to the η → ∞ asymptotes as the volatility increases due to the O(σ 2 K 2 ) scaling that emerges from (4. it enables us (with a little work) to deduce Since from (4.e. similar to the λ = 0 liquid options (as discussed in the previous section).e. without success. and it is our assertion that such solutions do not exist for this problem. smooth pasting (φ.10) which has the useful property of antisymmetry of φ∗ with respect to η = 0 (and so sgn(η) rK 2 as η → ∞). It is helpful to shift φ as follows: φ = φ∗ − which leads to the equation σ 2 K 2 φ∗ η ηη φ∗ − φ∗ − 2 η 2 1 − λφ∗ ηη φ∗ → 5 rK . Indeed. At η = 0. 2 (4.9) with φ → −rK as η → −∞.7) and (3. These results indicate that the option values all lie below the payoﬀ5 (the repercussions of this will be discussed below). . φη continuous) is appropriate.8) 1 with φ → 0 as η → ∞. In the region S < K (i.CHAPTER 4. In addition.3 (obtained via a straightforward RungeKutta fourthorder shooting method). Sample results for a put option are shown in ﬁgure 4.8) and (4. In the region S > K (i.
conﬁrming putcall parity for the nonlinear case. the chosen scaling for the inner region is given by S = K + ητ 2 . FULLFEEDBACK MODEL 95 the results for calls from the results for puts.2.. For the nonlinear problem. Substitution thus gives. Another point to be noted is that ﬁgure 4. provided that early exercise is not permitted. namely (4. V P = −ητ 2 H(−η) + τ φP (η) + o(τ ). . even in this highly nonlinear case (see section 4.11) indicates this is not the case with calls). Note that this symmetry of the local solutions is simply a manifestation of the putcall parity relationship which still holds for all time.3 indicates the possibility of negative put options values. The key observation in the above is the discontinuity in the delta (∆ = ∂V ) ∂S 1 1 1 1 at η = 0 as indicated in (4. (4. after a little rearranging ητ 2 [1 − H(−η) − H(η)] + τ φP (η) − φC (−η) = −rτ K. a somewhat undesirable property (although (4. namely φC (η) = φP (−η) + rK.CHAPTER 4. Note that the scaling for a call was obtained by replacing η by −η in the put scaling. The standard BlackScholes putcall parity is given by V P = V C − S + Ke−r(T −t) . and it is the neglect of this that is undoubtedly responsible for the apparent spurious results observed in ﬁgure 4.1).11).7). we can recover the local solution for calls from that of puts. e−rτ = 1 − rτ + o(τ ). V C = ητ 2 H(η) + τ φC (−η) + o(τ ). which reduces to φC (−η) = φP (η) + rK. which corresponds to the symmetry obtained for the inner equations.11) i.e.
r = 0.15 (taking the other parameters used in ﬁgure 4. 0.9) failed. where we have taken the negative root in order to satisfy the condition that φηη → 0 as η → ∞. For values of σ just below 0.2 η 0 0.01 0.95. Indeed. this is the form that was taken as the basis of the numerical treatment used to treat the results shown in ﬁgure 4.6 0. with the onset of negative roots in the computation.04 0.15.04 and σ = 1.02 σ decreasing 0. 0 0.015 φ 0.3: Local (τ → 0) solution of a fullfeedback put.4 0. ηη where ψ = φ − η φη + rKH(−η).e.2). To understand this.035 0. K = 1.025 0.8) and (4.9) in the form of a quadratic in φηη .. . FULLFEEDBACK MODEL 96 Before a consideration of the problem for calculations for nonsmall values of τ (i. this time in the limit as σ decreases (with other parameters held ﬁxed). σ2K 2 .3. λ = 0.8) and (4. and inspection of the results indicated that diﬃculties arose if 1+ 8λψ < 0. . 2λ2 ψφ2 − 4λψ + σ 2 K 2 φηη + 2ψ = 0.4 0.6 Figure 4. the numerical treatment applied to (4. it turns out that yet another anomaly occurs.CHAPTER 4. Using the quadratic formula we can write the 2 ‘solution’ for φηη as φηη 8λψ 1 σ2K 2 1− 1+ 2 2 = + 2 λ 4λ ψ σ K 1 2 . at times away from expiry). we rewrite (4.2 0.03 PSfrag replacements 0. 0. .005 0.1.
such that V − = V + and V1+ = V1− − 1. ﬁgure 4.CHAPTER 4.10).5 shows the corresponding distributions of the delta ( ∂V ). 0). FULLFEEDBACK MODEL 97 Hence. τ ) and V1 (S. Furthermore. we may expect this regime to arise for large values of the ratio λ/σ 2 K 2 . as evidenced in ﬁgure 4. This latter condition eﬀectively builds the proposed jump in the delta at the strike price into the numerical scheme.4. grid independent). These results also help to justify of the original form of the solution. τ ) = ∂V /∂S.e. . 4.e. We shall return to a consideration of this regime later in section 4. which adds signiﬁcant credence to the integrity of the results. τ ) − max(K − S. In order to incorporate this into our numerics.e. (4. a standard CrankNicolsontype scheme was adopted. an alternative strategy was adopted.7. for suﬃciently large λ.2. as evidenced by 4. based on the Keller (1978) scheme. At S = K two values of the option price and its delta were computed. clearly indicating the jump in its value at S = K. or suﬃciently small σ (since ψ must be an odd function about η = 0.1. and as such can be compared directly with the smalltimetomaturity solutions displayed in ﬁgure 4. The analysis in the previous section points to6 a discontinuity in the delta (∆) at the strike price (S = K) of +1 in the case of a put option. i. This modiﬁed procedure involved writing (4. ∂S The computations shown are highly robust (i.7). which concerns itself with the full problem.full problem We now revisit the choice of parameters employed in ﬁgure 4. showing distributions of V (S.7 Numerical results . i. 6 Since no inner solution with a continuous delta could be found.1) as a system of two ﬁrstorder equations namely in V (S. Calculations performed in this manner provided accurate and highly reliable results. which led to the aforementioned diﬃculties.3. the diﬀerence between the option value and payoﬀ. in particular to the correctness of the jump condition. The grid was then chosen in such a manner that the strike price K coincided with the S grid. namely V − and V1− (for S − = K) and V + and V1+ (for S + = K). In the timewise direction.
This could be avoided in practise by imposing the condition V ≥ 0 which eﬀectively creates another free boundary on the PDE at Sb where the conditions V (Sb (τ ). however. by a simple backward induction argument. at least under the dynamic hedging (replication) pricing paradigm. hence.4) that the model permits negative values for put options. because if the European option value is always below the payoﬀ immediately prior to expiry then. For example. provided it perfectly replicates the option payoﬀ. FULLFEEDBACK MODEL 98 There is. Essentially the hedging strategy should never force the hedger into an irrational position. A corollary to the above remarks is that it can also be seen (from ﬁgure 4. Whilst in certain extreme option valuations.4. It makes little ﬁnancial sense to allow negative option values in any model incorporating market frictions. this could also be regarded as a somewhat undesirable and unrealistic feature of the model. In doing this. in transaction cost models the writer would not rehedge the portfolio (and hence incur extra transaction costs) at times when he does not need to. such as those involving storage costs. the solution to the American put will trivially correspond to the payoﬀ for all time. Note that Bakstein and Howison (2003) make a similar observation and call this condition the ‘American’ constraint. indicates that there is no value in early exercise. this may be acceptable. but are also always above the payoﬀ and. the price being modelled is the cost of replicating the option by trading in the underlying. This has implications for the pricing of American options in this framework. a further issue relating to the results observed in ﬁgure 4. generally this may be regarded as an unwanted facet of the model. i. the corresponding American option will always be exercised immediately when the contract is initiated at t = 0 (or τ = T ). . we have freedom in our hedging strategy. provided the option is still perfectly hedged. This clearly reveals that call values not only remain positive.CHAPTER 4.6 shows results (option value . namely that this indicates the put option value (close to expiry) is always less then the option payoﬀ.e.payoﬀ) for the corresponding call. τ ) ∂S = 0 should be applied. τ ) = 0 and ∂V (Sb (τ ). Figure 4. The same is true for liquidity.
since here even the τ < 0.015 0.04.7. 1 regime is unclear.4: Full feedback put. modiﬁed numerical scheme.96 0. 4.CHAPTER 4.04 Figure 4. K = 1.1 0. 0. r = 0.8 1 0. FULLFEEDBACK MODEL 99 0 τ = . i.035 τ =1 0.2 and λ = 0.02 1.2 and λ = 0.e. when 1 + which turns out to be even more problematic.5 S 1 1. σ = 0.6. It was therefore decided to mount an homotopy type of approach in this regime.2 τ =1 τ = .5 2 Figure 4.1.04.4 0.04 0 0. modiﬁed numerical scheme. speciﬁcally by considering a payoﬀ function of the form V (S.12) .1. 0) = 1 K−S+ 2 (K − S)2 + ρ2 (4.1 0 0.025 0.005 0.1 A second solution regime 8λψ σ2 K 2 Returning now to the other regime outlined in section 4.5: Full feedback put.03 0.6 0.01 V payoﬀ PSfrag replacements 0. σ = 0.2 ∆ PSfrag replacements 0.02 0. K = 1. r = 0.98 S 1 1.
r = 0. Here its use is slightly diﬀerent.13) Ke−rτ − S for S < K.005 τ = . σ = 0. but were tested extensively for numerical grid convergence and found to be numerically consistent on the scale shown. (4.035 0. modiﬁed numerical scheme. These results were based on the method employed for ﬁgure 4.2 and λ = 0. In this way.7. which represented a welldiversiﬁed portfolio containing a multitude of diﬀerent payoﬀs with diﬀerent strikes which combine to produce a suﬃciently smooth payoﬀ to satisfy the smoothness assumptions imposed for existence and uniqueness. .025 0.1) and for three choices of ρ are shown in ﬁgure 4. it is possible to mimic a standard put payoﬀ as the smoothing parameter ρ → 0.01 0. it is merely a mathematical tool to investigate the limit of smoothness.5 2 Figure 4.1 and at a time shortly before expiry (τ = 0.1.03 V payoﬀ PSfrag replacements 0.02 0. In Frey and Stremme (1997) and Frey (1998) this smoothed payoﬀ proﬁle was used to represent an ‘idealised’ option payoﬀ. Results corresponding to the parameter choice of ﬁgure 4.1). τ ) = 0 for S > K. in conjunction with the full problem (4.6: Full feedback call.2.015 0. but instead with σ = 0.1 0 0. FULLFEEDBACK MODEL 100 0.CHAPTER 4. the solution for the put takes the trivial form: V (S. K = 1.5 0 S 1 1.4.04 τ =1 0. These calculations strongly indicate that in the limit as ρ → 0.04.
. K = 1.04.0002 0. Note also that (4. . for all time. Such a breakdown can be seen to be a direct result of the singular nature of the diﬀusion coeﬃcient in the the governing equation.13) indicates that American options in this regime will always be exercised immediately (at t = 0). One interpretation of the above results is that the eﬀect of the nonlinearity.96 0.00025 S 1 1. smoothed payoﬀ. The following chapter investigates another breakdown of the nonlinear PDE. is to suppress the diﬀusion term of the equation in regions of nonsmoothness. Here the diﬀusion term eﬀectively eliminates itself completely and the discontinuity at S = K and τ = 0 cannot propagate away from this point for τ > 0 since there is no diﬀusion.1..CHAPTER 4.1). for standard (nonsmooth) payoﬀ proﬁles. thereby failing to smooth out any discontinuities in the derivative of the payoﬀ proﬁle. for the same reasons expounded earlier for the other regime. σ = 0. which occurs for smoothed payoﬀ proﬁles.1).1 and τ = 0.0003 0. as would normally be the case with the BlackScholes equation.0002 V payoﬀ 1e04 ρ = .001 0 PSfrag replacements 0.0004 0. FULLFEEDBACK MODEL 101 0.98 ρ = . (4.04 Figure 4. 2008).7: Full feedback put.0001 0. λ = 0. Duck et al. solutions which do (trivially) satisfy (4.0003 ρ = . Note that this form of solution indicates discontinuous option values (compare the small volatility analysis of Widdicks et al. 2005. r = 0.0004 0.0005 0.01.02 1.
it can be seen that there is the potential for further diﬃculties to arise due to the vanishing of the denominator in (4. One can consider this as parameterising 1 Note that for simplicity here λ is a constant but can be generalised in what follows. note that for suﬃciently smooth payoﬀ proﬁles. attributed to the discontinuous delta (i. Secondly.1) Firstly. λ (5. note that for standard put and call payoﬀ proﬁles. so far. but in the limit as ρ → 0 recovers the discontinuous payoﬀ proﬁle of a put option. we will once again consider the smoothed payoﬀ proﬁle V (S. inﬁnite gamma) of the payoﬀ proﬁle.e. condition (5.2) where ρ ≥ 0. If we instead assume smoothness in the payoﬀ (i.e. when1 VSS = 1 . since the solution must pass from VSS = ∞ at (K. i. To illustrate the circumstances under which we should expect such singular behaviour. 102 .Chapter 5 Smoothed Payoﬀs . 0) = 1 K−S+ 2 (K − S)2 + ρ2 (5. condition (5.1) must always be satisﬁed somewhere in the domain D ⊆ R+ × [0.Another Breakdown The diﬃculties encountered in the previous chapter have been.e. T ]. ﬁnite gamma). 0) to VSS → 0 as τ → ∞. This function is smooth.1).1) may never be satisﬁed in the solution domain.
(S. the payoﬀ proﬁle. i. SMOOTHED PAYOFFS . i. τ ). 0) = 1 2 1+ ρ2 S)2 (K − S) .τ )∈D or alternatively as placing a restriction on the payoﬀ proﬁle.τ )∈D0 where D0 = R+ × {0}. τ ) (S. sup {VSS } ≤ 1 .ANOTHER BREAKDOWN 103 the smoothness of the payoﬀ proﬁle by ρ. (S. 2 (K − S)2 + ρ2 2 2 2 2 3ρ 4(K − S) − ρ . 0) = 7 2 2 + ρ2 2 (K − S) 2 (K − + 2 3ρ (K − S) (5. we must have λ(S. If we wish to prevent the denominator from vanishing then this can be seen as placing a restriction on the size of the liquidity function λ(S. Furthermore it can be shown. λ(S.3c) (K − S)2 + ρ2 ρ2 3 2 5 . Returning to the smoothed payoﬀ proﬁle (5.e. 0) = VSSS (S. With this in mind the crucial property in determining the existence of singular behaviour will be the maximum of the second derivative of the solution at τ = 0.3d) . VSSSS (S. that the maximum of the second Sderivative of the solution in the entire domain will coincide with the maximum at τ = 0. 0) = − VSS (S. In fact it is intuitively clear from the diﬀusive nature of equation (4.CHAPTER 5.e. i.τ )∈D For the analysis in the remainder of this chapter λ(S. τ ) ≤ sup 1 VSS .e.3a) (5. τ ) will be considered constant for simplicity.3b) (5. (5.1) for increasing τ that this should be so.τ )∈D (S. outlined in appendix A.2) direct computation gives VS (S. . in other words. with the limit ρ → 0 representing highly nonsmooth functions (in the sense of very large second derivatives in the region of the strike) and conversely the limit ρ → ∞ representing increasingly smooth functions (small second derivatives). sup {VSS } = sup {VSS } . via a judicious application of the maximum principle.
For example Sircar and Papanicolaou (1998) set the option value to be the BlackScholes price a small time prior to expiry.4) from which it can be seen that the smoother the payoﬀ proﬁle the more liquidity the model can handle. (5. It should also be mentioned that problems associated with the vanishing of the denominator have been highlighted previously in the literature.see section 7. SMOOTHED PAYOFFS .CHAPTER 5. when we are in the regime that λ > 2ρ. 0) = −3 ≤0 2ρ3 since ρ ≥ 0 by deﬁnition. It should be noted at this stage that the results for existence and uniqueness of a replicating portfolio provided by Frey (1998) only apply when the denominator is not allowed to vanish (here we impose no such restriction). Also a corollary to this result is that if we have a payoﬀ proﬁle with a discontinuous ﬁrst derivative (delta). which yields (obviously) that the maximum occurs at S = K.4) may seem rather restrictive and indeed it is when considering standard put and call payoﬀ proﬁles. which includes the majority of the payoﬀ proﬁles used in practise. where is determined to be suﬃciently large such that the denominator in the diﬀusion term is always positive . Therefore the maximum of the second derivative of the payoﬀ proﬁle is given by sup {VSS } = VSS (K. then to restrict the denominator from vanishing it is required to set λ = 0 and so this model cannot treat nonsmooth payoﬀ proﬁles. .ANOTHER BREAKDOWN 104 The maximum of VSS (S. i. 0) we can see that VSSSS (K.e. 0) = 0. To check that this is indeed a maximum of VSS (S. In what follows we shall assume smooth payoﬀ proﬁles and investigate the nature of the singularities that arise if this restriction is not imposed. Condition (5.3. 2ρ and we can exclude the denominator from vanishing if we place the restriction that λ ≤ 2ρ. 0) = D0 1 . 0) is determined by setting VSSS (S. but that this regime has deliberately been avoided.
CHAPTER 5. Along the same lines Frey and Patie (2002) modify the diﬀusion term of the equation in an ad hoc manner. To solve the full equation (4. Hence for the payoﬀ (5.2) we can calculate the explicit locations of any singularities (denoted S0 ) at τ = 0 by equating the second derivative (5.1) to a simpler ODE valid locally in the region close to the singularity.1) is satisﬁed. Doing so yields S0 = K ± λρ2 2 2 3 − ρ2 .3b) to 1/λ. Here we make no such modiﬁcations and attempt to fully investigate the nature of these singularities.1) near these singular points will be particularly diﬃcult.1) when the singularity condition (5. τ ) = max δ0 . hence ensuring the denominator is nonzero and that the denominator does not become too large to ‘annihilate’ the diﬀusion term.5. see section 7. Note that the solutions at τ = 0 are not themselves singular.2) there are two singularities. Finally Liu and Yong (2005) suggest a form of the liquidity function λ(S. This is done in order to bypass any problems associated with the limits σ → 0 or σ → ∞. each equally spaced either side of the strike K. This analysis is outlined in the next section. Instead a local similarity solution is to be attempted which reduces the PDE (4. λVSS } for suﬃciently large δ0 and suﬃciently small δ1 . ˆ σS 1 − min {δ1 . both analytically and numerically.ANOTHER BREAKDOWN 105 In addition Lipton (2001) states that to avoid any undesirable behaviour of the the option prices we have to limit the magnitude of the local volatility from above and from below. For the smooth payoﬀ proﬁle (5. this has the eﬀect of fundamentally changing the option price dynamics close to expiry. more speciﬁcally they set σ (S. We can expect singular behaviour of (4. although he gives no suggestions about how to do this. . SMOOTHED PAYOFFS . but what remains is to determine if and how singularities propagate through the solution for τ > 0. τ ) that is hoped to suppress such singular behaviour. due to the inherent singular behaviour.
0)+. the solution in the vicinity of the singular point S0 must be analytic (at least for τ = 0) and so can be expressed in the form of a Taylor series about S = S0 . . given the form of the payoﬀ proﬁle.e. 0).6) we therefore seek a similarity solution in the vicinity of the singularity of the form V (S. 0) + τ β V (η). i. . .ANOTHER BREAKDOWN 106 5.1). both of which are assumed to be known.e.2 A local expansion for small τ is sought about the point where the equation becomes singular.1) gives τ β−1 ˆ ˆ β V − αη Vη − 2 ˆ σ 2 S0 λ−1 + τ β−2α Vηη ˆ2 2λ2 τ 2β−4α Vηη − rS0 V1 + τ αη ˆ + τ β−α Vηη λ = 0. 0)+ (S − S0 )3 (S − S0 )2 VSS (S0 .8) +r V0 + τ α ηV1 + 2 τ η ˆ + τ βV 2λ 2α 2 Note that for the case of the put or call payoﬀ proﬁle (i.CHAPTER 5. τ ) = V (S.e. i. section 3. τ ) = V0 + τ α ηV1 + τ 2α η2 ˆ + τ β V (η) + . where VSS = 1/λ. 0) and V1 = VS (S0 . 2 6 (5. τα (5. which we have determined to have an isolated singularity in the payoﬀ proﬁle at S = S0 . isolated in the sense that there are no other singularities in the vicinity of S0 .7) into (4.1 Local analysis about the singularities We consider a general form of payoﬀ proﬁle. at the strike price.5) and the small τ expansion (5. 0) = V (S0 . SMOOTHED PAYOFFS . to remain close to the singular point at τ = 0 we also require that VSS (S0 . λ Next we seek a similarity solution for small τ of the form ˆ V (S.6) where α and β are to be determined from the appropriate balancing of terms and asymptotic matching (cf. (5. . 0)+ VSSS (S0 .5) In addition. Combining the Taylor series about S0 (5. ρ = 0) then the two singularities will coincide. 0) = 1 . 0)+(S−S0 )VS (S0 . and so can not be thought of as isolated anymore and the following analysis will not be appropriate. 2λ (5. .7) where V0 = V (S0 . V (S. with η= S − S0 . . Since we are assuming a smooth payoﬀ proﬁle. Direct substitution of (5. .
6 τ 3α η 3 βˆ V3 + . .1) and evaluating in the limit τ → 0 yields 2 5σ 2 S0 ˆ ˆ = 0. τ V (η) = 6 η→∞ ⇒ lim η→∞ ˆ ˆ For a nontrivial inner solution V (η) we require that V = O(1) as τ → 0. 0). To be consistent.7) in the limit as η → ∞ (i. as τ → 0) must match with the Taylor series expansion of the solution at τ = 0. (5. . . 2λ which after substitution into (4. the form of the solution in the vicinity of the singularity (5. τ ) = V0 + τ 5 ηV1 + τ 5 1 2 η2 3 ˆ + τ 5 V (η) + . 5 as η → ∞.e. .10) Note that the above matching procedure has also provided us with the appropriate boundary conditions (for large η) of the inner solution. 2λ 6 where we have deﬁned Vi = Since the ﬁrst three terms on both sides are identical (by construction) this reduces to lim ˆ τ β V (η) + . 2λ ∂iV (S0 . . (5. 5 3 β= . .12) . . 3 V − η Vη − ˆ2 2λ3 Vηη (5.ANOTHER BREAKDOWN 107 To balance the diﬀusion term and the time derivative term in (5.9) To ﬁx the values of α and β. . .8) it is clear that we require β − 1 = 4α − 2β ⇒ 3β − 4α = 1. this forces us to set β = 3α. i. .e. (5. . . Hence lim V0 + τ α ηV1 + τ 2α η2 ˆ + τ β V (η) + . .10) into (5. that η 3 V3 ˆ V (η) → 6 Substituting (5.CHAPTER 5. .9) we ﬁnd that 1 α= . the matching of the inner and outer solution is used.5). .e. (5. ∂S i η→∞ = V0 +(S−S0 )V1 + (S − S0 )2 (S − S0 )3 + V3 +. i.11) hence the appropriate form of the solution to try around the singularity is given by V (S. = (S − S0 )3 V3 + . SMOOTHED PAYOFFS .
CHAPTER 5. 2γ(γ − 1)A γ−3 η +.11) is not a solution of the inner equation (5.12) gives (3 − γ)Aη γ − 2 5σ 2 S0 2λ3 ηV3 + γ(γ − 1)Aη γ−2 2 = 0. Note that the second term is a higher order correction as η → ∞ iﬀ γ < 3. which must contain higher order matching terms.. rather it is only the leading order term in the asymptotic matching procedure. Substitution into (5. 2λ3 V32 It is clear that we have an inﬁnite asymptotic series solution as η → ∞ with each subsequent term corresponding to a higherorder term in the Taylor series expansion .1. We can determine the next order correction by seeking a solution to the inner equation (5.12) of the algebraic form 3 ˆ (η) = η V3 + Aη γ .13) where A and γ are constants to be found. V 6 (5. We can see that (5. V3 = 0. SMOOTHED PAYOFFS .11) that were derived from an asymptotic matching procedure.12).12) given the appropriate boundary conditions (5. Hence in the limit η → ∞ we must have γ = −2. 5. 2 η therefore we have that the constant A must be given by A= 2 σ 2 S0 .ANOTHER BREAKDOWN 108 Before attempting to solve (5.11). which leads to the equation 5A − 2 5σ 2 S0 2λ3 V32 1 = 0(η −5 ).1 Asymptotic matching Let us look again more closely at the boundary conditions (5. Since we are interested in the solution as η → ∞ we can approximate the denominator as follows (3 − γ)Aη γ − ⇒ (3 − γ)Aη γ − 2 5σ 2 S0 2λ3 η 2 V32 2 5σ 2 S0 2 2λ3 η 2 V32 1− 1+ γ(γ−1)A γ−3 η V3 = 0.. we ﬁrst look a little closer at the matching procedure performed above and the relevant boundary conditions that arise.
11) is modiﬁed to 3 2 2 ˆ (η) = η V3 + σ S0 + O(η −7 ) as η → ∞. these further terms can be determined by performing the same procedure adopted above.15) This shows that the matching procedure results in integer powers of τ .14) The signiﬁcance of the extra term can be seen if we transform the large η behaviour (5. τ ) =V0 + (S − S0 )V1 + + V3 1 (S − S0 )2 + (S − S0 )3 2λ 6 2 τ σ 2 S0 + O τ 2 (S − S0 )−7 . Returning to the smoothed payoﬀ for a moment we can see with a little work that V3 = 3ρ2 2 2 λρ2 5 3 λρ2 2 2 3 − ρ2 . To summarise. which shows that V3 can be either positive or negative depending on which of the two singularities we are seeking a local expansion near. Doing so we have that 7 σ2S 2τ 5 V3 (S − S0 )3 ˆ + O τ 5 (S − S0 )−7 .7) in terms of the outer variable is given by V (S. V 6 2λ3 V32 η 2 (5.14) back to the outer variables.14) to the inner equation (5. 2λ3 V32 (S − S0 )2 (5. SMOOTHED PAYOFFS . the matching condition (5. suggesting strongly that the scaling used is the correct scaling for this problem.CHAPTER 5. which is the focus of the next section. One ﬁnal point to note is that the sign of V3 may aﬀect the qualitative behaviour of the solution. V3 is identiﬁed as the third derivative of the payoﬀ proﬁle evaluated at the location of the singularity. It appears that the inner solution is a power series in η starting from η 3 and decreasing by powers of ﬁve.ANOTHER BREAKDOWN 109 of the outer solution. + 3 2 0 V (η) = 3 2 2λ V3 (S − S0 ) 6τ 5 2 and the form of the similarity solution (5. Now that we have the corrected boundary conditions (5. .12) we can attempt to solve this nonlinear system.
leads to χ = 4/3 and to the exact solution ˆ V (η) = 243κ 80 1 3 1 η = 4 3 3 2 5 3 (σS0 ) 3 4 η3. SMOOTHED PAYOFFS .1. The remainder of this chapter . i.16) subject to the boundary conditions (5. Unfortunately. Therefore in order to obtain a solution with the required boundary conditions we must turn to numerical techniques. If we are a little more sophisticated we could perform the transformation 3 ˆ ˆ V = κ5 V .CHAPTER 5.14). such as shooting or ﬁnite diﬀerence methods.14) may not exist. It is also noted that an exact solution of the ODE (5. however this scaling would then place the constant κ into the boundary conditions (5.16) exists.2 Properties of the inner solution For simplicity we shall rewrite the inner equation (5. numerical solutions of (5. η = κ 5 η.e. λ 2 (5. η 3 leadingorder behaviour as η → ∞.ANOTHER BREAKDOWN 110 5. Trying a solution of the form ˆ V (η) = Bη χ where B and χ are constants.16) The ﬁrst point to note is that κ can be scaled out of the problem 1 ˆ ˆ by setting V = κ 3 V .17) However this clearly does not satisfy the matching condition (5. however it would not be removed completely from the boundary condition and consequently we shall not make any such scalings. The previous statement suggests that a solution to equation (5. which would remove the constant from the equation and also from the leading term of the boundary condition. shooting methods ﬂoundered and ﬁnitediﬀerence schemes failed to converge.12) as ˆ2 ˆ ˆ Vηη 3V − η Vη = κ where κ = 2 5σ 2 S0 . 2λ3 (5.16) with the boundary conditions above proved fruitless.14).
v). An investigation of the phase portrait of equation (5. v). vx = g(u. x ux = f (u. Eliminating the x variable in (5.16) using phaseplane analysis can provide us with such a proof and also give us invaluable insight into the qualitative behaviour of the solutions to (5.ANOTHER BREAKDOWN 111 attempts to prove that this is.18a) (5. the case. SMOOTHED PAYOFFS . v = g(u.18b) The emphasis of phase portraits is on the general qualitative properties of diﬀerential equations and their solutions rather than ﬁnding a closed form solution.16).3 Introduction to phaseplane analysis A useful tool for the study of diﬀerential equations. This analysis will be outlined in the following subsection. is the socalled phase portrait.18) gives the following dv g(u. in fact.CHAPTER 5. v) = . v). ux) can be expressed as two coupled ﬁrst order ODEs by deﬁning the new variable v := ux . (5. v) . du f (u. Below we shall attempt to provide a brief overview of the main properties of such phase portraits. especially if they are in two dimensions.1. hence More generally we can have ux = v. Any general second order (autonomous) ODE of the form uxx = g(u. This indeed becomes extremely useful when such systems have no such closed form solutions. For a more detailed introduction see for example Jordan and Smith (1999) or Hirsch and Smale (1974) and the references therein. 5.
as it transpires. If we assume a linear system of the form ux = au + bv. v) plane.CHAPTER 5.19) and. One might assume that a zero in the denominator. In addition. v) or g(u. This derivative represents the ﬁeld direction in the phase portrait. The only problem that can arise is if we ever have f (u. v) = 0.u. dv g(u. then it remains at that point for all x since here ux = uxx = 0. vx = cu + dv. if they do then this would contradict uniqueness. v c d v x or more concisely ux = A. v) are singular. It can be shown that the nature (and stability) of the ﬁxed point is determined by the solution of the linear system (5. v) = 0 are identiﬁed as ﬁxed points (also called equilibrium or stationary points) of the system. u) plane. since at these points the equation becomes singular and nothing can be said about the direction of the ﬁeld at these points. SMOOTHED PAYOFFS . the determinant of the coeﬃcient matrix . however in this case we could simply shift the axis and consider du f (u. which corresponds to a vertical slope in the (u. which includes the ﬁxed points.ANOTHER BREAKDOWN 112 where the variables u and v provide the axes for the phase portrait. so called because if a solution starts at (or reaches) a ﬁxed point. v) = 0 alone would cause problems. v) = g(u.19) then the only possible ﬁxed point of such systems are at u = v = 0. f (u. i. (5. which can be better represented in matrix form as u a b u = . phase paths cannot normally intersect. v) = g(u. In fact the only place where phase paths can intersect is when either f (u. v) which would give zero gradient in the (v.e. v) = . The points where f (u.
A saddle point is stable along one direction (corresponding to the eigenvector of the negative eigenvalue) and unstable along another (corresponding to the positive eigenvalue) Spirals If β1.2 < 0 we have an stable node (sink) since in the limit x → ∞ the solution will tend to the ﬁxed point and if α1.2 = 0 but α1.2 and β1.2 + iβ1. then matters are nearly as simple as in the linear case outlined above. real eigenvalues and α1 and α2 have opposite sign then such a situation corresponds to a saddle point. α1 = α2 = α and β1 = −β2 = β.2 = α1.19) with the eigenvalues of A given by ω1. Nodes If β1. if we have a linear homogeneous system of the form (5.2 = 0 then the solutions become periodic but the trajectories are closed.e. 3 v(x) = v eα2 x (cos β x + i sin β x). however. 0 2 2 See for example Peixoto (1997). When considering a nonlinear system.CHAPTER 5. in this case the ﬁxed point is called a centre. due to the periodic functions (sin and cos) in the solution.2 > 0 then we have a unstable node (source) since the solution will move away from the ﬁxed point as x → ∞. SMOOTHED PAYOFFS . Three qualitatively diﬀerent classes of ﬁxed points can be identiﬁed.2 . then the general solution is given by u(x) = u0 eα1 x (cos β1 x + i sin β1 x). Furthermore if α1.2 .2 have the same sign then we have a nodal ﬁxed point. socalled because. since there can be multiple ﬁxed points which may interact.e. namely nodes. Saddle Points If again β1. In this case we have a spiral ﬁxed point. However it can be shown3 that provided the system is structurally stable.2 = 0. As an example. i. the solutions will spiral into or out of these ﬁxed points. In addition if β1.2 = 0 we have a complex conjugate pair of eigenvalues. Furthermore if α < 0 then we have a stable spiral (sink) and if α > 0 at unstable spiral (source). For a nonlinear system. which are entirely determined by the values of α1. the eigenvalues are real and if both α1. i.ANOTHER BREAKDOWN 113 A. . spirals and saddle points.e.2 = 0. the structure of the phase portrait is not obvious. i.
the combination of which entirely determines the qualitative behaviour of the ODE and its solutions.16) yields (for both transformations) (Vxx − Vx )2 (3V − Vx ) = κe4x .e. Consider the nonlinear system (5. where the number of solutions increases or decrease. however. SMOOTHED PAYOFFS . there is the additional possibility of socalled limit cycles in the phase portrait. . Moreover each singularity (ﬁxed point) is of the same elementary type as for linear systems (e.g. nodes.16) can be made autonomous. For nonlinear systems. However it turns out that we can produce an autonomous system by making an appropriate transformation.ANOTHER BREAKDOWN 114 multiple singularities can exist (but only a ﬁnite number).20) Immediately we can see that the choice of the positive sign corresponds to a mapping from x ∈ (−∞.16) does not exhibit such nonlinear behaviour and so this will not be discussed further.CHAPTER 5. Firstly we make the variable transformation η = ±ex . spirals and saddles) and in this case linear stability implies nonlinear stability (provided none of the eigenvalues have zero real parts.20) into (5. Hence we are eﬀectively making two separate transformations on two diﬀerent Riemann surfaces. the solution could undergo bifurcations at critical values of the parameter. a centre).1.4 Deriving an autonomous system Equation (5. Furthermore. Substituting (5. ∞) and the negative sign corresponds to the negative semiinﬁnite plane in η. 5.16). ∞) to η ∈ (0. i. where the nonlinear system is dependent on some parameter. the observant reader may have noticed that this equation is not of the autonomous type (since the independent variable η appear explicitly in the equation) and so will not have a two dimensional phase portrait. (5. These are ﬁxed orbits that attract (or repel) nearby paths and correspond to ﬁxed oscillatory solutions of the ODE. however. the nonlinear system (5. Fortunately. which will be outlined in the following subsection.
However before proceeding we note that a further simpliﬁcation can be made . this second order ODE can be expressed as a system of coupled ﬁrst order equations.CHAPTER 5. dv 4u 5 1 =− − ± du 9v 3 v κ . Setting v = ux we arrive at the coupled system v u = 4 −9u − 5v ± v 3 x κ 5 u−v 3 from which we can eliminate the independent variable x to produce the equation for the ﬁeld lines of the phase portrait.e. The nature and stability of the ﬁxed point can be determined by investigating the linearised system in the vicinity of the ﬁxed point. the behaviour of this nonlinear system is entirely determined by the location and behaviour of its ﬁxed points. where the ﬁeld direction cannot be determined. where x = ln η and u = e− 3 V . from which it is clear that the equation will become autonomous if we set µ = 4 .ANOTHER BREAKDOWN 115 and further making the transformation V = eµx u(x). 5 4 uxx = − u − ux ± 9 3 5 u 3 κ . 3 resulting in 4 5 u + ux + uxx 9 3 4x 2 5 u − ux 3 = κ.e. for further details see for example Peixoto (1997).22). the points where ux = uxx = 0. (5. − ux As outlined in the previous section. SMOOTHED PAYOFFS . i. −v (5. gives (after substitution) µ(µ − 1)u + (2µ − 1)ux + uxx 2 (3 − µ)u − ux e3µx = κe4x .21) 5 u 3 Consistent with standard phaseplane theory. i. i. The next section outlines the linearisation procedure and the classiﬁcation of the ﬁxed points in more detail.e.22) . This can be rearranged to make uxx the argument. These points correspond to the singular points of (5.
80 Note that if we were to take the negative squareroot of the equation then the ﬁxed point equation becomes 4u 3 − = 0. Recall that the phase plane’s structure is entirely determined by the location and nature of its ﬁxed points. v) = v.ANOTHER BREAKDOWN 116 to the system (5. v0 ) = 0 and so it can be seen from (5. u= 243 80 1 3 e± 2iπ 3 .CHAPTER 5. v )space will be independent of the u ˆ parameter κ. ˆ (5. Hence the positive root seems the correct choice.23b) then it can be shown that the equation in (ˆ.5 Behaviour of the ﬁxed points We wish to ﬁnd and classify the ﬁxed points of the system ux = f (u. The ﬁxed points are deﬁned by f (u0 . If we seek a solution of the form u = κα1 u. ˆ v = κ α2 v . Now that we have our autonomous system in the simplest possible form we can begin to investigate the structure of the phase plane.24) that the ﬁxed points correspond to v0 = 0 with u0 the solution of the following equation4 u3 = 0 4 243 .21) and we recover the correctly scaled solution via the transformation (5. . SMOOTHED PAYOFFS . the parameter κ can be scaled out of the equation entirely by another appropriate transformation.24) where we have dropped the hats for simplicity of notation.23).e.23a) (5. 3 Hence for all intents and purposes we can set κ = 1 in the original system (5. −v (5. if we choose 1 α1 = α 2 = .21).21). v0 ) = g(u0 . − 9 5u which only has two solutions. 5.1. hence the aim of the next section is to ﬁnd and classify the ﬁxed points of the nonlinear system (5. i. both of which are complex. 5 4 vx = g(u. v) = − u − v + 9 3 1 5 u 3 .
25b) (5. v0 ) + v (u0 .17). v0 ) ∂v (u0 . and this can be performed by linearising the nonlinear equation around these points. . v) = g(u0 + u. ¯ ∂u ∂v ∂g ∂g g(u. v) and g(u. v0 ) + u (u0 . (5. the real part of the ﬁxed points above in (u. ¯ (5. v0 ) v0 v ¯ v ¯ g(u0 . v0 ) + v ∂v (u0 .ANOTHER BREAKDOWN 117 Obviously this has three roots. 1 3 (5.26a) (5. we need to undertake analysis in the local neighbourhood of the ﬁxed point. In order to determine the nature of these ﬁxed points. namely (5. v0 ) ∂u x x The ﬁrst term in the above equation disappears as it is a derivative of a constant (ﬁxed point). v0 ) + u ¯ ¯ ¯ ∂f ∂f (u0 . ¯ ¯ ¯ ¯ ∂u ∂v Hence the system becomes ∂f ∂f ¯ u + u ¯ f (u0 . 0 = ∂g ∂g ¯ g(u0 .14). v0 ) + u ∂u (u0 . v0 ) + o( ) ¯ v0 + v ¯ x ∂f (u0 . v0 ) + o( ). v0 ) + v (u0 . v) about the ﬁxed points (u0 . + ∂u ∂g ∂g (u0 . v0 + v ) = f (u0 . v0 ) + v ∂v (u0 .26b) where is a small parameter (corresponding to a small perturbation) and performing a Taylor series expansion on the functions f (u. v0 ) ∂f (u0 . SMOOTHED PAYOFFS . by setting u = u0 + u. however recall that this solution does not satisfy the required boundary conditions (5.12) in (V. They are u01 = u02 = u03 = 243 80 243 80 243 80 1 3 . In addition the other O(1) term also disappears as it is the functions . v0 + v ) = g(u0 .25c) e− 2iπ 3 Interestingly. v0 ) + u ∂u (u0 . v0 ) + o( ).25a) 2iπ 3 e 1 3 . η)space.CHAPTER 5. v0 ) + o( ) ¯ . v0 ) u0 u ¯ u ¯ f (u0 . ¯ v = v0 + v . v) = f (u0 + u. one real and a complex conjugate pair. v0 ) ∂v + = + o( ). v0 ) which gives f (u. v)space corresponds to the exact solution of (5.
v0 ) ∂v v ¯ . 30 (5. ∂v ∂g 4 5 =− − ∂u 9 6 ∂g 5 1 =− + ∂v 3 2 5 u−v 3 5 u−v 3 3 −2 . SMOOTHED PAYOFFS . the long term behaviour of the linearised system near a ﬁxed point can diﬀer qualitatively from the long term behaviour near a ﬁxed point of the fully nonlinear system. the system becomes u ¯ u ¯ 0 1 . This leads to the following linear system for (¯. which are determined by the solution to the following characteristic equation 0−ω 1 = 0. ∂u ∂f = 1. v ).ANOTHER BREAKDOWN 118 f and g evaluated at the ﬁxed points which by deﬁnition is equal to zero. Considering ﬁrst u01 . The linearisation performed here does not always work however. = v ¯ v ¯ − 2 − 23 3 15 x The behaviour of this linear system is determined by its eigenvalues. In all other cases the local picture of the nonlinear system near a ﬁxed point looks like its linearisation. 3 −2 We now evaluate these derivatives at each ﬁxed point in turn. One is when the ﬁxed point of the linearised system is a centre and the other when the linearised system has zero as an eigenvalue.CHAPTER 5. v0 ) v ¯ ∂u x ∂f (u0 . u ¯ ∂f (u .28) . Fortunately however there are only two situations where this can occur. v ) u ¯ = ∂u 0 0 ∂g (u0 . at this point.2 = √ 1 −23 ± i 71 . For the nonlinear system (5. v0 ) u ¯ ∂v ∂g (u0 . .24) we have ∂f = 0. det 23 2 − 3 − 15 − ω which has two (complex) solutions ω1.
in the complex domain) to fully determine their behaviour.24) the two complex ﬁxed points can be omitted from our analysis. (cf. Recall that two negative real eigenvalues correspond to a stable node. To conclude we have determined that the (real) ﬁxed point is a (stable) spiral sink.1. assume that we 5 Copyright John C. .29b) It also turns out that the ﬁnal ﬁxed point u03 has the same linearised system as above and so the same eigenvalues.rice. and so will be unable to satisfy the boundary conditions as η → ∞ and η → −∞. usually called a spiral sink. negative eigenvalues Next consider the ﬁxed point u02 .edu/∼dfield/.24) (with κ scaled out of the problem completely).e.14) is that it too must pass through this ﬁxed point. 3 ω1 = − (5.CHAPTER 5. Dormand and Prince. What this implies for the solution of (5.12) subject to the boundary condition (5. Further. 15 5 ω2 = − . For more information see http://math.ANOTHER BREAKDOWN 119 Systems with complex eigenvalues correspond to spiral node ﬁxed points and since the real part of the eigenvalue is negative (see section 5. However since these ﬁxed points are in the complex domain we are strictly required to perform a full (fourdimensional) complex stability analysis about these ﬁxed points (i. which has two real.29a) (5. Figure 5. The results were obtained using MATLAB and the pplane ODE software package5 which employed the DormandPrince modiﬁcation to the standard RungeKutta shooting technique. SMOOTHED PAYOFFS . 1980). Polking.3) we have a stable ﬁxed point. at this point the local linear system corresponds to 0 1 u ¯ u ¯ = 9 −2 −5 v ¯ v ¯ 9 x 2 . every path (each corresponding to a diﬀerent boundary condition) passes through this ﬁxed point. The important point to note is that in the absence of any other ﬁxed points.1 shows the phase portrait of the the autonomous system (5. However since we are only really interested in real solutions to the nonlinear system (5.
cf.12) exists (and hence no smooth inner solution about the point S0 ) satisfying the boundary condition (5.e. 4 1 5.17)). we must ﬁrst check that all ﬁxed points of the system have been found and indeed that every path must pass through the ﬁxed point we have previously found. equation (5. 3 have obtained the solution for η ∈ (−∞. 0) and ‘jumped’ to the current phase plane.ANOTHER BREAKDOWN rag replacements 120 8 Complex Region 6 4 2 v 0 2 4 6 8 0 2 4 u 6 8 10 Figure 5. Solving from η = 0 corresponds to shooting in the current phase plane (Riemann surface) from x = −∞. Before we can conclude that no smooth solution to the ODE (5. Note the ﬁxed point at u = 243 3 . SMOOTHED PAYOFFS . (5. since the solution at the ﬁxed point is ﬁxed as η → ∞. This path (like all paths) must pass through the ﬁxed point and by deﬁnition it must remain there for all x as x → ∞.14) as η → ∞.24). The following subsection investigates the structure of the phase plane in yet more detail.1.24) only has real solutions for 5u − v > 0. v = 0 and the ﬁeld direction lines.14). The dotted line represents an analytic 80 envelope for the phase portrait close to the singular line v = 5u . However.6 Structure of the phase portrait Firstly it is clear that the equation (5. we have no hope of satisfying the boundary condition (5.31). 3 . corresponding to η → ∞.1: Phase portrait of the autonomous system (5. and in fact takes the form V ∼ η 3 (i.CHAPTER 5.
CHAPTER 5. However the extra term in the matching condition (5.14) for large η has the form V = A0 η 3 + A1 +. To see this recall that the boundary behaviour (5.14) means that the boundary condition is to be applied slightly below this singular line. but we can make an approximation for large x that x= 3 ln 5 u A0 .ANOTHER BREAKDOWN 5u 3 121 hence the line v = is a limiting point for imaginary solutions and is in fact a singular line.30) 5x 10x 3 .30) gives v= Recalling that A0 = V3 6 5u 10A1 A2 0 − . which after substituting into (5. We ﬁrst transform variables to give u = A 0 e 3 + A 1 e− and so v = ux = 5A0 5x 10A1 − 10x e3 + e 3 . We wish to determine whereabouts on the phase plane this condition corresponds to. correspond to the line v = − 4u on 3 the phase plane.. 3 54u2 (5. The singular line thus corresponds to where we must apply the boundary conditions of the ODE (5. η2 4 where A0 and A1 are known constants. Similarly solutions of the form V ∼ η 3 correspond to the line v = 0 on the phase plane and solutions of the form V ∼ const. 3 3u2 yields 5u 1 − . . and so there is a chance that shooting methods will work here. which corresponds to solutions of the form V ∼ η 3 in the original variables.12) and so it should be clear that we have no chance of applying the boundary condition on this singular line. Clearly this cannot be expressed explicitly in terms of just u and v.31) and A1 = 1 2 5V3 v= Hence this is just below the singular line.. SMOOTHED PAYOFFS . 3 3 (5.
for example v ∼ (u − u1 ) 2 . 2 which also gives the value of C2 to be 8 C 2 = ± − u1 + 9 12 5u1 1 2 (u − u1 )α2 provided 3 . there may still be ﬁxed points which we have .32) The constant C2 found in equation (5. SMOOTHED PAYOFFS . This gives 4 2 α2 C2 (u − u1 )2α2 −1 ≈ − u1 + 9 and so we must have 1 2α2 − 1 = 0 ⇒ α2 = . 5u1 . here we have that u1 α2 > 0. However. We are interested in the limit u → u1 . (5.1.7 Other ﬁxed points Recall that in phase plane analysis. 5.32) only has real solutions provided u1 < 243 80 1 3 = u0 . once we have determined the location and behaviour of all the ﬁxed points of the system we have entirely determined the qualitative behaviour of the solution. If we are interested in calculating the behaviour to the right of the ﬁxed point then a solution of the form v = C2 (u1 − u)α2 should be used. It can be shown that this is indeed the correct behaviour by seeking a solution of the form v = C2 (u − u1 )α2 to give 4 5 2 α2 C2 (u − u1 )2α2 −1 = − u − C2 (u − u1 )α2 + 9 3 1 5 u 3 1 − C2 (u − u1 )α2 . the location of the ﬁxed point.1 that the ﬁeld lines appear to be approaching v = 0 in a square root fashion.ANOTHER BREAKDOWN 122 It can also be seen from ﬁgure 5. where u1 is the v = 0 intercept.CHAPTER 5.
v = r sin θ. − sin2 θ.24) into the above yields 5 5 rx = r cos θ sin θ − r sin2 θ + 1 9 3 r2 5 4 θx = − cos2 θ − cos θ sin θ + 3 9 3 r2 sin θ 5 3 cos θ − sin θ cos θ cos θ − sin θ 1 2 . (5.34a) (5. and diﬀerentiating gives ρx = −ρ2 rx . φ)space 5 5 ρx = f (ρ. cos φ + sin φ Note that this corresponds to the transformation z = z −1 in complex space z = reiθ . r φ = −θ. To investigate any ﬁxed points at inﬁnity it is convenient to transform the problem into plane polar coordinates. and diﬀerentiation gives uux + vvx . namely those at u = v = ∞. SMOOTHED PAYOFFS .34b) 5 3 + sin2 φ.33b) 5 3 1 2 Next in order to investigate the behaviour at inﬁnity we make the transformation6 1 ρ= .CHAPTER 5. φ) = ρ cos φ sin φ + ρ sin2 φ + 9 3 4 5 φx = g(ρ. This is done via the transformation u = r cos θ. (5. r uvx − vux θx = .ANOTHER BREAKDOWN 123 still not considered. φx = −θx . This leads to the following system in (ρ. φ) = cos2 φ − cos φ sin φ − 9 3 6 ρ 2 sin φ 5 3 5 cos φ + sin φ ρ 2 cos φ 3 1 2 1 2 . ˆ .33a) (5. r2 rx = Substitution of equation (5.
CHAPTER 5. φ). 3 ∂φ 9 3 2 5 cos φ + sin φ 2 3 . 9 3 (5. Again to investigate the nature of the ﬁxed point it is required to perform a linearisation about the ﬁxed points. v) coordinates it can be shown that these two ﬁxed points corresponds to 4 1 the point at inﬁnity along the lines u = − 3 v and u = − 3 v. however we are only interested in the principle branch when 0 ≤ φ ≤ 2π. The nature of the ﬁxed points in this space will remain unchanged under the transformation back to the original coordinate system. For simplicity we shall remain in the transformed space (ρ. All that remains is to determine the nature of these ﬁxed points.35) can be factorised as follows 4 cos φ − sin φ 3 1 cos φ − sin φ 3 4 3 = 0. 1 3 Hence this has solutions when either tan φ = or tan φ = which have inﬁnitely many solutions. 3 5 ρ 2 1 + sin2 φ + 3 cos φ sin φ 10 5 ∂g 2 2 = cos φ sin φ − cos φ − sin φ + .ANOTHER BREAKDOWN 124 Clearly the only ﬁxed point of this system is when ρ = 0 and when φ satisﬁes the following equation φx = 5 4 cos2 φ − cos φ sin φ + sin2 φ = 0. Evaluating the partial derivatives of our system yields ∂f 5 5 = cos φ sin φ + sin2 φ + ∂ρ 9 3 2 5ρ 2 sin φ 5 3 3 cos φ + sin φ 5 1 2 . φ0 ) φ ∂ρ 0 x ∂f (ρ . φ0 ) ρ ¯ ∂φ 0 ∂g (ρ . ρ 2 1 + sin2 φ + 5 cos φ sin φ ∂f 10ρ 5ρ 3 = cos φ sin φ + cos2 φ − sin2 φ − . Transforming these ﬁxed points back to the original (u. SMOOTHED PAYOFFS . Doing so leads to the following linear system in (¯. φ ) ρ ¯ = ∂ρ 0 0 ∂g ¯ (ρ .35) Equation (5. 3 5 ∂φ 3 9 2 3 cos φ + sin φ 2 ∂g =− ∂ρ 2 3ρ 2 cos φ 5 3 1 cos φ + sin φ 1 2 . and so we have only two solutions. φ) ρ ¯ ∂f (ρ . φ0 ) ∂φ 0 ¯ φ .
These eigenvalues are both real and positive which corresponds to a nodal source ﬁxed point. In other words any perturbation in the ρdirection will result in the solutions being pushed away from this ﬁxed point.9273). corresponding to a saddle node.3218) leads to the 3 following linearised system 1 ρ ¯ 0 ρ ¯ = 3 ¯ ¯ φ 0 −1 φ x which has eigenvalues 1 ω1 = . SMOOTHED PAYOFFS . 0. Finally considering the ﬁxed point at (ρ. which is unstable. φ) = (0. 3 ω2 = −1.37a) (5. 3 ω2 = 1.37b) Here we have real eigenvalues. which has a stable direction (here corresponding to the φdirection) and an unstable direction (the ρdirection). 0. Here the 3 linearised system becomes 4 ρ ¯ 0 ρ ¯ = 3 ¯ ¯ 0 1 φ φ x which.ANOTHER BREAKDOWN 125 First we will consider the ﬁxed point at (ρ.36b) with eigenvectors along the ρdirection and the φdirection.CHAPTER 5. One ﬁnal point of interest is that the analysis of the ﬁxed points at inﬁnity has revealed a path which does not terminate at the ﬁxed point near the origin. This path however is not realistic in the context of numerical solutions of the original ODE. We can move from the nodal source at inﬁnity along the φ direction and arrive at the saddle node (also at inﬁnity) provided there is no movement in the radial direction.36a) (5. (5. since the matrix is diagonal clearly has eigenvalues 4 ω1 = . arctan 1 ) ≈ (0. (5. as . arctan 4 ) ≈ (0. but of opposite sign. φ) = (0.
ANOTHER BREAKDOWN 126 any solution method would introduce small numerical perturbations and the solution would always terminate at the spiral sink near the origin. to prevent such behaviour. therefore. However. namely ρ itself.12). despite applying a smoothed payoﬀ proﬁle. η3 4 in the original variables. . insofar as in this limit yet further asymptotic analysis is applicable. however. the two values of S0 will coincide and in this limit the problems associated with the vanishing of the denominator are in some ways mediated by the problems of the inﬁnite gamma. indicating (another) failure in the underlying modelling.7 are in the range ρ < λ . 2 as discussed in the present chapter. in guiding the asymptotics described earlier. suggesting that. Hence any solution of a numerical shooting method given any condition at any boundary will asymptote to the ﬁxed point solution u = which corresponds to the solution V = 243κ 80 1 3 1 3 243 80 .7 is still useful. The corollary to this is. via phase plane analysis. there is a further subtlety as ρ → 0 (which we do not explore). corresponding to solutions of the second order ODE with any given boundary condition (at a point) will always terminate at the ﬁxed node near the origin. even using smallscale analysis. it should be pointed out that strictly the parameter values taken in ﬁgure 4. Hence the ﬁxed point in u corresponds to the exact solution of the original equation. Figure 4. involving another small parameter. In this section we have used phaseplane analysis to show that it does not appear possible to resolve the singular behaviour of equation (5. nonsmoothness has been ‘induced’ into the solution for τ > 0. for standard option payoﬀ proﬁles. Finally. that there is insuﬃcient ﬁnancial modelling in (4. SMOOTHED PAYOFFS .1). Hence we have shown. Note too that as ρ → 0. that every numerically simulated path in the phase space of the nonlinear system under consideration.CHAPTER 5.
The next set of results (obtained using a straightforward PSOR scheme). shown in ﬁgure 6. r = 0. or rather on time left to maturity. VP DE . 127 .04.4)). is removed. σ = 0. This chapter investigates such perpetual options in the context of the nonlinear models described in chapter 2. To be consistent with the ﬁnal payoﬀ conditions. This set of results (for an Americanstyle put option) corresponds to the payoﬀ condition with ρ = 0.2).25). in this parameter regime the denominator does not vanish since λ < 2ρ (cf. correspond to a calculation obtained taking the smoothed payoﬀ condition (5. The situation is quite diﬀerent. however.Chapter 6 Perpetual Options Explicit solutions to parabolic freeboundary problems are rare. Although section 4. For these options the dependence on time. the question that naturally arises (given the results of the previous chapter) is what of other payoﬀ conditions. in particular those which do not have discontinuous deltas (and assuming the difﬁculties raised in chapter 5 can be bypassed).2.1.7 has demonstrated that the fullfeedback model with early exercise leads to what amounts to a trivial problem for puts. if we consider perpetual American options. the earlyexercise condition was imposed by taking V = max 1 K −S+ 2 (K − S)2 + ρ2 . (5. so the partial diﬀerential equation is reduced to an ordinary diﬀerential equation.15 (together with K = 1. λ = 0.
which will certainly be the case for standard payoﬀ functions on account of the discontinuous deltas.1). PERPETUAL OPTIONS 128 at all S and τ at each iteration of the PSOR algorithm. Figure 6. .2 1 0. unabated.4 0.2. and evaluate Vτ at ﬁnal maturity (directly from the PDE) then this will provide us with an indication of whether there will exist any earlyexercise regions or not. The computation was permitted to continue until a near steady state had been attained (i. Note that we are in the regime λ < 2ρ and so we should expect no singular behaviour.2) to perpetual options that should be noted and that we shall attempt to outline below.6 0. more speciﬁcally if Vτ ever changes sign. . K = 1. indicating that there is an earlyexercise region.CHAPTER 6. σ = 0.1).1 clearly indicates that the computation could be extended.1: Full feedback American put. However for ρ large enough then this is not the case and Vτ always remains negative and thus the option will always be exercised immediately . for long maturities. λ = 0. r = 0. For a smoothed put numerical investigation shows that Vτ at expiry can be both positive and negative. the asymptote to a perpetual valuation). This does emphasise. If we apply the smoothed payoﬀ proﬁle to the standard BlackScholes equation. ρ = 0. . of course.e. 10.5 1 τ = 10 S 1.15 (smoothed payoﬀ).2).04.5 2 2.5 3 0 Figure 6. (3. that much of the diﬃculty reported above with standard payoﬀ conditions is associated with the vanishing of the denominator in (4. 1. There is a subtlety with the application of the smoothed payoﬀ proﬁle (5.25.8 V PSfrag replacements 0. where VP DE is the solution to (4. τ = 0. . 1.2 τ =0 0 0.
Given that longterm solutions to (4. ρ > ρmax . we have demonstrated in chapter 4. that in chapter 5 it was shown that we should expect nonsmooth behaviour for the European option in this regime and hence it is likely that the American counterpart will exhibit similar solution diﬃculties. that American options are always early exercised. i. Numerical investigations reveal that the same behaviour is also seen for the nonlinear equation (4. since the (CrankNicolson) ﬁnite diﬀerence scheme successfully employed to the system in regime III proved unsuccessful in regime II. 2 (1 − λVSS )2 V = K − S.1) subject to (the standard earlyexercise put conditions) V → 0 as S → ∞. For regime I. 0 < ρ ≤ III. 1 2 2 VSS σ S + rSVS − rV = 0.e. This system was solved using a straightforward RungeKutta algorithm. which performed an iteration procedure to . λ . using a local expansion about τ = 0. II. i. kinked payoﬀ proﬁles.1) where we can now identify four regimes I. ρ = 0. 2 λ < ρ < ρmax . However. the behaviour of the solution to the American option problem in regime II still remains unclear. PERPETUAL OPTIONS 129 (at t = 0). Regime IV is of little interest (since we would never optimally exercise the option) and also ﬁgure 6.CHAPTER 6. dV dS (6. Hence for the BlackScholes equation the earlyexercise boundary exists only for 0 ≤ ρ < ρmax . it is of some interest to investigate the behaviour of this system with the time variation omitted.1 indicates that in regime III there exists a wellposed American option problem. We shall not investigate this regime further but recall however. 2 IV. and = −1 on the free boundary S = Sf .1) (with early exercise) can exist under certain parameter regimes.e.
2.2: Perpetual fullfeedback American put.5 0. As a ﬁnal cautionary note on the numerical solution of the nonlinear ODE (6.5 5 Figure 6.1) are shown in ﬁgure 6. r = 0. freeboundary location as indicated.9 0.5 4 4. σ = 0.1.8 0. corresponding to the smooth solutions shown in ﬁgure 6.1 0 0 0.7 0.2 for a range of values of the parameter λ.CHAPTER 6. 1 0. Results. The location of the free boundary is also clearly marked.5 2 λ = 1.1 λ=0 S 2. 0. . σ = 0. and thus reveals yet another interesting feature.3 0.2.4 0.1. However these solution branches appear to be merely a numerical artifact of the equation (and the solution technique) as increasing the resolution of the grid saw these solution branches collapse down onto the ‘stable’ branch.2. such as ﬁnite diﬀerence methods and the socalled bodyﬁtted coordinate system (described further in section 9. it would appear that it is never optimal to early exercise the perpetual option (the free boundary reaches S = 0 at λ ≈ 1. r = 0.2.5 3 3. based on (6. . 0. it was observed that multiple solution branches could be found using certain numerical techniques.1. 1. For λ 1. .1).1). These solutions exhibited nonsmooth behaviour and were thought to be a possible steadystate solution of the time dependent PDE (4. for the choice of parameters taken above.6 V PSfrag replacements 0.5 1 1. with K = 1.3).04.1). PERPETUAL OPTIONS 130 evaluate Sf .04. . λ = 0. K = 1.2 0. . namely the approach of the free boundary towards S = 0.
5 the American call option value is coincident with the European call option value and so there exists no (optimal) earlyexercise region for the American call. Sf = βK .3) Furthermore. As an alternative to resorting to fully numerical solutions we can utilise perturbation methods to obtain .CHAPTER 6. we have β → 1 and so the free boundary for the American call tends to inﬁnity and the solution reduces to the trivial solution V C (S) ≡ S. as stated in section 1. In fact the value of the perpetual American call option on an underlying paying a constant dividend yield D is given by V (S) = (Sf − K) where β > 0 is given by 1 1 β = 2 − r − D − σ 2 + σ 2 C S Sf β . however. Unfortunately the nonlinear ODE (6. Hence the value of a perpetual American call option will be trivially equal to the current value of the stock. α+1 (6. (6.1) has no analytical solution. PERPETUAL OPTIONS 131 6. If. we include the payment of a constant dividend yield in the underlying then the optimal exercise boundary becomes nontrivial and thus its perpetual counterpart will have a nonzero value. β−1 Interestingly we can see that as D → 0.2) > 0 and Sf is the (ﬁxed) location of the free boundary determined by Sf = αK .1 Analytic solutions and perturbation methods It is well known that the BlackScholes perpetual American put option admits an exact analytical solution V (S) = (K − Sf ) where α = 2r σ2 P S Sf −α (6.3.4) 1 r − D − σ2 2 2 and the free boundary Sf given by + 2rσ 2 .
V = V0 + λV1 + λ2 V2 + . (6.e. . To deal with the nonhomogeneity we seek a particular solution of the form V1 (S) = kS −2α−2 .6c) The solution to the leading order equation (6.6b).5) and collecting together the powers of λ gives the following asymptotic set of equations (cf. with α as previously deﬁned and constants A and B are to be determined from the appropriate boundary conditions. .5) (6. .4) O(λ0 ) : O(λ1 ) : O(λ2 ) : 1 2 2 σ S V0SS + rSV0S − rV0 = 0. 2 1 2 2 σ S V2SS + rSV2S − rV2 = 2V0SS (rSV1S − rV1 ) 2 2 − V0SS − 2V1SS (rSV0S − rV0 ) . where again the constants C and D are to be determined from the boundary conditions. (6.2). which can also be solved analytically. where k is to be determined.6b) becomes 1 2 2 σ S V1SS + rSV1S − rV1 = −2rα (α + 1)2 B 2 S −2α−2 . The general solution to this equation is thus V1 (S) = CS + DS −α . Hence equation (6. 2 1 2 2 σ S V1SS + rSV1S − rV1 = 2V0SS (rSV0S − rV0 ) . sections 1.1) as 1 2 2 σ S VSS + (rSVS − rV ) (1 − λVSS )2 = 0 2 and we proceed by trying a regular expansion of V in powers of λ.6a) is simply the solution to the BlackScholes perpetual option.7) .8 and 4. 2 ⇒ S 2 V1SS + αSV1S − αV1 = −2α2 (α + 1)2 B 2 S −2α−2 . Substitution into (6. Note that we are required to use this solution in order to solve the next order equation (6.CHAPTER 6. or more generally V0 (S) = AS + BS −α .7) gives (2α + 2) (2α + 3) k − α (2α + 2) k − αk = −2α2 (α + 1)2 B 2 . Substituting the above into equation (6.6a) (6. (6.6b) (6. PERPETUAL OPTIONS 132 an approximation of the solution for small values of the parameter λ. now a nonhomogeneous BlackScholes equation. i. It is useful to rewrite (6.
) = V1 (Sf0 ) + λSf1 V1S (Sf0 ) + O(λ2 ). . (6.8b) (6. . Exploiting the smallness of λ and recalling Taylor’s theorem we have V0 (Sf0 + λSf1 + . .8c) First we must also apply an asymptotic expansion to the location of the free boundary Sf . . . V1 (Sf0 + λSf1 + . .8b) thus becomes V0 (Sf0 + λSf1 + .) + λV1 (Sf0 + λSf1 + . = K − Sf0 − λSf1 − . . these shall now be determined for the case of a (nondividend paying) perpetual American put option.8a) (6. . V1 (Sf0 ) = −Sf1 1 + V0S (Sf0 ) . .9b) . PERPETUAL OPTIONS 133 which can be solved for k to yield k=− 2α2 (α + 1)2 B 2 . − (2α + 3) (α + 2) The constant B is found from the boundary conditions on the leading order equation V0 and C and D are found from the boundary conditions on the ﬁrst order correction V1 . .9a) (6.) = V0 (Sf0 ) + λSf1 V0S (Sf0 ) + O(λ2 ). V (Sf ) = K − Sf . . (2α + 3) (α + 2) Hence the solution of the ﬁrst order correction is V1 (S) = CS + DS −α 2α2 (α + 1)2 B 2 −2α−2 S . Along with the perturbation in V (S) the boundary condition (6. VS (Sf ) = −1. . . (6.CHAPTER 6. . hence equating powers of λ we can see that this boundary condition becomes O(λ0 ) : O(λ1 ) : V0 (Sf0 ) = K − Sf0 .) + . In this case the boundary conditions are given by V (S) → 0 as S → ∞. namely Sf = Sf0 + λSf1 + . . .
0 V1 (S → ∞) = 0. 0 V0 (Sf ) = K − Sf . Boundary condition 2.CHAPTER 6. Bringing things together we have that the V (S) = AS + BS −α . PERPETUAL OPTIONS 134 A similar application of Taylor’s theorem to the smooth pasting condition (6. V1S (Sf0 ) = −Sf1 V0SS (Sf0 ). V0S (Sf ) = −1.9b) results in the condition at the free boundary for the ﬁrstorder correction reducing to zero.2). This coincides exactly with the BlackScholes Perpetual Put whose solution was given by (6. 1 V0SS (Sf ) 0 General solution. 1 (2α+3)(α+2) V1 (Sf ) = 0. Boundary condition 1. However we still have the smooth pasting condition on this correction which can be exploited to give us an estimate of the correction to the free boundary. Sf1 . V0SS (Sf0 ) leading order system is given by General solution. Perturbed free boundary.10a) (6.10a) when substituted into (6. . Boundary condition 2. Sf = − V1S (Sf0 ) . This implies that the solution of V1 (S) can be determined without knowledge of the correction to the free boundary.10b) Interestingly we can see that the smooth pasting condition (6. hence A = 0.8c) yields O(λ0 ) : O(λ1 ) : V0S (Sf0 ) = −1. 0 Sf 0 = αK . therefore rearranging (6. 0 0 V0 (S → ∞) = 0. (6. α+1 The system for the ﬁrst order correction V1 is V (S) = CS + DS −α − 2α2 (α+1)2 B 2 S −2α−2 . Smooth pasting. α B = (K − Sf0 ) Sf0 . Boundary condition 1.10b) gives Sf 1 = − V1S (Sf0 ) .
i.e. . after some laborious algebra. V − V BS .11) is valid in the region S < Sf0 . i. 2α + 3 Hence our approximate solution can be written as Sf V (S) = 0 α S Sf 0 −α 2λ(α + 1)2 + (2α + 3)(α + 2) S Sf 0 −α − S Sf 0 −2α−2 + O λ2 .4 shows the same comparison as in ﬁgure 6. Figure 6. results in Sf 1 = − 2(α + 1) . ∞) since it is not entirely clear whether the approximate solution (6. Note that V1 has only been calculated in the region S ∈ (Sf0 .3 but for increasing values of the parameter λ.3 shows the ﬁrst order correction term. expectedly the agreement with the ‘exact’ solution worsens for larger values of λ.1) and the BlackScholes value. λV1 and compares it to the diﬀerence of the (numerical) solution to the full problem (6.CHAPTER 6.11) with the free boundary now located at Sf = S f 0 − 2λ(α + 1) + O λ2 . PERPETUAL OPTIONS 135 Clearly from the boundary condition at inﬁnity we have C = 0 and the condition at the free boundary leads to D= α 2(α + 1)2 Sf0 .e. (2α + 3)(α + 2) To determine the position of the perturbed free boundary involves evaluating the derivatives of V0 and V1 which. 2α + 3 Figure 6. it can be seen that there is a good agreement in the solutions. (6.
2 λ = 1.5 1 1.5 1 1.5 4 Figure 6.5.5 S 2 2.5 3 3.01 0.1 0 0 0.025 0.015 λV1 0. r = 0.2 and λ = 0. σ = 0.4: The ﬁrst order correction to the BlackScholes perpetual American put option (solid line) compared to the diﬀerence of the fully numerical option value with the BlackScholes (dotted line) for various values of λ. σ = 0. K = 1.02 0.04. PERPETUAL OPTIONS PSfrag replacements 136 0. r = 0. PSfrag replacements 0.3: The ﬁrst order correction to the BlackScholes perpetual American put option (solid line) compared to the diﬀerence of the fully numerical option value with the BlackScholes (dotted line).5 3 3.1 Sf 0 0.5 λ = 0.1.5 Figure 6.CHAPTER 6.5 S 2 2.25 0.05 λ = 0. 0.15 λV1 0.0 0. . 1.2 and λ = 0.04.1.005 Sf Sf 0 0 0 0. K = 1.
2000) is the most similar to that discussed earlier. We will not present derivations of the models below. which is more consistent when using geometric Brownian motion as the reference process. in particular by investigating their small τ behaviour.7) can be used here.Chapter 7 Other Models We have shown that signiﬁcant diﬃculties arise when the form of the function λ(S. τ ) = λS where λ ∈ R. In fact it can be seen that the Frey model is in some sense a more consistent model of price impact as this form of λ(S. namely (4. τ ) eﬀectively models the price impact on the percentage price change rather than the absolute price change.1 Frey (1998. We now discuss some of these. t) in (4. The solution turns out to be very similar. There have been alternative models proposed. t) as was brieﬂy discussed in chapter 2.1) is taken to be constant. the interested reader is referred to the appropriate references. and for a put is given 137 .10) but with λ(S. and leads ˆ ˆ to the same PDE as (2. including those which involve nonconstant λ(S. The same scaling as was employed for the fullfeedback model as τ → 0. 7. 2000) The model of Frey (1998. the model as introduced by Sch¨nbucher and Wilmott o (2000). from which we will be able to ascertain whether or not the diﬃculties outlined in the preceding chapters (for constant λ) are also present.
1) + r + 1 σ2 τ 2 √ . However. τ ) = log S K 1 ∗ 2 (7. The same argument holds in the case of call options. 2000) model vanishes when λe− 2 d1 (S . σ τ and S ∗ (τ ) is the location of the singular denominator.1) can be solved explicitly to obtain 1 2 S ∗ (τ ) = Ke−(r+ 2 σ )τ exp ±σ τ log λ2 2πσ 2 τ . Again it can be seen that there exists no solution past a critical value of τ . since the addition of S is not suﬃcient to alter the qualitative behaviour of the solution . the arguments expounded earlier (in section 3.2) Figure 7. OTHER MODELS 138 by the equation η φ − φη − 2 and for a call η φ − φη − 2 σ 2 K 2 φηη ˆ 2 1 − λKφηη 2 σ 2 K 2 φηη ˆ 2 1 − λKφηη 2 + rKH(−η) = 0. It is clear that this model will exhibit the same behaviour as the model discussed previously. Similar to the Sch¨nbucher o and Wilmott (2000) model equation (7. (7. 2πσ 2 τ hence when τ > τcrit where τcrit = λ2 .1 shows these locations for the same parameters as ﬁgure 3. 2000) is when we consider ﬁrstorder feedback. one interesting diﬀerence in the model of Frey (1998. This can be seen directly from equation (7. − rKH(η) = 0.3 and chapter 5) regarding the zero in the denominator of (2.τ ) √ =0 1− σ 2πτ where d1 (S.10) remain applicable. When considering the location of the vanishing of the denominator similar to the analysis in section 3. 2πσ 2 .it merely leads to a rescaling of λ.CHAPTER 7. Equally.7.3 the denominator of the ﬁrstorder Frey (1998.2) since there exists no real solution when λ2 < 1.
i. a1 and a2 are found by minimising the squared distance of the observed price from the model. PSfrag replacements 0 139 1.015 0. then we have τcrit ≈ 0. t) increases if the stock price drops.CHAPTER 7. o K = 1. 7. However this additional structure is not seen on the small scale close to expiry where the scaling (4.1 and σ = 0. This form of the liquidity structure incorporates so called liquidity drops.2 Frey and Patie (2002) Here an asset dependent liquidity is introduced in order to reproduce the volatility smile with ˆ λ(S.97 0. OTHER MODELS For the parameters used in ﬁgure 3.045 Figure 7. i.005 0. o where the previous analysis is thus applicable.98 0.e. (7. the model of Sch¨nbucher and Wilmott (2000).1: Location of the vanishing of the denominator of the Frey (1998.02 τ 0. λ = 0.2.1.e.e.7.02 1.039789.025 0.01 S 1 0.3) ˆ where I denotes the indicator function and λ.035 0.7) reduces (7. . ˆ which as τ → 0 reduces to λ.04 0. r = 0. τ ) = λ 1 + τ (η − η0 )2 a1 I{η≤η0 } + a2 I{η≥η0 } . 2000) (solid line) and Sch¨nbucher and Wilmott (2000) (dotted line) model with λ = 0. t) = λ 1 + (S − S0 )2 a1 I{S≤S0 } + a2 I{S≥S0 } .99 0.03 0.04 and σ = 0.2.3) to ˆ λ(S. that λ(S. i.01 0.
They do.7) leads to the following smallτ equation for a put: η φ − φη − 2 and.4).6. (7. who arrived at the following PDE 1 Vτ − σ 2 S 2 2 ˆ 1 − λVS ˆ ˆ 1 − λVS − λSVSS 2 VSS − rSVS + rV = 0. Sircar and Papanicolaou set the value of the option close to expiry to be equal to the BlackScholes analytical value. (7. It turns out that the small τ analysis considered earlier is quite similar to (7. Indeed.4) remains positive. similarly. (7.5) φηη − rKH(η) = 0. τ .3 Sircar and Papanicolaou (1998) Another interesting model is that of Sircar and Papanicolaou (1998). is switching oﬀ the eﬀect of the price impact close to expiry.4) ˆ where again λ ∈ R. note that these are the same form considered in section 4.6 but with a discontinuous jump in the value of the elasticity/liquidity parameter . oﬀer the ﬁnancial argument that transaction costs act as a natural smoothing close to strike and close to expiry and so the cost of replication (hence the price) would naturally be smoothed.10) with λ(S.6) The two equations above are subject to the same boundary conditions as employed in section 4. however. for a call η φ − φη − 2 σ 2 K 2 φηη 2 1− ˆ λK ˆ 1−λH(η) 2 σ 2 K 2 φηη 2 1− ˆ λK ˆ 1+λH(−η) φηη 2 + rKH(−η) = 0. More speciﬁcally this is done in the region 0 < τ < . This is eﬀectively introducing a smoothing in the payoﬀ function and.CHAPTER 7. where is chosen such that for τ > the denominator of (7. in particular using the scaling (4. OTHER MODELS 140 7. ˆ 1 − λVS The authors state that there will be diﬃculties with the PDE when the denominator passes through zero (as has been outlined in chapter 5 of the present study). indeed. To circumvent this diﬃculty. VS ) = ˆ λS . With a little rearranging we can see that this equates to (2.
r = 0.025 ˆ λ = 0.2: Local (τ → 0) call solution of the Sircar and Papanicolaou (1998) model ˆ K = 1. 0.10).2 0.2.2 η 0 0.9). .8) and (4.6 has now been broken. Although in ˆ both cases there appears to be very little variation with λ.2.04.015 0. Equations (7. OTHER MODELS 141 (equivalent to λ) at S = K (η = 0). suggesting yet again another solution regime. 0.05.4 0. φcall (η) = φput (−η) + rK. σ = 0. . r = 0. as is our discussion regarding the vanishing of the denominator in (2.6 is again relevant.04.005 0 0.CHAPTER 7. This matter was not pursued further but it would appear that a treatment along the lines of section 4. with K = 1. Another interesting point to note is that the symmetry seen in the fullfeedback model of section 4.035 0.6) were then solved in a manner similar to that employed on (4. and λ = 0.2. . it was found for values in excess of those shown (up to 0.02 0. this is as a consequence of the inclusion of VS into the function λ(S.4 Figure 7.5) and (7.01 0. and results for a call and a put are presented for a range of values ˆ of λ in ﬁgures 7. with the occurrence of negative square roots in the calculation. in a manner described in section 4.2 ˆ λ=0 φ PSfrag replacements 0.03 0. 7.2 in the case of calls) that the calculation failed.4 Bakstein and Howison (2003) Bakstein and Howison (2003) developed a model for liquidity eﬀects which results in . σ = 0.04 0. τ ). 0.3 in the cases of puts. .3 respectively.6. 0.2 and 7.
σ = 0. However. this is what Bakstein and Howison (2003) term price slippage. α = 1 corresponding to no ˆ slippage. Assuming that a market maker provides the quotes. hence a permanent impact will be felt by the market.4 0. but we shall return to this later.2.3.2 η 0 0.7) can be applied to the Bakstein and Howison (2003) model to obtain a valid local solution. then it is not unreasonable to expect that in the next period the market maker will quote diﬀerent prices in order to neutralise his position. and indeed this shall be done .015 ˆ λ = 0. When a large order has been placed. . the following PDE. The ﬁrst point to note is that this equation looks uncannily like a small λ expansion of the previously discussed models. First we shed a little light onto the term price slippage. 1 2 ˆ Vτ − σ 2 S 2 VSS − λσ 2 S 3 VSS − 2 1 ˆ2 3 λ (1 − α)2 σ 2 S 4 VSS − rSVS + rV = 0.3 ˆ λ=0 φ PSfrag replacements 0. and that he has also provided the other side of the large trade.05. the large trader will inevitably obtain a worse price for the order than the quoted prices.005 0. .035 0.. It can be shown that the nonsmooth scaling (4.04 0.4 Figure 7.2 0. r = 0. 2 (7.7) ˆ where λ ∈ R is a measure of the market’s liquidity and α is a measure of the price slippage impact of a trade felt by all market participants.04. and λ = 0 0.02 0.3: Local (τ → 0) put solution of the Sircar and Papanicolaou (1998) model ˆ K = 1.025 0.CHAPTER 7. . 0. a question that may be asked is whether this large order should have a permanent aﬀect on the price process or not.01 0.03 0. OTHER MODELS 142 0 0.
unlike in the Sch¨nbucher o and Wilmott (2000). However it transpires that. For the standard payoﬀ proﬁle (puts and calls) we have that. 2000) and Sircar and Papanicolaou (1998) cases. The following outlines the procedure adopted to ﬁnd such a scaling.8) we have 1ˆ 3 γτ γ−1 (f − ηfη ) ∼ λ2 (1 − α)2 σ 2 K 4 fηη τ −3γ 2 we can see that in order to obtain an appropriate balance of terms (for all τ ) we are forced to set γ − 1 = −3γ ⇒ 1 γ= .7) will dominate over the other lower order terms in VSS . τ ) = τ γ1 f (η) where η = S−K τ γ2 (7. i. Frey (1998. a suitable scaling (applicable to standard put and call payoﬀ proﬁles) can be found in the class of smooth functions. OTHER MODELS 143 in the following subsection. 1ˆ 3 Vτ ∼ λ2 (1 − α)2 σ 2 S 4 VSS .CHAPTER 7. 1 ˆ where ν1 = 2λ2 (1 − α)2 σ 2 K 4 −1 . section 3.8) where we require γ1 = γ2 = γ in order to obtain an O(1) inner solution and correctly match with the standard payoﬀ proﬁles (cf.7) it can be shown that in the limit τ → 0 the equation for the inner solution becomes the cubically nonlinear ODE1 3 fηη + ν1 (ηfη − f ) = 0 (7.e. Note that ν1 can be scaled out by setting f (η) = place the constant ν1 into the boundary conditions. 4 Hence applying this scaling to the full equation (7. Substituting this into equation (7. 2 We seek a solution of the form V (S. Hence to ﬁnd an appropriate scaling we require a balancing between this cubic term and the time derivative. the cubic VSS term in equation (7.1). due to the increasing gamma in the region close to strike and expiry.9) ˆ totic match which is the same as for the corresponding BlackScholes (λ = 0) case. The boundary conditions arise from an asymp√ ˆ ν1 f(η) but we will not do so as this would .
r = 0. α = 1.2.10) can be solved using standard ﬁnitediﬀerence techniques.10a) (7. K = 1.04. (7.4 0. the parameter ν1 increased) the solution appears to collapse down onto the standard put option payoﬀ proﬁle. in the absence of price slippage impact.7) it is clear that the scaling used above is no longer appropriate when α → 1.4: Solution to equation (7. f → 0 as η → ∞. r = 0. with the solution becoming increasingly focused around η = 0. Here the inner equation (7.7 0. 5.5 and with λ = 0. However there is another subtlety that arises with the above scaling in the limit α → 1.e. 1 0. and α = 1. the smallness of 1 − α 2 must be considered.3 0. in this case the quadratic VSS term in the equation can no longer be neglected and will be of comparable size to the cubic term for certain ranges of τ . 0. PSfrag replacements f → −η as η → −∞ for calls.8 0. the cubic term in equation .01. ν1 ↑ η 0 0.. 1. Note that as λ is decreased (i. . this can be seen by balancing the quadratic and cubic VSS terms in equation (7.01. suggesting that an asymptotic breakdown occurs when τ = O (1 − α)8 . i. f → 0 as η → −∞ for puts.1 1 0.6 f (η) 0.CHAPTER 7. When α = 1 however.9) and (7. . σ = 0. .9) becomes degenerate and looking again at the full PDE (7.9 0. 0.5. f → η as η → ∞.1 0 0. Figure 7.4 shows sample results for the inner put solution with σ = 0.e.7). In this limit.04.5.2.2 0.5 0.10b) The system (7.5 1 Figure 7. K = 1. OTHER MODELS 144 i.9) for a put option with λ = 0.e. 5.5 λ ↓. 1.5. Therefore the limit α → 1 is not a trivial one.
2 145 (7. Performing the same power balancing analysis as above leads to the conclusion that the appropriate scaling for the problem when α = 1 is given by V (S. In addition. ˆ 2 . This suggests that either the chosen form of the solution is inappropriate or that a smooth solution to the equation may not exist. Furthermore the numerics indicate that the solution most likely takes on the form of the (nonsmooth) payoﬀ proﬁle.10). rather we leave this as a subject of future research. Note that if we diﬀerentiate equation (7.CHAPTER 7.11) In this case the appropriate balancing of terms is now between the quadratic term in VSS and the time derivative. (2004).12) ˆ where ν2 = 3λσ 2 K 3 −1 and it is once again coupled with the boundary conditions (7. however neither of these satisfy the required boundary conditions (7. The results were not consistent with grid reﬁnement and kinked solutions started to appear. It is (very) interesting to note that this PDE also arises in the quadratic transaction cost model of Cetin et al.10). (7. This shall not be investigated further. τ ) = τ 3 g(ξ) where ξ = leading to the equation (in the limit τ → 0)3 2 gξξ + ν2 (ξgξ − g) = 0 1 S−K τ3 1 .9) proved ineﬀective.12) then we obtain gξξ (2gξξξ + ν2 ξ) = 0 which suggests two solutions to equation (7. 48 2 ν2 g(ξ) = D1 ξ.7) disappears and the equation is reduced to2 1 2 ˆ Vτ − σ 2 S 2 VSS − λσ 2 S 3 VSS − rSVS + rV = 0.12) g(ξ) = − C2 ν2 ξ 4 C1 ξ 2 + + C2 ξ + 1 . attempts to solve the system numerically using the same techniques employed on equation (7. 3 Note again that ν2 can be scaled out by setting g(ξ) = ν2 g (ξ) but we will not do this. OTHER MODELS (7.
if we apply the scaling (4. with the boundary conditions described in section 4.13) can be approximated by 1 2 ˆ Vτ − σ 2 S 2 VSS − λσ 2 S 3 VSS − rSVS + rV = 0. let us recall the equation arising from the Frey (2000) model.e.15) . Clearly the assumption that λSVSS 1 prohibits the denominator of the equation (7. 2 1 then we can which coincides with the Bakstein and Howison (2003) model for α = 1.6. and hence equation (7. i.13) ˆ ˆ Making the assumption that λ is small and further that λSVSS make the approximation ˆ ˆ (1 − λSVSS )−2 ≈ 1 + 2λSVSS + .1 Nonsmooth solutions to the Bakstein and Howison (2003) model As mentioned in the previous section.11).7). .CHAPTER 7. i. . 7.e. The cubic nonlinearity in the secondorder derivative φηη demands a somewhat different numerical approach from that adopted for the Sircar and Papanicolaou (1998) (7. we can obtain a valid local solution and furthermore doing so for a put we obtain 1 η ˆ φ − φη − σ 2 K 2 φηη − λσ 2 K 3 (φηη )2 2 2 ˆ λ2 − (1 − α)2 σ 2 K 4 (φηη )3 + rKH(−η) = 0.13) vanishing. (7. 2 and 1 η ˆ φ − φη − σ 2 K 2 φηη − λσ 2 K 3 (φηη )2 2 2 ˆ λ2 − (1 − α)2 σ 2 K 4 (φηη )3 − rKH(η) = 0. Vτ − σ 2 S 2 VSS ˆ 2 1 − λSVSS 2 − rSVS + rV = 0.14) (7. 2 for a call. but also leads to a regime in which standard put and call payoﬀ proﬁles are not permitted due to their inﬁnite second derivatives. equation ˆ (7. OTHER MODELS 146 Finally.4.7) to equation (7.
.5. and λ = −5.03 ˆ λ=5 0.2 0. ˆ K = 1. r = 0. Consequently.75.6)). r = 0.015 φ PSfrag replacements 0.04 0. . with σ = 0.15) do not exist (which in turn suggests the possibility of a regime akin to that described in section (4. . in some regimes the numerical scheme failed to converge. . ﬁgures 7. the Bakstein and Howison model degenerates to the standard BlackScholes equation and. .2. thereby resulting in two ﬁrstorder equations. α = 1.14) and (7.02 0.005 0. 4.4 0. 5.. σ = 0.4 0.CHAPTER 7. coupled with Newton iteration was then employed.5 and 7.6 Figure 7. Sample results are shown in ﬁgures 7. suggesting the possibility of regimes for which solutions of (7. Secondorder ﬁnite diﬀerencing.6 indicate that a nontrivial solution to the classical BlackScholes problem does .5: Local (τ → 0) put solution of the Bakstein and Howison (2003) model ˆ K = 1. . .4.01 ˆ λ = −5 0. 7.5 and 7. 0 ˆ discuss possible values for these latter two parameters (it appears that λ is likely 0.6 for puts and calls respectively. OTHER MODELS 147 model.2 ˆ λ=0 η 0 0.04. somewhat intriguingly. Bakstein and Howison (2003) do very small).04. 4.6 0.75.025 0.2. and with λ = −5.035 0. Other computations performed suggested that although the variation ˆ of option values with α is generally quite small (with ﬁxed λ).2 New nonsmooth solutions to the BlackScholes equation ˆ It should be noted that when λ = 0.5. 5. we followed a treatment which considered the problem as a system in φ and φ1 = φη . α = 1.
4 0.2 0.035 0.5.025 φ PSfrag replacements 0. exist with a discontinuous delta.e. using the Keller (1978) scheme (described in section 4. i.015 0. r = 0.005 0 0.6 Figure 7. Although this new solution is found as the limiting case of the nonlinear Bakstein and Howison model it can be illustrated by directly applying the scaling (4. equation (7.. σ = 0.02 0. . 4. quite distinct from the classical solution.15) with λ = 0. OTHER MODELS 148 0. Doing so we arrive at the equation for the inner solution of a put η 1 φBS − φBS − σ 2 K 2 φBS + rKH(−η) = 0.03 ˆ λ = −5 0.7) to build in the jump in the ﬁrst derivative.2).75.2. seeking a solution of the BlackScholes equation (3.2) of the form V BS = −τ 2 ηH(−η) + τ φBS (η).CHAPTER 7.01 ˆ λ=5 0. 5. (3.6 0.04 0. and λ = −5. η ηη 2 2 ˆ i.e. Note that this does not contradict the standard and well known uniqueness results of the (linear) BlackScholes equation as these results are only valid in a restricted class of smooth (classical) solutions to the PDE. .7 gives sample numerical solutions of the BlackScholes equation. In fact it should also be noted that there also exists inﬁnitely many smooth solutions to the BlackScholes equation but ones which do not satisfy the required growth conditions on the coeﬃcients of 1 . To illustrate the nonsmooth behaviour of these solutions ﬁgure 7.4 0.04. α = 1. .2 ˆ λ=0 η 0 0.7) to the BlackScholes equation.6: Local (τ → 0) call solution of the Bakstein and Howison (2003) model ˆ K = 1.
5 2 Figure 7. the PDE. the private information about the asset value is gradually revealed so that the price impact gradually decreases to zero at maturity. The authors’ rationale behind this choice of the liquidity function is stated that as time passes.2.7: Nonsmooth solution of the BlackScholes equation. Uniqueness of the BlackScholes PDE relies on its ability to be reduced to the heat equation where uniqueness results are well known. previously discussed in chapter 2. attempts to overcome the undesirable asymptotic behaviour at expiry by eﬀectively ‘switching oﬀ’ the eﬀect of liquidity as we approach it.5 Liu and Yong (2005) The model of Liu and Yong (2005). namely the volatility term. This is achieved by adding a time dependency to the function λ(S. β ∈ R. OTHER MODELS 149 1 0. τ ) of the form ˆ λ(S. which will .8 0. r = 0. furthermore uniqueness of the heat equation is only ensured if we prescribe the behaviour of the solution for large x.2 0 0.25 0.2 PSfrag replacements 0 0.16) ˆ where λ. Also Widder (1975) shows that there is at most one solution that is nonnegative for t ≥ 0 and all x.6 τ = .5 S 1 1. K = 1.CHAPTER 7. σ = 0. t) models absolute temperature or option prices. 7.4 τ =1 V 0.04. τ ) = λ(1 − e−βτ ) (7. a reasonable assumption when u(x.
τ ) deﬁned by (7. The best we can do . τ ) may not completely circumvent the issues associated with the vanishing of the denominator.τ )2 √ = 0.e.1 Vanishing of the denominator First we consider the ﬁrstorder feedback case. Note that this has the same structure as the standard BlackScholes model as τ → 0. equation (3. but it may for the full feedback case which we now consider. e 2 σS ∗ 2πτ where d1 (S.CHAPTER 7. To determine if the denominator vanishes in this simple case reduces to determining whether there exists a real solution S ∗ (τ ) to the equation 1− ˆ λ(1 − e−βτ ) − 1 d1 (S ∗ . The nonsmooth scaling (4.7) is no longer appropriate for this PDE and in fact it transpires that the standard BlackScholes scaling (3.5. τ ) has been deﬁned previously. it is of interest to check to see if the denominator vanishes for τ > 0. The ﬁrst point to note is that the Liu and Yong (2005) model reduces to the Sch¨nbucher and Wilmott (2000) model in o the limit β → ∞.5) and hence the asymptotic behaviour of the Liu and Yong (2005) model close to expiry will be close to the BlackScholes model.1) with λ(S. leading to σ 2 K 2 fηη + ηfη − f = 0. as such the location of the vanishing denominator in this limit was shown in ﬁgure 3. it appears that this form of the liquidity function λ(S.8 shows the location of the singular denominator for various values of β. as we shall outline next. OTHER MODELS 150 also prevents any stock price manipulation at maturity. equation (3. Figure 7. Conversely when considering the limit β → 0.16). i. i. hence we can see that the denominator does still vanish for suﬃciently large values of β. In addition. This will not cause any problems in the ﬁrstorder feedback model (as was outlined in chapter 3).3) is the relevant one to use as τ → 0. However. In the full feedback case of the Liu and Yong (2005) model at τ = 0 the denominator never vanishes because the denominator reduces to 1 at τ = 0.7.e. 7. the denominator will never vanish since here the denominator becomes uniformly unity.
035 0.1. 2π The exponential is bounded above by one and so the denominator will not vanish provided β< σK λ 2π . is investigate the region close to strike and expiry.015 0. hence if β is suﬃciently large.2.98 0. and β = 1 × 105 . 1 × 106 . λ = 0.005 0. Using this.e.CHAPTER 7.04 0.. i. if this is also in the region τ 1. In this region the Liu and Yong (2005) model is identical to the BlackScholes local solution and hence we can approximate VSS in this region by the BlackScholes local solution.025 0. i. τ in other words there is a potential for model failure (breakdown) in the region τ = O(β −2 ). r = 0.02 1. OTHER MODELS rag replacements 151 1.99 0.045 Figure 7. . .01 S 1 β increasing 0. Liu and Yong (2005) use the value β = 100 in the numerics giving τ = O(10−5 ) which is within the region where τ 1. VSS e− 2τ σ2 K 2 √ = . and the knowledge that in this region e−βτ ≈ 1 − βτ the denominator will thus vanish if and when 1− βλ σK 2 τ − (S−K)2 e 2τ σ2 K = 0. σ = 0. τ 1 and S ≈ K.02 τ 0. .8: Location of the vanishing of the denominator for the Liu and Yong (2005) model for various value of β.97 0 0.03 0. σK 2πτ (S−K)2 valid for both puts and calls. 2 × 105 . K = 1. .01 0.e.04.
. . OTHER MODELS 152 7. Note that. and for a = −1. It transpires that the relevant scaling in this case is also that of BlackScholes (3. S s (7. ds dS = egf (S. f → 0 as η → ∞.e. with σ = 0.6 Jonsson and Keppo (2002) Another model (in addition to Bakstein and Howison.9 and 7. the call option value is monotonically increasing in the parameter a.2 of Jonsson and Keppo (2002).18) admits wellbehaved solutions at times close to expiry. 0.17) where f (S.9. namely 1 Vτ − σ 2 S 2 e2aVS VSS − rSVS + rV = 0. 1. i. 2003) that can be found to be free of the problems with nonsmooth solutions for standard put and call payoﬀ proﬁles. (7.t) . Note that in this case there is no discontinuity in the derivatives and we appear to have a classical solution. However it is clear that this model will encounter diﬃculties if we have discontinuous payoﬀ proﬁles.10 respectively. S. Here the price observed in the market. 2 (7.CHAPTER 7. in agreement with proposition 3. . results for calls and puts are shown in ﬁgures 7.19) for a put.. K = 1. However the . The above system was solved using a straightforward fourthorder shooting scheme. This leads to a very diﬀerent nonlinear pricing PDE. We therefore conclude that equation (7. f → 0 as η → −∞. for a call and fη → −1 as η → −∞.3). is dependent on an exogenous Brownian motion (ds = µsdt + σsdWt ) via an exponential price impact function. such as binary options. is the Jonsson and Keppo (2002) model. which leads to the equation σ 2 K 2 e2afη fηη + ηfη − f = 0. t) again denotes the number of stocks held by the hedger and g a so called eﬀect parameter. and conversely the put monotonically decreasing.2. subject to fη → 1 as η → ∞.18) where a ∈ R.
2 η 0 0.9.2 η 0 0. .10: Local (τ → 0) put solution of the Jonsson and Keppo (2002) model K = 1.3 a=1 0. . there does.6 0.1 a = −1 0 0.1 0 0. 1. 0. a link which is outlined below.2 0. 1. 0.. OTHER MODELS 153 modelling assumptions used to arrive at this equation may be questioned. . σ = 0. σ = 0. however.CHAPTER 7.4 0.2 a = −1 a=1 PSfrag replacements 0.3 0. 0.4 0.5 0.4 f 0.4 f 0.6 0.. . and a = −1. .4 Figure 7. .4 Figure 7.2.2.2 PSfrag replacements 0.9: Local (τ → 0) call solution of the Jonsson and Keppo (2002) model K = 1. .2 0.5 0. 0.9. exist a link to the other models discussed in this chapter. and a = −1.
t) S s ds ds + gf (S. t) ds . The models.CHAPTER 7. therefore. In this case equation (7. or more speciﬁcally if gf (S. OTHER MODELS 154 7. s It is now possible to equate this model to the reduced form SDE model. t) df (S. section 1. L → ∞ the relationship (7.1 Connections with the other modelling frameworks A connection of the Jonsson and Keppo (2002) model with the models of Frey (1998. then it is possible to deﬁne the quantity s df f ds = L as the price elasticity of the market to trades (cf. S s df If we identify f as the number of shares traded and s as their price.6. . t) f ds =g .20) gives that λ → 0 and hence the liquidity of the market also increases.4). can be seen if we consider small values of the eﬀect parameter g. t) = g S. i.e. 2000) etc. t) = s s = µdt + σdWt + gf (S. if the elasticity of the market increases. become equivalent for small gf (S. t) . t) can be chosen such that λ(S. L (7. (2.17) can be approximated to dS ds ≈ 1 + gf (S. and in order for the two processes to agree we require that gf (S. t) if we allow the parameter λ to be λ(S. s S in other words the liquidity function λ(S. in line with our understanding of price elasticity.2). t) 1.20) This is intuitive since. t) ds = λ(S.
e. clustering appeared when nonoptionable stocks became optionable and disappeared when optionable stocks became nonoptionable. more speciﬁcally that on expiration dates. Krishnan and Nelken (2001) observed Microsoft stocks (during the period 19902001) and concluded that the probability of the stock price pinning at strike (i. however. despite understandable interest from market professionals. (2005). and on nonexpiration 155 . It transpires that the models outlined in this thesis can be used to partly explain and quantify such interesting market behaviour. we start by discussing some of the empirical evidence for stock pinning. The most comprehensive empirical investigation of stock price pinning was undertaken by Ni et al. (2005) suggested that this pinning eﬀect is most likely due to delta hedgers moving the price and possibly due to intentional market manipulation. Stock pinning is deﬁned as the tendency of stock prices to move to the strike price of heavily traded option contracts as the options approach expiry.Chapter 8 Explaining the Stock Pinning Phenomenon This chapter aims to illustrate one of the most documented forms of price impact. who provided striking evidence that the presence of options perturbs the prices of underlying stocks. Prior to this there had been little indication of any signiﬁcant impact. being within a small of strike) on expiration days was 23. First. Ni et al. that of the stock pinning phenomenon.29%.
e.e. Since the market makers will be long options. Avellaneda and Lipkin (2003) described anecdotal evidence of the pinning of J. In addition. This can be calculated via the Kolmogorov backward equation which is given by1 ∂P 1 2 ∂2P ∂P + µ(S. If the large institution is not also hedging the sale of the options. t) + σ (S. undertaken to maintain delta hedges on existing netpurchasedoption positions is allowed to impact the price and as a consequence pushes the stock price toward the strike price as expiration approaches.2) See for example proposition 5. but rather on calculating the socalled pinning probability. note the negative sign. With this in mind Avellaneda and Lipkin (2003) therefore make the assumption that the underlying stochastic process is modiﬁed to dS = S µ − nE ∂∆ ∂t dt + σdWt .CHAPTER 8. pinning. ∂t ∂S 2 ∂S 1 (8. counter to the assumptions made in chapter 2.6). 8. that the net position (in stock) of delta hedgers in the market is short rather than long. that the stock price will end up at S = K at t = T . t) 2 = 0. then the market makers will inevitably hedge their position. The rational behind this is as follows: If large institutions such as hedge funds sell options to market makers (deﬁned in section 1.1 Linear price impact Avellaneda and Lipkin (2003) developed a model in which stock trading. The focus of the work by Avellaneda and Lipkin (2003) is not on option pricing.1) where n is the open interest and E a constant price elasticity term. (8.D.11 of Bj¨rk (2004) o . Edwards stocks in 2003. deﬁned as the probability. denoted P(S. t). EXPLAINING THE STOCK PINNING PHENOMENON 156 days was only 13. The model assumes that the price impact is proportional to the ‘rate of change’ of the BlackScholes delta.52%. they must become short the underlying in order to hedge their position. i. then the net position (of the buyer and seller of the option) in the market for the underlying will be short. S ∂t Furthermore they argue. ∂∆ dS ∼ dt. i.
Jeannin et al. EXPLAINING THE STOCK PINNING PHENOMENON 157 where µ(S.3) numerically.3b) Furthermore. the above model can be seen as simply an approximation to the form of price impact considered in this thesis with any terms from the quadratic variation of the process and the eﬀects on the volatility ignored. the system (8. t) = 1 − exp − log(S/K) nE exp − 2σ 2 (T − t) 2π(T − t) 2 σ . correspondingly. Subsequent to the work of Avellaneda and Lipkin (2003).1 shows the solution to equation (8. the process will drift upwards more on average. To determine this probability however we are required to solve (8.CHAPTER 8.5) Figure 8. For the case when µ = 0. we assume for simplicity (and to be consistent with the original paper) that the delta is given by the delta of a call option.5) from which it is clear that there is a greater pinning probability when nE is increased.3) admits an exact solution given by P(S. P(S. (8. t) is the drift of the process and σ(S. (8. t) the volatility.1) this corresponds to solving the system ∂∆ ∂P + µS − nE ∂t ∂t ∂P 1 2 2 ∂ 2 P + σ S = 0. which can calculate directly as ∂∆ = ∂t 1 log(S/K) − r + 2 σ 2 (T − t) 3√ 2σ(T − t) 2 2π 1 log(S/K) + r + 2 σ 2 (T − t) exp − 2σ 2 (T − t) 2 . for the special case µ = 0 and r + 2 σ 2 = 0. This can be seen more clearly by considering . (8. For the model of (8. (8. Doing so we see that an increase in r results in an increased pinning probability below the strike price and. T ) = 0. a decreased probability above. and more importantly r + 1 σ 2 = 0 the authors showed that there still exists a positive 2 probability of pinning. ∂S 2 ∂S 2 1.3a) if S = K. In fact.4) 1 Avellaneda and Lipkin (2003) showed that. (2006) commented on the similarities of the model with those outlined in this thesis. for an increase in r. this might be expected since. otherwise.
e. t) S ∂∆ ∂∆ 1 ∂2∆ = µdt + σdWt − nE dt + dS + (dS)2 2 ∂t ∂S 2 ∂S = ⇒ 1 µ − nE 1 + nES ∂∆ ∂S ∂∆ σ2S 2 + ∂t 2 1 + nES ∂∆ ∂S ∂2∆ 2 ∂S 2 dt + σdWt .1 0 0.5. 5.3 0. .7 0.9 0. . K = 1.8 0. which for the case investigated above (i. 1.. this is due to the net long ∂S option position assumption. . This can be attributed to the fact that not only does the price impact aﬀect the drift of the process it also aﬀects the volatility. since 1 + nES ∂∆ > 0 in the entire domain.1: The pinning probability (8. t)dWt . 1 + nES ∂∆ ∂S (8. whereas the model outlined above includes the time derivative and the delta convexity term.6) and showed that the pinning probability was higher in this case than the corresponding Avellaneda and Lipkin (2003) model.5) for values of nE = 0.CHAPTER 8. and σ = 0.6) dS = µ(S. (2006) solve the associated Kolmogorov equation (numerically) to determine the pinning probabilities for the more accurate model (8. EXPLAINING THE STOCK PINNING PHENOMENON rag replacements 158 1 0.1.5 0.2.3 S Figure 8. hedging a long call option position) is decreased in the region of the strike. The Note that for (8.7 0.9 nE increasing 1 1.6 P 0.8 0. the following dS = µdt + σdWt − nEd∆(S.6) the problems associated with the vanishing of the denominator highlighted in this thesis are not present.1 1. ˆ ˆ S Hence the model of Avellaneda and Lipkin (2003) only includes the time derivative term of the change in delta.4 0. 2 .2 Jeannin et al. t)dt + σ (S.2 0. T −t = 0.2 1.
(2003) concluded that price impact is well described by a power .2: Comparing the pinning probability associated with (8. σ = 0. T − t = 0.2 Nonlinear price impact Recently. empirical work by Kempf and Korn (1999) brought into question the reliability of the assumption of linear price impact using empirical evidence on German index futures.2 Figure 8. Bouchaud et al.8 0. and r + 1 σ 2 = 0.85 0. (2002) showed that this relationship is sublinear (i.2 shows comparisons of the pinning probabilities for the model (8. Figure 8.4 0. EXPLAINING THE STOCK PINNING PHENOMENON 159 combined eﬀect is that the drift pushes the underlying towards the strike and the reduced volatility eﬀectively keeps the underlying in the vicinity of the strike. obtained from the associated Kolmogorov backward equation. ﬁnding that the average order book has its maximum away from the best bid (or ask).8 PSfrag replacements 0. K = 1.9 0.CHAPTER 8.6) (solid line) with the model of Avellaneda and Lipkin (2003) (dotted line) for nE = 0. Lillo et al. 0. concave) and further still.15 1. This motivated several studies to determine quantitatively how a market order of a given size would aﬀect the price of the underlying. and the Avellaneda and Lipkin (2003) model.6 0.1 0 0.1.e.2.1 1. for the same value of nE.95 S 1 1. Furthermore. Plerou et al.05 1.1. (2002) studied the distribution of the order ﬂow and the resulting average order book.2 0. possibly helping to explain the concavity of the price impact function.7 0.3 0.6). 2 8.5 P 0.
Avellaneda et al. . which is at odds with their numerical solutions of the model (8. (2003) hypothesised that p = 1/2.5.1 and 0. EXPLAINING THE STOCK PINNING PHENOMENON 160 law where the change in price (dS) caused by an order of volume w is given by dS(w) ∼ w p where p is found to be between 0. These agentbased models aim to reproduce the main stylized facts observed in real ﬁnancial markets. (2005) proposed a 3/5 power dependence whereas Gabaix et al. Almgren et al.CHAPTER 8. see LeBaron (2000) for a review of work in the ﬁeld. In addition. (2006) proposed such an agentbased model (based on a double auction) in an attempt to model stock pinning whilst also incorporating the aforementioned nonlinear price impact. again a concave function. However the value of p is greatly contested in the literature and there has been many attempts to determine its value. such as fattailed distributions of returns and volatility clustering. In recent years a number of computersimulated. This still remains an open question in the empirical literature. (2007) attempted to incorporate the empirical evidence of nonlinear price impact into the linear model ﬁrst introduced in Avellaneda and Lipkin (2003). Jeannin et al. recent work by Potters and Bouchaud (2003) suggested that this relationship may be better described by a logarithmic price impact function.6) but with increased volatility around strike rather than decreased. More recently. Simulations of the model give the same qualitative eﬀect on the drift as the model described by (8. artiﬁcial ﬁnancial markets have been constructed.6). They do so by assuming a price impact of the form dS ∼ sgn S ∂∆ ∂t ∂∆ ∂t p dt hence they assume the stochastic process for the underlying in the presence of delta hedgers hedging a long call position to take the form dS = S µ − nE sgn ∂∆ ∂t ∂∆ ∂t p dt + σdWt .
7a) if S = K. . 1.1 1. Avellaneda et al. The most consistent way to extend the model .1 0 0.7b) and conversely a nonzero probability if p > 1 .8 0.9 0.1. P(S.15 1.6).85 0.4). . 0.3: Solution to (8. (8. K = 1. Figure 8.2 and r + 1 σ 2 = 0. T − t = 0. ∂S 2 ∂S 2 1.7) for p = 0. (2007) showed that there is a zero probability of pinning if p ≤ PSfrag replacements 1 2 ∂P 1 2 2 ∂ 2 P + σ S = 0. σ = 0. EXPLAINING THE STOCK PINNING PHENOMENON 161 It is clear that the pinning probability under such assumptions will be given by the solution to the following Kolmogorov backward equation ∂P + µS − nE sgn ∂t ∂∆ ∂t ∂∆ ∂t p in conjunction with the previously calculated delta (8.2 Figure 8.7 0.4 0.2 0.95 S 1 1. otherwise.8. (8.2. .3 A new nonlinear price impact model The aim of this section is to suggest a possible extension to the model which would incorporate the empirically observed powerlaw price impact into the more accurate (linear) price impact model (8. 0.8 0.7) for ﬁve values 2 of the exponent p. 2 8. .3 shows the numerical solution to (8.5 p=1 P 0.CHAPTER 8.6 0. it can be clearly seen that the pinning probability is monotonic increasing in p.3 0. T ) = 0.05 1.9.
µ µ dt d log S p ∼ sgn(ˆ ) ˆ  2 .8a) if S = K. Note that for p1 = p2 = 1 this reduces to the model of Jeannin et al. (2006).6). is left as the subject of future research. t) are given by (8.8) for various combinations of p1 and p2 . (8. (8.8b) Figure 8. however. namely (8.6) and the parameters p1 and p2 model the ˆ ˆ degree of nonlinearity. P(S. The corresponding Kolmogorov backward equation is thus given by ∂P 1 ∂2P ∂P + sgn(ˆ) ˆp1 µ µ + sgn(ˆ ) ˆ 2p2 σ σ = 0.CHAPTER 8. σ σ dWt .4 shows the solution to equation (8. EXPLAINING THE STOCK PINNING PHENOMENON 162 would be to assume a price impact of the form dS ∼ (d∆)p S ∂∆ ∂∆ 1 ∂2∆ dS ∼ dt + dS + (dS)2 S ∂t ∂S 2 ∂S 2 p however clearly this makes little mathematical sense. t) and σ (S. We instead choose to model nonlinear price impact as3 dS = sgn(ˆ) ˆp1 dt + sgn(ˆ ) ˆ p2 dWt µ µ σ σ S where µ(S. T ) = 0. 3 Note that this is eﬀectively assuming that the change of the log price with respect to time and the Brownian motion is given by d log S p ∼ sgn(ˆ) ˆ 1 . otherwise. ∂t ∂S 2 ∂S 2 1. It appears that p2 has a greater aﬀect on the solution than p1 . further investigation of this model.
8 p2 = 1 p2 = 1.6 0. and r + 1 σ 2 = 0.4 0.05 1.6 0.9 0.2 p2 = 0.1 1.5 0. K = 1.2 .1 p1 = 1.7 p2 = 1.2 0.8 p1 = 1 p1 = 1.8 0.8 0.3 0.5) (dotted line) for T = 0. p1 = 0.2 0.2.2 0.8 0.8 0. σ = 0.8 p2 = 1 0.9 0.2 0.4: Solution to equation (8.2.2 0.4 0.8) (solid line) compared to (8. 2 p2 = 0.1. EXPLAINING THE STOCK PINNING PHENOMENON PSfrag replacements 163 P P Figure 8.8 p1 = 1 PSfrag replacements p1 = 0. 1 and 1.CHAPTER 8.7 0.2. 1 and 1.1 1. p2 = 0.2 p1 = 0.95 (a) p2 = 1.05 1.1 0 0. S 1 1.15 1.5 0.85 0.95 (b) p1 = 1.8.85 0 0. S 1 1.15 1.8.3 0.
Chapter 9 The British Option
The most important questions of life are, for the most part, really only problems of probability.  Pierre Simon Laplace (17491827) Th´orie Analytique des Probabilit´s (1812) e e In this chapter we aim to investigate the mathematical properties and ﬁnancial motivations of the newly introduced British option (see Peskir and Samee, 2008a,b), a new nonstandard class of early exercise option; such options can help to mediate the illiquidity eﬀects discussed in the preceding chapters.
9.1
Introduction
The noarbitrage price of an earlyexercise option uses the implicit assumption that the holder of the option will act optimally in the sense of following the optimal strategy of exercising upon ﬁrst hitting the rational exercise free boundary (cf. section 1.3.5). If the holder of the option chooses to deviate from this strategy, then the expected discounted payout (under the riskneutral measure) will be less than the amount the writer received for the option at the start of the contract.
164
CHAPTER 9. THE BRITISH OPTION
165
Recall that the riskneutral pricing measure is used to determine the noarbitrage price of the option, since the writer of the option can completely hedge away risk. Indeed, if we make the assumption that the writer of an earlyexercise option will hedge away the risk exposure associated with selling such an option, then from the writer’s perspective at least, the exercise strategy followed by the holder is irrelevant. For the purposes of this chapter, we shall assume that writers of earlyexercise options perfectly hedge their positions, hence eliminating any risk associated with the options. The holder of such an option is assumed not to be hedging the position; instead, he is interested in maximising proﬁt given his view on the market. Peskir and Samee (2008b) deﬁne such investors as true buyers, i.e. those who have no ability or desire to sell the option; in short holders of ‘naked’ (unhedged) options. In this case, the exercise strategy for the true buyer is of paramount importance. Furthermore, since the true buyer has no interest in hedging his position, the real world drift of the underlying will play a role in determining the rational exercise strategy. Let us suppose initially that the holder of an earlyexercise option knows with certainty the true drift of the underlying µ, and that it diﬀers from the riskneutral drift rate r. In such a situation, the optimal exercise boundary for the true buyer would deviate from the optimal exercise boundary under the assumption of a riskfree drift. In fact, the true buyer’s rational exercise boundary would be given by the free boundary of the following optimal stopping problem1 V (S, t; µ) = sup EP e−r(γ−t) (K − ST )+ , S,t
t≤γ≤T
recall that the indices of the expectation denotes that the process is started at S at time t and that the expectation is taken under the real world probability measure P . Furthermore, such a buyer would presumably not invest in a put option unless the true drift µ were less than the riskfree interest rate r, otherwise the true buyer would purchase a call option.
1
Note that here we are making the assumption of a riskneutral investor (in the utility sense).
CHAPTER 9. THE BRITISH OPTION
166
Unfortunately, the drift of a stochastic process is notoriously diﬃcult to measure, and so investors would not be able to know the true drift with any degree of certainty. They can however have an estimate of the true drift, or a ‘view’ on the future direction of the underlying, which they wish to take advantage of by buying such options. In this case the true buyer has become a speculator in the sense of seeking to maximise gains or minimise losses given his particular sentiment of future market conditions. In other words a speculator who chooses to invest in an earlyexercise option must be under the belief that the true drift of the underlying process is less than the riskneutral drift r. The British option, recently proposed by Peskir and Samee (2008a,b), is a new class of earlyexercise option that attempts to utilise the idea of optimal prediction in order to provide the true buyer with an inherent protection mechanism should the true buyer’s beliefs on the future price movements not transpire. More speciﬁcally, at any time γ during the term of the contract, the investor can choose to exercise the option, upon which he receives the best prediction of the European put payoﬀ (K − ST )+ (given all the information up to the stopping time γ) under the assumption that the drift of the underlying is µc , the socalled contract drift, for the remaining term of the contract. Hence the (now timedependent) payoﬀ proﬁle of the earlyexercise British put option is given by G(S, γ; µc ) = ER (K − ST )+ Fγ , where the expectation is taken with respect to a new probability measure R, under which the stock price evolves according to dSt = (µc − D)St dt + σSt dWtR , where we have included (in anticipation of what will follow) a constant dividend yield D. The value of the contract drift is chosen by the holder of the option at the start of the contract and is selected to represent the level of protection (from adverse realised drifts) that the holder requires; a higher µc corresponds to a lower level of protection.
CHAPTER 9. THE BRITISH OPTION
167
9.2
The noarbitrage price
Let us ﬁx a probability space (Ω, F , Q), where Ω describes a ﬁnancial market with a ﬁltration (Ft )t≥0 , which represents the information structure of the ﬁnancial market (see Harrison and Kreps, 1979) with the unique riskneutral measure Q. Assume that St is an Ft adapted stochastic process that describes the stock price process. Analogous with the American option deﬁned in subsection 1.3.5, the noarbitrage price of the British put option is given by the supremum over all stopping times γ (adapted to the ﬁltration Ft generated by the process St ) of the expected discounted future payoﬀ. In contrast with an American option, the future payoﬀ is now itself an expectation, i.e. V (S, t) = sup EQ e−r(γ−t) ER (K − ST )+ Fγ S,t
t≤γ≤T
.
Recall that the future payoﬀ is deﬁned as the best prediction of the European payoﬀ, conditional on all the information available up to the stopping time γ. There are numerous approaches to evaluating this expectation, for example we could directly use the probability density function of the process under the measure R, to give ER (K − ST )+ Fγ =
∞ 0
(K − z)+ f R (S, γ; z, T )dz,
where f R (S, γ; z, T ) is the transitional probability density function of the process started at time γ at the position S and ﬁnishing at time T at the position z and is given by2 f (S, γ; z, T ) =
R
σz
1 exp 2π(T − γ)
− log
z S
1 − µc − D − 2 σ 2 (T − γ) 2σ 2 (T − γ)
2
.
Tackling the above integral would provide us with the required expectation; however, we shall adopt an alternative approach in order to give some intuition behind the resulting expression. We can evaluate the expectation using direct integration with respect to the probability measure R over the probability space Ω, i.e. ER (K − ST )+ Fγ =
2
Ω
(K − ST )+ dR.
See appendix C for a derivation.
t) = sup t≤γ≤T The British option EQ S. log K−y − µc − D − 1 σ 2 (T − γ) Sγ 2 . Finally we have our expression for the conditional expectation K log K−y − µc − D − 1 σ 2 (T − γ) Sγ 2 dy √ ER (K − ST )+ Fγ = Φ σ T −γ 0 = Sγ 0 K Sγ Φ 1 log u − µc − D − 2 σ 2 (T − γ) √ σ T −γ K−y .CHAPTER 9. Further simpliﬁcation gives K 0 (K − z)R(z ∈ dzFγ ).t e −r(γ−t) Sγ 0 K Sγ Φ log z − µc − D − 1 σ 2 (T − γ) 2 √ σ T −γ dz . . In order to evaluate this integral we rewrite the function K − z as an integral to give K z=0 K−z K K−y dyR(dz) = y=0 y=0 z=0 R(dz)dy. where we have also changed the order of integration. √ = Φ σ T −γ 1 2 e− 2 y dy. Sγ du. ∞) and so the integral can be written as ∞ 0 ∞ 0 ∞ 0 (K − z)+ dR(z) = (K − z)+ R(dz) = (K − z)+ R(z ∈ dzFγ ). where the probability is now conditional on the ﬁltration of information up to the stopping time γ. We can now exploit the fact that (see appendix C) K−y 0 where Φ denotes the standard normal distribution function 1 Φ(z) = √ 2π z −∞ R(z ∈ dzFγ ) = PR [z ≤ K − yFγ ] . THE BRITISH OPTION 168 The random variable is a real valued function whose domain is given by ST ∈ [0. where at the ﬁnal step we have made the substitution u = value is thus given by V (S.
3 It remains to ﬁnd an analytical expression for G(S. t. µc ) . with the optimal stopping time deﬁned as γ∗ = inf{t ∈ [0. µc ) − Se(µc −D)(T −t) Φ d2 (S. t) > G(S. we have that the problem now reads V (S. t. section 1. t) = KΦ d1 (S. GA (S. General optimal stopping theory can now be applied to this problem analogous with the American option problem (cf. t) : V (S. t. µc ) = log log K S K S (9.5) we have that C = {(S.2a) (9. equation 1.e. T ] : St ∈ D}. 3 . µc ) = d2 (S. t.2b) Note that the dependency on the contract drift rate µc has been stated explicitly. t)} (continuation set). t) = G(S. σ T −t (9. G(S. µc ) − σ T − t. t) = KΦ log K S − µc − D − 1 σ 2 (T − t) 2 √ σ T −t log K S − Se(µc −D)(T −t) Φ Alternatively this can be written as 1 − µc − D + 2 σ 2 (T − t) √ σ T −t . t) which can be done easily with integration by parts to give G(S. D = {(S.t t≤γ≤T This is the standard form of an Americantype option (cf. γ) .3. S. For the standard American put option the gain function is simply the (time homogeneous) payoﬀ proﬁle i. t. t) = sup EQ e−r(γ−t) G(S. σ T −t √ − µc − D + 1 σ 2 (T − t) 2 √ = d1 (S. t) = S 0 K S Φ log z − µc − D − 1 σ 2 (T − t) 2 √ σ T −t dz.11) and so we can directly see the links with the existing American option theory.CHAPTER 9.1) − µc − D − 1 σ 2 (T − t) 2 √ . t)} (stopping set). THE BRITISH OPTION 169 Deﬁning the function G(S. where d1 (S. t) = (K − S)+ . t) : V (S.
∞)}. and hence C = {(S. again analogous to the American put option. THE BRITISH OPTION 170 i. the ﬁrst time that the stock price enters the stopping region.CHAPTER 9. The above identity can also be used to check that we are calculating G(S.3) with the gain function G(S. VS (S. V > G in C.2. t) correctly. t) = G(S. t) = Ke−r(T −t) Φ d1 (S. t). t) and the analytical expressions of the corresponding European option values.1). t. t. t) given by equation (9. t) on S = b(t). let us ﬁrst take a closer look at the gain function. t SS S 2 V (S.1 The gain function This section brieﬂy comments on the links between the gain function G(S. where d1 (S. Note that this looks very much like the gain function and indeed if µc = r. This is intuitive as the payoﬀ is the best prediction received now and so involves no discounting. t) on S = b(t) (smooth ﬁt). Also note that as t → T we have that Φ(d2 ) → Φ(d1 ) → 0 for S > K and conversely Φ(d2 ) → Φ(d1 ) → 1 for S < K . then it is clear that P G(S. V = G in D.e. 9. t) : S ∈ (b(t). r) − Se−D(T −t) Φ d2 (S. Now applying standard optimal stopping and Markovian arguments. (9. Note that the BlackScholes European put value is given by P VE (S. t. r) emphasises the dependence on the parameter r and is given by equation (9. It can be shown (see Peskir and Samee. t)µc =r = er(T −t) VE (S. r) . t) = GS (S. Before we continue. ∞). unlike the European option value. 2008b) that the stopping and continuation regions are separated by a smooth function b(t).2a) for µc = r. the problem can be conveniently expressed as the following freeboundary problem: V + 1 σ 2 S 2 V + (r − D)SV − rV = 0 for S ∈ [b(t). the free boundary.
In addition it is advantageous to make the transformation τ= √ T − t. For example. this suggests that the British option may be considered as a compound option. transforming to logspace via ˆ the transform S = log S K will simplify the resulting equations. In addition. In fact. t) at every node and iteration.CHAPTER 9. subject to the constraint V (S. T ) = (K − S)+ . This type of option has been studied extensively.3) can be solved numerically via a number of diﬀerent methods. namely G(S. i. 9. i. which is consistent with the fact that at t = T clearly the best prediction of the put payoﬀ will be the payoﬀ itself. (9. which attempts to solve the appropriately discretised and linearised system (typically based on a CrankNicolson discretisation scheme). Whaley (1982) and Hodges and Selby (1987). an option on an option.3 Numerical treatment The (nonlinear) system (9. By far the easiest to implement is the Projected Successive Over Relaxation (PSOR) algorithm. a standard American option written on an underlying European option has simply the same value as the underlying European option. the entire . Geske (1979). However the innovation with the British option considered here is that the two options are priced under diﬀerent measures. t) ≥ G(S. compare Geske (1977). THE BRITISH OPTION 171 which indicates that in this limit the gain function takes on the form of the standard put payoﬀ condition.e. resulting in nontrivial solutions. the region where the solution is changing most rapidly. and often leads to trivial solutions. there exists no optimal earlyexercise region.4) which has the eﬀect of concentrating the grid points close to expiry.e. Furthermore.
6) where the gain function must also be transformed via (9. ∞) onto the ﬁxed domain (9. but it . ∞).5). where the free boundary now becomes an additional variable in the problem.5) ˆ S ∈ [1.3) can be nondimensionalised via an appropriate set of transformations.CHAPTER 9. increasing the computation time at a disproportionate rate to the increased accuracy gained in the freeboundary estimate. 2 ∂τ V (S. ˆ This eﬀectively maps the continuation region S ∈ [b(τ ). The drawback of this method is that we now have a more complicated equation to solve. S= b(τ ) (9. The idea behind this technique is to make the transform (in addition to the time change (9. ˆ ˆ ˆ ˆ ˆ ˆ ˆ VS (S.4)). t) on S = 1. t) = GS (S. its major drawback (for the purposes of this exposition) is that in order to determine the location of the free boundary accurately. ∞). which for the interested reader can be found in appendix B. Nondimensionalising makes the system more parsimonious. then the number of grid points must be increased. t) on S = 1 (smooth ﬁt). but to retain ﬁnancial intuition. Despite the PSOR algorithm’s ease of implementation. For these reasons a more sophisticated method. 1]. was employed for the results given in the present chapter. ˆ V > G for S ∈ (1. V = G for S ∈ [0. and further if a more accurate estimate of the freeboundary location is needed. S ˆ . ∞). the bodyﬁtted coordinate system. t) = G(S. in what follows we choose to retain the dimensional form. Bodyﬁtted coordinates where ﬁrst proposed by Landau (1950) and later applied to ﬁnitediﬀerence schemes by Crank (1957). THE BRITISH OPTION 172 system (9. Making these changes of variables yields the following modiﬁed ﬁxed boundary problem 1 1 ˆ ˆˆ ˆ ˆ ˆ 2τ Vτ − 1 σ 2 S 2 VS S − r + 2τ b(τ ) ∂b(τ ) SVS + rV = 0 for S ∈ [1.4) and (9. More recently Widdicks (2002) adapted the scheme to solve the standard American option problem and Johnson (2007) to more complex options (involving multiple free boundaries). some form of interpolation is needed.
2: The British put option free boundary for varying volatilities.11.CHAPTER 9. 0.115.5 t 0..95 0.2 0. µc = 0.125.1.1 0.1 shows the location of the British put option free boundary for varying values of the contract drift µc and in addition ﬁgure 9.5. 0. K = 1. 1. ..05 1 0.65 0 0. THE BRITISH OPTION PSfrag replacements 173 is now in a ﬁxed domain and standard ﬁnitediﬀerence methods for such problems are well understood and easily applicable.9 S 0.1 1 0. T = 1.7 0. K = 1.75 0.1.8 0. r = 0.4.05.1. . D = 0. T = 1.1: The British put option free boundary for varying values of the contract drift. Figure 9. D = 0. σ = 0. 1. Note that this method provides highly accurate results for the freeboundary locations. . . and µc = 0.16.12. and σ = 0. .9 1 Figure 9. 0.9 S 0.2 0.5 0 0.9 1 Figure 9.7 0.3 0.7 0.1 1.6 0. .4 0.8 0. since no interpolation to calculate the free boundary is required.4 0.3 0.8 µc increasing 0.85 0.5 t 0.8 σ increasing 0. r = 0. 0.6 0.6 PSfrag replacements 0.2 the variation with the volatility .1 0. 0.7 0.
It can be seen that as the contract drift parameter increases the free boundary (and its value at expiry) collapses downwards monotonically to zero. In fact the exact location to which the free boundary asymptotes and its behaviour close to expiry is determined in section 9.e. i. The most striking diﬀerence with the American option free boundary is that. Furthermore. regularity and smoothness shall be implicitly assumed. Also the volatility appears to have a signiﬁcant inﬂuence on the shape of the free boundary.5. THE BRITISH OPTION 174 (for a ﬁxed µc ). Another interesting point to note is that for a given investor at t = 0 with current stock price S0 . . although in the latter case all free boundaries asymptoted to the same value at expiry.. also resulted in diﬀering free boundaries. Finally.e. Note that some parameters (such as volatility) have been chosen artiﬁcially high in order to illustrate the distinct behaviour of the British option free boundaries. r µc = const. More information about the relationship between the parameters could be gleaned from a consideration of the fully nondimensionalised problem (see appendix B) but this shall be left as the subject of future research. there exists a value of µc below which all British option free boundaries at t = 0 lie above the current stock price. hence the optimal investor would choose to exercise immediately. in particular in the asymptotic analysis. numerical investigations showed that varying µc but keeping the diﬀerence between µc and interest rate r constant resulted in distinct free boundaries. b(0) > S0 and so the investor is automatically placed in the exercise region at the initiation of the contract. varying µc and keeping the ratio of the contract drift to interest rate constant. This nonmonotonic behaviour can lead to complications in the convergence of the numerical schemes employed and in addition can result in subtle diﬃculties when trying to prove the regularity and smoothness of the free boundary (see Peskir and Samee.CHAPTER 9. In what follows. i. the boundary behaviour is no longer monotonic. for some values of the contract drift at least. 2008b).
t) = L G(S. t) + E Q 0 γB −t e−ru H(St+u . we will ultimately be interested in the behaviour of Sh in this limit. Φ (d2 ) = 4 K −(µc −D)(T −t) e Φ (d1 ).e.1 suggest that the free boundary for the British put option either tends to inﬁnity or tends to zero in this limit. S (9.8) it is clear that the expected future gain (the lefthandside) must be greater than the current gain G(S. It is a standard result from optimal stopping theory that the free boundary cannot be contained in the region in which H(S. From (9. t) ∈ B}. indicating it would not be optimal to stop at (S. dependent on the choice of the parameter µc . / i. t) > 0. t) = Gt + σ 2 S 2 GSS + (r − D)SGS − rG. t) > 0.9) See for example Peskir and Shiryaev (2006). the ﬁrst exit time of the process from the domain B deﬁned as an arbitrary small halfcircular domain around the point (S.CHAPTER 9. t) = 0 will act as an analytical proxy for the free boundary. (9. Making extensive use of the identity. the numerical results showed in ﬁgure 9. or at least provide an upper bound on the location of the free boundary at any given instant. . The following analysis will try to shed some light on these two distinct regimes of behaviour. 2 (9. t). First we introduce the following function 1 H(S.e.8) where γB is deﬁned as γB = inf{t ∈ [t. γB ) = G(S.4 Free boundary analysis far from expiry Far away from expiry. As such the solution to H(Sh . i. Since this section is concerned with the large time to expiry behaviour of the free boundary. T ] : (St . t) if H(S. t + u)du . t) > 0. THE BRITISH OPTION 175 9.4 This can be easily seen by directly applying Itˆ’s formula to the discounted gain function and o taking expectations to obtain E Q e −r(γB −t) G(SγB .7) which will be of use in our analysis. t) and so the free boundary cannot be located in the region where H(S. T − t → ∞ (eﬀectively the perpetual limit). since it is a known analytic function which we shall calculate shortly.
. It can also be seen that for large enough values of the contract drift. namely tending to inﬁnity or to zero.10) GS = −e(µc −D)(T −t) Φ(d2 ). Note that the converse is not necessarily true. t) − rKΦ d1 (Sh . Sh (t) appears to tend to zero with no visible turning point. the same qualitative behaviour that can be seen for the true free boundary.11) This pleasingly simple expression for the Hfunction can now be used to provide us with an upper bound on the location of the free boundary. (9. This is an important point. we are able to infer that the free boundary must also tend to zero in that limit.12) for Sh = Sh (t). since if we can show under what circumstances the value of Sh (t) tends to zero for large times to maturity. we can directly compute each term in the Hfunction as Gt = − σKe− 2 d1 2 2π(T − t) Ke− 2 d1 σS 2 2π(T − t) 1 2 1 2 + (µc − D)Se(µc −D)(T −t) Φ(d2 ). GSS = . and furthermore in the limit T − t → ∞.7) yields5 H(S. Guided by numerical diﬀerentiation it appears that the solution for Sh for large values 5 Note also that the dividends do not occur anywhere but in the exponents.CHAPTER 9. The ﬁrst point to note is that Sh is not monotonic for some values of the contract drift µc . It is hypothesised that if Sh → 0 as T − t → ∞ then since the free boundary must lie below Sh (t) then the free boundary must also tend to zero as T − t → ∞. t) = µc Se(µc −D)(T −t) Φ(d2 ) − rKΦ(d1 ). To do this we are interested in the solution to the equation µc Sh e(µc −D)(T −t) Φ d2 (Sh . suggesting a critical value of µc which separates two distinct regions of asymptotic behaviour of Sh (t). (9. t) = 0 (9. which after substitution into (9. THE BRITISH OPTION 176 where primes denote derivatives.12) obtained using standard NewtonRaphson iteration. Figure 9.3 shows the solution of equation (9.
5 Sh 1 µc increasing 0. µc = 0. i. Sh (t) ∼ Aeβ(T −t) as T − t → ∞.2 (Sh .1.e. r = 0. T = 50. 2 which as T − t → ∞ the second term dominates giving T −t→∞ lim d1. Sh (t). σ Now it is clear that in the limit T − t → ∞. of T − t takes on an exponential form.102.2 (Sh . K = 1. and σ = 0. for varying values of the contract drift. Therefore making this ansatz we can see that d1.14) For convenience in what follows we shall deﬁne 1 µc − D ± 2 σ 2 + β ν± = . the functions d1. In order to provide an appropriate . 0.13) − µc − D ± 1 σ 2 + β (T − t) 2 √ .e. D = 0. i. .4.104.2 will tend to ±∞ depending on the sign of the parameter ν± . .3: The zero of the Hfunction. . σ T −t √ 1 µc − D ± σ 2 + β T − t. THE BRITISH OPTION rag replacements 177 2 1. t) = − 1 σ The equation for Sh (t) in the limit T − t → ∞ thus becomes µc Ae(µc −D+β)(T −t) Φ − 1 (µc − D + 2 σ 2 + β) √ T −t σ 1 (µc − D − 2 σ 2 + β) √ = rKΦ − T −t . t) = log K A (9. 1. . σ (9.CHAPTER 9.5 0 0 10 20 t 30 40 50 Figure 9.
THE BRITISH OPTION 178 balancing of terms in the following analysis we assume that both functions d1. noting that Sh is a function of t. 2 1 2 rK e− 2 ν− (T −t) + . if 1 1 β > σ 2 − µc + D > − σ 2 − µc + D...15) 1− 3 (−1)n 1.+ +.CHAPTER 9. only provides us with one equation for two unknown constants A and β.. Next we use the well known result (see for example Abramowitz and Stegun. Equation (9. ν+ ν− (9.12).. 2 2 a condition which can (and will) be checked a posteriori. . . .. to obtain µc e(µc −D)(T −t) dSh ∂d1 σµc − (µc − D)Sh Φ(d2 ) + KΦ (d1 ) (µc − r) + √ dt ∂t 2 T −t = 0. .. the cumulative normal distribution function Φ(z) has the following asymptotic expansion e− 2 z Φ(z) = − √ z 2π 1 2 (9.14) becomes µc Ae(µc −D+β)(T −t) − 1 ν+ (T −t) 2 = e 2 ν+ 2π(T − t) ν− that 1 (µc − D + 2 σ 2 + β)2 1 2 µc − D + β − ν + = µ c − D + β − 2 2σ 2 (µc − D − 1 σ 2 + β)2 2 =− 2σ 2 1 2 = − ν− . .17) Finally considering the exponent of the exponentials on the lefthandside we can see which remarkably allows us to cancel through all exponential terms leaving the much simpler equation rK µc A = . however. ν± 2π(T − t) 1 2 (9. z2 z z 2n (9. i.3 . In order to determine their value uniquely we can obtain another equation by diﬀerentiating (9. This can only be achieved if ν± > 0. 1968) that for large negative values of the arguments z.16) note that a similar expression exists for large positive z. 2π(T − t) e− 2 ν± (T −t) + .e.2 → −∞ in the limit.18) This.. (2n − 1) 1 + 4 +. .16) gives to leading order that √ Φ −ν± T − t = therefore equation (9. .
19) Now eliminating the unknown A from equations (9.21).13) and performing a similar analysis to the above.20) Finally solving the (quadratic) equation (9. Also note that since µc +r µc −r > 1 for µc > r.15). (9.CHAPTER 9. c . again provided ν± > 0 µc K (µc − D + β) A = (µc − r)ν− + σµc . substitution of the found value for β.20) for β yields 1 β = σ2 2 µc + r µc − r − µc + D (9.19) leads to the following 2 2r (µc − D + β) = (µc − r)ν− + σµc ν− .2) that the volatility appears to greatly aﬀect the behaviour of the free boundary. back into equation (9. THE BRITISH OPTION 179 Once again making the ansatz (9.e. suggesting that it is always optimal to early exercise immediately. Equation (9. will behave as Sh (t) = Keβ(T −t) where β is given by equation (9. i. equation (9. Finally. 2 (9.18) gives immediately that A = K.15).18) and (9.21) is also consistent with the observation (as seen in ﬁgure 9.e.21). ν+ 2 (9. Also note that this result is consistent with the fact that as µc → r the free boundary appears to be tending to inﬁnity at an increasing rate. The more useful corollary of this result however is that clearly if β > 0 then Sh (t) → ∞ for large values of T − t and conversely if β < 0 then Sh → 0. we can arrive at the following expression. The critical value of the parameter µc which separates these two distinct regimes is given by the solution to the equation 1 β = σ2 2 µ∗ + r c ∗ −r µc − µ∗ + D = 0.21) is consistent with the initial assumption (9. 2 1 ⇒ 2r (µc − D + β) = (µc − r) µc − D − σ 2 + β 2 1 + σµc µc − D − σ 2 + β . adding credence to the obtained result. and so to summarise we have found that for T − t → ∞ the zero of the Hfunction.21) where we have taken the positive root in order not to violate the assumption (9. the positive root. Sh . i.
Furthermore any numerical treatment of the freeboundary ˆ problem in (S. The large time to expiry behaviour can be used to make an appropriate transformation that removes the observed blow up to inﬁnity of the free boundary for large T − t.5 Analysis close to expiry Also of great interest is the behaviour of the free boundary close to expiry.CHAPTER 9. however only c 2 the positive root will be greater than r. In fact as the function Sh (t) ˆ provides an upper bound for the location of the free boundary. t)space will be better behaved in the absence of any numerical breakdowns. will tend to ﬁnite or zero values.23) may help remove any exponentially growing behaviour. 9. Knowledge about the large time to expiry behaviour of the Hfunction is not only useful in determining diﬀerent regimes of behaviour of the free boundary. all free boundaries. but can also be used to improve the eﬃciency of the numerical calculations of the option value and its corresponding freeboundary location. + 4r 1 2 σ −D 2 1 2 . it is clear that in (S. If D > 1 σ 2 then we could have two positive solutions. We have shown that in the limit Sh (t) takes on the form Keβ(T −t) . We must determine the value to which the free boundary asymptotes and also the functional . t)space. regardless of the value of the contract drift µc . (9. THE BRITISH OPTION 180 hence 1 µ∗ = σ 2 c 2 or 1 1 r + D + σ2 µ∗ = c 2 2 1 + 2 1 r + D + σ2 2 2 µ∗ + r c µ∗ − r c + D. suggesting that the transformation ˆ S = Se−β(T −t) (9. In addition the square root will always stay positive for all value of D and so we will always have a critical value.22) 1 If D < 2 σ 2 then we are required to take the positive square root in order to obtain a positive µ∗ .
When investigating the small time to expiry behaviour of free boundaries arising from derivative securities with earlyexercise features. THE BRITISH OPTION 181 form in which it asymptotes to that value. exploits the Green’s function for the heat equation to convert the resulting boundary value problem to an integral equation. Using this information as t → T we have that the Hfunction is trivially zero for S > K. In addition to being interesting in its own right. Johnson (2007) or the recent survey article by Howison (2005).3. (1995). Hence we have that Sh (T ) = rK . Recall that the region in which H(S. knowledge of the free boundary behaviour close to expiry can be exploited to improve the eﬃciency of numerical schemes used in determining the free boundary.CHAPTER 9. which is then solved asymptotically for times close to expiry. providing the correct location of the free boundary at expiry can be crucial in the scheme’s success. there are a multitude of approaches that can be taken. The value to which the British option free boundary asymptotes can be obtained simply via a cursory inspection of the Hfunction deﬁned in the previous section and given by (9. In order to convince ourselves that it should in fact coincide with Sh (T ) we consider a situation in which the process is an arbitrarily small time away from expiry and that the current price is below the . for example see Wilmott et al. which investigates the standard American put option.11). For example when using the bodyﬁtted coordinate system described in section 9. µc rK µc and that the Hfunction is positive in the region S > and so we can conclude that the free boundary must lie at or below this value. For example the analysis of Kuske and Keller (1998). t) > 0 cannot contain the free boundary. T ) = µc S − rK. Alternatively the diﬀerential form of the freeboundary problem can be tackled directly and matched asymptotic expansions can be used to investigate the solution behaviour close to expiry. and for S < K is given by H(S.
since the process is extremely close to expiry. µc (9. hence the optimal investor would stop in this region leading to the conclusion that b(T ) = rK . The above result can be shown using a diﬀerent approach which we shall brieﬂy outline below. One of the hallmarks of the existence of an earlyexercise region is the existence of a region in which the value of the earlyexercise option’s European counterpart (i. In addition rearranging the governing PDE directly to obtain an expression for Vt and using the fact that V (S. Close to expiry this amounts to determining if the dynamics of the PDE move the option price below the gain function at a small time prior to expiry. T ) and we can also directly evaluate the expression for Gt . this fact is easily conﬁrmed numerically. ∂t ∂t By deﬁnition V (S. with no possibility of the process reaching the region in which H(S. rK . More speciﬁcally if the quantity ∂ V (S. Hence there exists a (possible) earlyexercise region when rK − µc S > 0 ⇒ S < in agreement with (9. namely (9.CHAPTER 9.24). T − δt) ∂t ever becomes positive. T − δt) − G(S. t) > 0.24) Note that the inclusion of dividends does not aﬀect the location to which the free boundary asymptotes as one might expect. T ) − (S. T − δt) − G(S.10). THE BRITISH OPTION rK . T ) + δt ∂t ∂G ∂V (S. µc 182 value In this region H(S. at t = T . T ) = G(S.e. T − δt) ≈ δt(rK − µc S) ∂t for S < K and trivially zero otherwise. µc . T ) − G(S. with no early exercise) lies below the payoﬀ function (gain function). T − δt) − G(S. T ) = (K − S)+ leads to ∂ V (S. T ) . A simple Taylor expansion leads to ∂ V (S. t) < 0 and so the investor is accumulating negative gain. T − δt) ≈ V (S.
e. we are therefore interested in the behaviour of the gain function as we approach expiry. the leadingorder behaviour is identical to that of the standard American put option with dividends. THE BRITISH OPTION 183 However. t) = . t → T ) the ﬁrst term in d1 (S. for small values of T − t. The ﬁrst point to note is that in this limit ˆ that as T − t → 0 (i. with 1 + 2 σ 2 (T − t) ˆ √ d1 (S. t) . then the governing PDE becomes 1 ˆ ˆ ˆ Vt + σ 2 S 2 VS S + (r − µc ) SVS − rV = 0.CHAPTER 9. namely ˆ S = Se(µc −D)(T −t) . Also note that the actual dividend yield D has been completely removed from the problem by the scaling (9. t) = KΦ d1 (S. t) will dominate giving t→T d2 → d1 and so it will suﬃce to determine the behaviour of d1 in this limit. σ T −t √ ˆ ˆ d2 (S.e. ˆˆ 2 and the gain function is transformed to ˆ ˆ ˆ ˆ G(S. It transpires that the answer is pleasingly simple.25) log K ˆ S Notice that the parameter µc has been taken out of the gain function (and hence the boundary conditions) and placed into the PDE. t) − σ T − t.25). (9. t) = d1 (S. t) − SΦ d2 (S. where it appears as a pseudodividend. It is clear 1 ˆ log lim d1 (S. of more interest is the functional form in which the free boundary approaches (9.24). the proof of which is given below. If we make a subtle change of variable. indicating that dividends have a rather benign aﬀect on the dynamics of the British put option. i. Since we are interested in the behaviour of the free boundary as we approach expiry. t) = lim √ t→T σ T − t K ˆ S . .
such that these breakdowns do not interfere with the free boundary at expiry. Hence to leading order as t → T we have that the British put option freeboundary problem (9. the standard put option payoﬀ.26) decays as (T − t) 2 e− T −t in the limit t → T . In the region S > K if d1 → −∞ then we can exploit the asymptotic expansion of the cumulative normal distribution function (9. under the transformation (9. t) = K − S i. which as t → T tend to 0 and ∞ respectively. Clearly this decays much faster than the terms arising from the asymptotic (power series) expansion of the PDE and consequently the gain function can be approximated to6 ˆ ˆ G(S. d1 2π ⇒ Φ(d1 ) = −σ T −t log 2π K ˆ S −1 1 2 ˆ < 0 for S > K. ..25). t) = (K−S)+ −σ T −t log 2π K ˆ S −1 2 1 log exp − 2 2σ (T − t) K ˆ S ˆ (K−S)+..16) to give e − 2 d1 Φ(d1 ) = − √ + . ˆ Similarly for the region S < K where d1 → +∞ we can use the symmetry of Φ(·) to obtain e − 2 d1 Φ(d1 ) = 1 − √ + . . ..e. THE BRITISH OPTION 184 Since log K ˆ S ˆ ˆ and conversely for S < K then d1 → +∞ as t → T . Using the above and further the fact that Φ(d1 ) → Φ(d2 ) as t → T . d1 2π ⇒ Φ(d1 ) = 1 − σ T −t log 2π K ˆ S −1 1 2 exp − 1 log 2 (T − t) 2σ K ˆ S 2 + . where A must be a positive constant. . is identical to the standard American put option freeboundary problem with dividends. . . (9. where the dividend amount (yield) Note that there is a breakdown if S < Ke−B T −t or S > Ke+B T −t .3). Note that the secondorder term in (9. .26) in the region close to expiry.. .. . 6 √ √ 1 A + . the gain function can be approximated by ˆ ˆ G(S.CHAPTER 9. then in this region it is clear that d1 → −∞ as t → T 1 log exp − 2 2σ (T − t) K ˆ S 2 + ..
4517. Fortuitously the asymptotic form of the American put with a constant dividend yield close to expiry has been studied extensively.27) i.25). where the constant α0 is the solution to the transcendental equation 3 2 α0 e α0 ∞ α0 e−u du = 2 1 2 2α0 − 1 .e. the free boundary approaches expiry parabolically. see for example Evans et al. Note. Note that if we are interested in higherorder asymptotics then the behaviour will diﬀer to that of the American put with dividends. (2002). Also note that for the American put the free boundary takes on a very ˆ diﬀerent functional form for diﬀerent values of the dividend parameter.CHAPTER 9. i.e.25). however. is given by b(t) ∼ rK 1 − σα0 µc 2(T − t) e−(µc −D)(T −t) . after transforming back to the original variables via (9. Figure 9. and the leading order behaviour is given by b(t) ∼ rK 1 − σα0 ˆ D 2(T − t) ˆ for constant dividend yield D > r. Hence the behaviour of the British put option free boundary will coincide with that of the American Put option with dividends. for D = r . numerical investigations performed by the author indicate that (9. 4 which can be calculated to give α0 ≈ 0.4 shows the above approximation compared with the free boundary obtained via a full numerical treatment using the techniques described in section 9.27) becomes a better approximation (over the same scale) as the volatility is decreased.3. (9. In addition. that to leading order the behaviour of the free boundary remains unchanged by the transform (9. THE BRITISH OPTION 185 is given by the contract drift rate µc . Therefore the asymptotic form of the British put free boundary. This is most likely due to the fact that for larger σ the turning point of the free boundary occurs closer to expiry and the free boundary would diverge from the parabolic approximation over a shorter timescale. This is due to the various terms we have neglected as t → T from the gain function and also from the exponential transform.
6 Financial analysis of the British put option The protection feature of the British option that was so crucial in motivating its introduction led Peskir and Samee (2008a.775 0 0.8 0.125.b) to investigate further the motivations of a British option investor. rag replacements 186 t 0. the value of the dividend D does not qualitatively change the asymptotic behaviour of the British put free boundary since the scaling (9.01. Here we reproduce this analysis in more detail by exploiting the advanced numerical techniques employed to obtain relativereturn surfaces of the options. and D = 0.4. r = 0. or more speciﬁcally that the realised drift will be less than the risk neutral rate r. in particular the potential return on investment in the British (and other) options.004 Figure 9.79 0.1.008 0.4: The asymptotic approximation for the British put option free boundary close to expiry. T = 0. Furthermore.01 9. THE BRITISH OPTION 0.CHAPTER 9. σ = 0. µc = 0.27) (dotted line) compared with fully numerical value (solid line).78 0. however for the British put option we have the natural restriction that µc > r and so the parabolic regime is the only one of interest. i.002 0.795 b(t) 0.006 0. it was with these very traders .785 0. (9. ˆ or D < r.25) has removed the parameter completely from the system. Any speculator (as opposed to a hedger) who chooses to invest in a putstyle option will inevitably hold the view that the underlying will experience downward price movements at some time in the future. Indeed.e.
given the current stock price S0 .b). Furthermore. . To further illustrate possible situations in which a British option might be favourable over other options available in the market. For example. t) × 100%. option price paid at t = 0 The idea of this section is to produce a return surface for all the available options in the market. whilst maintaining a protection feature should these price movements not transpire. let us construct an idealised market consisting solely of a European. say one year and strike price K which for simplicity we shall assume is equal to the current stock price and scaled to unity. American and British put option written on the same underlying. as noted by Peskir and Samee (2008a. THE BRITISH OPTION 187 in mind that the British option was developed. The option price of all three options is easily calculated at t = 0 given the current price S0 .e. i. however any rational investor would never choose to exercise the option ‘out of the money’. the British option provides an instrument that can be used to ‘milk’ proﬁts from the speculators view of future price movements.CHAPTER 9. but there is some ambiguity in the value of the money received at a future time. For a European option we are unable to exercise the option until expiry and so there can be no payout of the option at any time prior to maturity. Alternatively for an American option we can choose to exercise at any time prior to maturity and receive the payoﬀ (K − St )+ . Of great interest from the point of view of such an investor is the expected return on their investment for all possible future stock prices at any point in time prior to maturity. then which available option will provide the greatest return on the price paid for the option. t) = money received at (S. deﬁned as R(S. with the same expiry date. if an investor believes that the stock price will fall by approximately 20% in the next 6 to 9 months. Now an investor in such a market has the choice of investing in either of the three options and as such has to pay the corresponding option price. S0 = K = 1 and hence all options are ‘at the money’. The rational strategy of such an investor would be to choose the option that has the greatest return under the future price movements that are most inline with the sentiment of that particular investor.
VB (S0 . but that the initial price paid for the option will have varied. i. t) = max GB (S. despite the option having a zero gain function in such a situation. (K −ST )+ . THE BRITISH OPTION 188 i. it still has some intrinsic value. t). t) × 100%.e.CHAPTER 9. 0) > VB (S0 . if St > K for any time before maturity. 0) > VE (S0 . since if the investor is exposed to market frictions then they will inevitably obtain a worse price than the noarbitrage values used in the above calculations. t) = VE (S. We note that at maturity the return on all three investments are not equal. t) × 100%. However. 0) hence the European option investor’s only choice is to sell the option at times prior to maturity. . 0). VE (S0 . VA (S. Note that this assumption assumes that the investor has access to the market in order to sell and that he incurs no transaction or liquidation costs. VB (S. The British put on the other hand will not incur such costs (in the exercise region) as there is no need for the investor to enter the market. For this reason the most consistent measure of the expected return on an American option is given by RA (S. more speciﬁcally we must have VA (S0 . 0) hence we are allowing the investor to sell the option in the market and receive the current ‘noarbitrage’ value of the option in addition to allowing for the exercise of the option upon which the payoﬀ is received. the money received will be the same. 0) and the return on the European option as RE (S. In some sense this corresponds to a ‘best case’ scenario for the American and European option. Similar assumptions allow us to deﬁne the return on the British option as RB (S. VA (S0 .e. t) = max (K − S)+ . Also note that for these purposes we are assuming the investor to behave rationally in the sense that the option should be exercised at a stock price in the rational exercise region and to sell the option at a stock price in the continuation region. t) × 100%. precisely the noarbitrage price of the American option at that particular price and time.
then they have ‘wasted’ the earlyexercise premium priced into the American and British option prices. r = 0. For comparison the rational exercise boundaries can be found in ﬁgure 9.2 0.3 0. whereas in some regions the British put option can provide up to 60% µc = 0.4 0. σ = 0.6 is that the British put option appears to be providing a greater expected return in the majority of the stopping regions.8 0. This surface indicates the diﬀerences in the returns (as a percentage of the initial investment) should the investor close out their position by exercising or selling.8 compares the American and European put option.5.4.1.75 0.95 0. Even then the diﬀerence in the American put return is no greater than 20%. and D = 0. The most striking feature of ﬁgure 9. 9. 1 PSfrag replacements b(t) Figure 9.5: Location of the free boundary for the British (solid line) and American (dotted line) put option under investigation in ﬁgures 9.6 0.CHAPTER 9. THE BRITISH OPTION 189 This is natural since if an investor waits until maturity to exercise an earlyexercise option.6. i. K = 1.7 0.7 compares the return of the British put to the European put and ﬁnally for reference ﬁgure 9. t).135 0.125.125 0.125. T = 1.8 0. The American option only provides the investor with a better return on their investment if it transpires that the investor chooses to stop and close out their position in a ‘wedge’ shaped region below the current stock price and extending just beyond half way to maturity. µc = 0.5 t 0.9 1 .65 0 0. RB (S. t) − RA (S.85 µc = 0.7 0. whichever provides the greatest return (or is permitted). Similarly ﬁgure 9.e. for a contract drift of µc = 0.7 and 9.8.9 0.6 shows the diﬀerence in returns of the British put option and the American put option.1 0. Figure 9.
CHAPTER 9. THE BRITISH OPTION
Figure 9.6: The diﬀerence in the percentage return of the British put option and the American put option at every possible stopping location. The solid lines denote contours at increments of 10% from 10% to 60%. The dotted line represents the zero contour. S0 = 1, T = 1, K = 1, σ = 0.4, r = 0.1, D = 0, and µc = 0.125. PSfrag replacements
Figure 9.7: The diﬀerence in the percentage return of the British put option and the European put option. Again the solid lines denote contours at increments of 10% from 0% to 70%. The dotted line represents the zero contour. S0 = 1, T = 1, K = 1, σ = 0.4, r = 0.1, D = 0, and µc = 0.125.
PSfrag replacements
190
80 60 40 20 0 20 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 1.4 1.6 0
t
0
0.2
0.4
0.6
0.8
1
1.2
S
80 60 40 20 0 20 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 1.4 1.6 0
t
0
0.2
0.4
0.6
0.8
1
1.2
S
CHAPTER 9. THE BRITISH OPTION
Figure 9.8: The diﬀerence in the percentage return of the American put option and the European put option. The solid lines denote contours at increments of 10% from 70% to 30%. The dotted line represents the zero contour. S0 = 1, T = 1, K = 1, σ = 0.4, r = 0.1, and D = 0. Note the change of orientation. extra return over the American put option (although these regions are many standard deviations away from the current price of the underlying). More importantly, numerical investigations have shown that the return proﬁle seen in ﬁgure 9.6 and indeed the region in which the expected return on the American is greater than the British put option remains relatively constant in shape and size for varying values of the contract drift µc . Note that the above observations are in total agreement with Peskir and Samee (2008a,b). When comparing the British with the corresponding European put option (ﬁgure 9.7) the return diﬀerentials are the greatest for low values of the stock price and large times to expiry, as we might expect. However it can be seen that rather surprisingly the British put option will provide a greater return for the vast majority of stopping locations below strike K; with the only exception being relatively close to maturity where the European option would provide (at most) a 10% greater return. Figure 9.9 attempts to encapsulate all of the above observations into schematic form. It plots the regions in which the ordering of the expected returns (given any stopping
rag replacements
191
40 20 0 20 40 60 80
1.6
1.4
1.2
1
0.8
0.6
S
0.4
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 0.2 0 1
t
CHAPTER 9. THE BRITISH OPTION
rag replacements
192
1.4
1.2
RE > R B > R A
1
S
0.8
0.6
RA > R B > R E RB > R A > R E
0 0.1 0.2 0.3 0.4 0.5
0.4
0.2
RB > R E > R A t
0.6 0.7 0.8 0.9 1
0
Figure 9.9: Schematic representation of the regions in which atthemoney European, American and British put option would provide the greatest return on an investment. The dotted lines represent the free boundaries of the American and British put option for reference. T = 1, K = 1, σ = 0.4, r = 0.1 and D = 0. location) for the British, American and European put option are diﬀerent.7 The freeboundary location of the British and American put have been included for reference. It appears that there is a region, centred around the American earlyexercise boundary in which the American option would provide a greater return on the initial investment than the other two options. This region gets smaller as maturity approaches and disappears at approximately threeﬁfths of the way to maturity. This also highlights what could be described as an ‘unexpected’ additional beneﬁt of the British option, which is somewhat counter to the protection feature for which the option was originally created. If the stock price falls sharply, then the British option is able to ‘milk’ more money out of such a situation than the holder of the corresponding American option. In fact any stopping location below strike in the second half of the contract will inevitably result in a higher return for the British option holder, over the American option holder. Considering this fact, in conjunction with the the empirical observation that the majority of American options are exercised in the second half of
7 The three curves correspond to the zero contours of the three ﬁgures 9.6, 9.7, and 9.8 obtained for the parameter value µc = 0.125, although the same qualitative behaviour is seen for other values.
CHAPTER 9. THE BRITISH OPTION
193
the contract’s term,8 the British option can be considered an attractive alternative. In addition, if the options are far out of the money, i.e. St K then the European
option would provide the greatest return, however the values of all three options are so small in this region that any comparison becomes a purely academic issue. Finally, ﬁgure 9.10 shows how the location of the wedgeshaped region (in which the American put option provides greater returns over the British put option) changes as the initial value of the stock price is varied, whilst keeping the strike constant. In other words as we increase the moneyness of the option. Figure 9.10(a) shows that as we increase the moneyness by decreasing the ratio S0 /K from 1 to 0.7 in steps of 0.1 this region becomes progressively smaller until it actually disappears for a ratio of 0.6, at which point the British option is guaranteed to provide a greater return on an investment. Figure 9.10(b) shows the shape of this region when we continue to increase the moneyness by changing the ratio from 0.5 down to 0.2. In this case the region grows steadily, however this time any immediate downwards stock price movements will result in the British option providing the greatest return, whereas an upward movement would be better served by an American put option. This is the opposite of the behaviour seen in ﬁgure 9.10(a). It should be noted that for the majority of the values of S0 /K presented in ﬁgure 9.10 the option holder is placed immediately in the exercise region, and so should ‘optimally’ exercise. However, the comparisons are still of interest, since the investor need not follow the optimally strategy.
9.7
The British call option
We now turn our attention to the British call option (cf. Peskir and Samee, 2008a). Unlike the American call option (without dividends) the British call option (with or without dividends) no longer has a trivial solution; there exists an earlyexercise
See for example Diz and Finucane (1993), who investigate the earlyexercise behaviour of options on the the S&P 100 index. They show that over 82% of all call options and 77% of all put options that are exercised are done so during the ﬁnal week before maturity (inclusive of maturity).
8
µc ) − 1 − Se(µc −D)(T −t) Φ d2 (S. .4 0.1 and D = 0.4 0 0. .8 0.4 1. i.9 Again we can note that the gain function approaches the standard call option payoﬀ as t → T . T = 1.29) where all of the notation is as previously deﬁned. VE = Se−D(T −t) − Ke−r(T −t) + VE . t.8 and 0.8 S t 0.6 0.10: Figures representing the region in which American put options would provide a greater expected return that its British option counterpart. t) = GS (S.2 and in fact leads to the following freeboundaryproblem representation of the British call option price: V + 1 σ 2 S 2 V + (r − D)SV − rV = 0 for S ∈ (0.1.2.5.8 1 (a) S0 K = 1.2 Increasing Moneyness 1.28) where now the gain function is given by (9. t) = K Φ d1 (S. .9.8 1 = 0. THE BRITISH OPTION rag replacements rag replacements 194 1. b(t)]. The noarbitrage pricing procedure is identical to that of the British put described in section 9.4 0. t) on S = b(t) (smooth ﬁt) V > G in C V = G in D GC (S.4. (b) S0 K t 0. t) where GC and GP denote the gain functions of the British call and put respectively.. µc ) − 1 (9. 9 G(S.2 1 1 S 0. for increasing moneyness. r = 0. σ = 0. K = 1. t.7. t) on S = b(t) VS (S.2 Increasing Moneyness 0 0.2 0. Notice the similarity of this expression with the putcall parity relationship described in section C P 4.4 1.6 0.4 0 0 0. t) = G(S. Note the relationship between the call and put option gain functions GC (S.e. SS S t 2 V (S.6 0. boundary for all µc < r. Figure 9. 0. i. 0. t) = Se(µc −D)(T −t) − K + GP (S. T ) = (S − K)+ . 0.2 0.2 0.e. .6 0.CHAPTER 9.
11: The British call option free boundary for varying values of the contract drift. 2. r = 0. .2 σ increasing b(t) 1.5 t 0.8 0.5.11 shows the variation of the free boundary for varying values of the contract drift µc showing again a monotonic collapse of the free boundary as µc is increased to r.CHAPTER 9. is no longer monotonic.9 0. Again note that the free boundary of the British call option. Figure 9. 1. . .1. again showing that the nature of the free boundary is highly dependent the volatility of the underlying stock.6 0.1 1 0.8 b(t) 1.9 1 Figure 9.9 1 Figure 9.4 0.8 0 0.7 0.12 also shows its variation with the volatility parameter σ.8 0.2 0. . THE BRITISH OPTION 195 Figure 9. and µc = 0. 0. T = 1.2 PSfrag replacements 1 0. like the British put.12: The British call option free boundary for varying volatilities.4 0.3 0.08. T = 1. 0. µc = 0.06.4 1.1 0.09..1.4.4 1.1. . σ = 0. D = 0. 0.4 PSfrag replacements 2. K = 1. . .8 0 0.1 0.05.6 µc increasing 1. and σ = 0.2 0.3 1.3 0.6 0. K = 1.055..2 2 1.05. 0. r = 0. D = 0.7 0.5 t 0. 0.
the exponent of the exponential term 2 . If ν± > 0 then both the cumulative normal distributions ¯ tend to zero in this limit and the leading order terms become ¯ ¯ µc Ae(µc −D+β)(T −t) = rK + . t) = µc Se(µc −D)(T −t) [Φ(d2 ) − 1] − rK [Φ(d1 ) − 1] .1 Analysis far from expiry For the British call we can apply the same analysis as for the British put in order to determine the large time from expiry behaviour of the free boundary. t) − 1 − rK Φ d1 (Sh .CHAPTER 9.e. t) − 1 = 0.7. 2 2 2 2 ¯ and since in this regime β > 1 σ 2 − µc + D. hence our ansatz for the form of the solution does not permit β > 1 σ 2 − µc + D. t) = 0. ν where we have deﬁned ν± = ¯ 1 ¯ µc − D ± 2 σ 2 + β σ (9. For the British call the Hfunction can be shown to be H(S. THE BRITISH OPTION 196 9. i. (9.31) for convenience of algebra. Again making the ansatz that the solution for Sh (t) behaves as ¯ ¯ Sh (t) = Aeβ(T −t) ¯ ¯ as t → T we wish to solve for the constants A and β. Substitution into the equation (9. . hence µc Sh e(µc −D)(T −t) Φ d2 (Sh . when ¯ ¯ 1 1 ¯ 1 ¯ 1 − σ 2 − µc + D < β < σ 2 − µc + D ⇒ − σ 2 < β < σ 2 . indicating that there can be no solution to this equation for β in the ¯ limit. 2 The second regime corresponds to a situation when ν+ > 0 and ν− < 0. . . We wish to look at the asymptotic behaviour as T − t → ∞ and it transpires that there are three diﬀerent regimes in which the solution is quantitatively diﬀerent. ¯ must be positive.30) and hence we are interested in the large T −t behaviour of the solution to H(Sh .30) yields √ √ ¯ ¯ ν µc Ae(µc −D+β)(T −t) Φ −¯+ T − t − 1 = rK Φ −¯− T − t − 1 .
if D = 0 then the function Sh (t) will tend to inﬁnity (and so must the free boundary) provided that the dividend is greater than some critical value.e. THE BRITISH OPTION 197 in this regime the ﬁrst cumulative normal distribution function tends to zero but the second tends to unity in the limit.CHAPTER 9. that ν± < 0. . Finally we have that ν± < 0. On the other hand. i.31) under the assumption that ν± < 0 is identical to ¯ (9. i.. ¯ An interesting corollary to this result is that when the dividend yield D is zero. 1 D > Dcrit = µc + σ 2 2 r + µc r − µc . we ¯ have that β < 0 for all possible values of µc . and so in this regime we cannot say anything about the freeboundary behaviour for large times to maturity. µc − r 1 ¯ which conﬁrms our original assumption β < − 2 σ 2 − µc + D. i.19) and so it is is clear that a balancing of terms does exist for the call option in ¯ the limit T − t → ∞ and moreover that β will be given by ¯ 1 β = σ2 2 µc + r µc − r − µc + D. Recall that for a British call we must have that µc < r and all parameters must be positive so we have that µc + r < −1. 2π(T − t) ¯ solution for β. . since for the British call option the Hfunction is positive below Sh (t).e. the same behaviour as for the British put.. hence the leadingorder terms become ¯ rKe− 2 ν− (T −t) ¯ ¯ µc Ae(µc −D+β)(T −t) = + . ν− 2π(T − t) ¯ 1 2 where we have used the approximations for Φ(·) introduced earlier. in this regime the leadingorder terms ¯ which is identical to equation (9.. hence the function Sh (t) for the British call must always decay to zero for large times to expiry in this regime. . However this is not helpful from the viewpoint of determining the behaviour of the free boundary in this limit.17).e. . Furthermore the large T − t behaviour of the time derivative of equation (9. and so there can be no become ¯ ¯ µc Ae(µc −D+β)(T −t) − 1 ν+ (T −t) 2 e 2¯ = ν+ 2π(T − t) ¯ ν− ¯ 1 2 rK ¯ e− 2 ν− (T −t) + . Again it is clear that there can be no balancing in this regime as T − t → ∞.
22). (2002) state that the American call option with dividends behaves as b(t) ∼ rK 1 + σα0 ˆ D 2(T − t) . where α0 is as before.7. The best we have done is to use the behaviour of the Hfunction in this limit as an analytical proxy of the free boundary and infer that the free boundary must tend to zero in certain parameter regimes.8. Jacka (1991) and Carr et al. THE BRITISH OPTION 198 If the dividend is greater than this value. Using the same transformation we can transform the British call option to the standard American call option problem with dividends (in the limit t → T ). the function Sh (t) can show positive exponential growth for large times to expiry. c 9.32) when compared with the fully numerical free boundary.CHAPTER 9.1 Integral representations of the free boundary The American put option The value of an American option can be written using the socalled earlyexercise premium representation.13 once again shows the accurately of the approximation (9. So far we have been unable to determine the large time to expiry behaviour of the British option free boundary directly.32) Figure 9. more speciﬁcally if the contract drift is less than the critical value µ∗ given by (9. the following considerations may prove to be useful. due to Kim (1990).2 Analysis close to expiry The close to expiry analysis for the British call option is identical to that of the British put option. Again the results of Evans et al. 9. Hence the British call free boundary behaves as b(t) ∼ rK 1 + σα0 µc 2(T − t) e−(µc −D)(T −t) . (1992) . (9. As a step to determining the asymptotic behaviour of the free boundary itself.8 9.
t (9.27 1.13: The asymptotic approximation for the British call option free boundary close to expiry. An explicit expression for the probability in the integral is well known and furthermore is derived in appendix C and using this expression (with D = 0 for simplicity) reduces the above representation to T −t 0 V (S.08 and D = 0.26 1.004 t 0.25 0. r = 0.34) at S = b(t) for which e−ru Φ log b(t+u) S − r − 1 σ2 u 2 du. for a full exposition of this representation (including existence and uniqueness results) see Peskir and Shiryaev (2006). µc = 0.01.008 0. T ) + rK 0 T −t e−ru PQ [St+u ≤ b(t + u)] du.e. (9.e. S. To determine the location of the free boundary we evaluate equation (9.285 1. the second term can be seen intuitively as the extra value of the option due to the ability to exercise early. G(St .28 1. i. amongst others. t) = VE (S.265 1.t free boundary at time t + u (conditional on the information available up to time t). This representation is given by V (S. σ = 0. K = 1. t) + rK Note that in order to evaluate the integral above and hence determine the option value we require knowledge of the location of the free boundary b(t).255 1.CHAPTER 9. and PQ [St+u ≤ b(t + u)] is the probability that the process is below the S. i. T = 0.32) (dotted line) compared with fully numerical value (solid line). t) = (K − St )+ for a put. t) = EQ S.34) .33) where G(ST .006 0. T ) is the gain function of the American option.002 0. Identifying the ﬁrst term in the equation as the European put value (without any early exercise). √ σ u (9.1.t e −r(T −t) G(ST .4. THE BRITISH OPTION 1.01 Figure 9.275 rag replacements 0 199 b(t) 1.
The integral representation thus reduces to K − b(t) = rK ∞ 0 e −ru Φ √ This integral can be solved (with a little work) by ﬁrstly setting s = k1 u where k1 = σ 2 −2r 2σ √ (σ 2 − 2r) u 2σ du. and further integrating by parts yields 2rK K − b(t) = 2 k1 2 2 k1 k1 + √ 4r 2r 2π ∞ 0 e − 1 + r 2 2 k1 s2 ds . Firstly we note that in the limit T − t → ∞ the value of a European option trivially tends to zero (due to discounting). since if we are at time t then the free boundary at any time in the future will be the same as it is now. Therefore we would expect the ratio b(t + u)/b(t) to be equal to one. THE BRITISH OPTION 200 we know the value of the option (by deﬁnition) must be equal to K − b(t). . leading to 2rK K − b(t) = 2 k1 ∞ 0 se − rs 2 k1 2 Φ(s)ds. Hence lim log b(t + u) b(t) =0 T −t→∞ This is eﬀectively the same as making the assumption (or ansatz) that the free boundary to be found is a constant. Setting VE (b(t). To illustrate that this equation does indeed lead to the location of the free boundary. This leads to the socalled freeboundary equation which completely characterises the free boundary K − b(t) = VE b(t). however it can be reduced to an explicit equation by exploiting the fact that the American free boundary at large times to expiry tends to a constant value.35) √ σ u (implicit) equation for b(t) will give the location of the free boundary.35) in the limit T − t → ∞. hence we are interested in the behaviour of equation (9.CHAPTER 9. b(t) = b∞ say. t) = 0 still leaves an implicit equation for b(t). ∞). (9. t + rK T −t 0 Note that this is a Volterra integral equation (of the second type) and solving this e−ru Φ log b(t+u) b(t) − r − 1 σ2 u 2 du. we shall consider the simple case of a perpetual American put option. for all values of u ∈ [0.
t + s) = V (S.1.CHAPTER 9. ∂u 2 ∂S ∂S Applying the operator to the discounted process we can simplify the above expression to s e−rs V (St+s . t + s) = V (S. 2 2r + σ α+1 This agrees exactly with the well known value of the perpetual American put free boundary. where Mt is a local martingale. LSt is the inﬁnitesimal generator of the process deﬁned by L St = ∂ 1 ∂2 ∂ + σ 2 S 2 2 + (r − D)S . found in section 6. 2008a.3). = V (S. t) + 0 LSt e−ru V (St+u .8. Peskir and Samee. t) + 0 s e−ru (LSt − r) V (St+u . THE BRITISH OPTION 1 2 r 2 k1 201 Finally setting k2 = + and using the identity ∞ 0 e−k2 s ds = 2 1 2 π . In order to show this we can apply Itˆ’s formula to the discounted option value to obtain o s e−rs V (St+s . equation (6. σ2 2rK αK = . In terms of the original parameters we can see that k2 = 1 2 σ 2 + 2r σ 2 − 2r .b). t + u)du + Mt e−ru L V (St+u . t) + 0 . 9. t + u) du + Mt . k2 1 2k2 2 we arrive at b(t) = K 2 1− √ . t + u) du + Mt .2 The British put option Analogous with the American option. and so substitution gives the location of the free boundary as b(t) = where α = 2r . the optimal stopping boundary of the British option can be characterised as the unique solution of a nonlinear Volterra integral equation of the second type (cf.
t + u) = √ σz 2πu z S − r − D − 1 σ2 u 2 2σ 2 u 2 . t + u). t + u) du. THE BRITISH OPTION 202 where we have used the fact that (LSt − r) V (S. t + u)du (9. t) − 10 T −t 0 e−ru L(S. t) = J(S. t) where L is the BlackScholes diﬀerential operator. I(·) is the indicator function and b(·) denotes the location of the free boundary separating the continuation and stopping regions. where f (S. t. T ) − =E Q T −t 0 T −t 0 e−ru EQ L G(St+u . du e −r(T −t) G(ST . z. t) = EQ e−r(T −t) G(ST . This modiﬁes the expression to V (S. where we have taken the expectation under the integral sign and used the martingale property that E [Mt ] = 0. Hence we have the following representation of the British put option value. t + u) = G(St+u . Finally letting s = T − t and rearranging yields an expression for the option value V (S.36) For a derivation see appendix C. t + u) = 0 in the continuation region and is only nonzero in the stopping region where we have V (St+u . t. b(t + u). t + u)I St+u ≤ b(t + u) du. t + u)I z ≤ b(t + u) f (S. t) = EQ e−r(T −t) V (ST . t) = L V (S. t + u)dz. t + u) I St+u ≤ b(t + u) e−ru EQ H(St+u . t + u)I z ≤ b(t + u) = 0 ∞ H(z. T ) and also we have that L V (St+u . . t) + 0 e−ru EQ L V (St+u . t + s) = V (S. t + u) is the transitional probability density function of the process started at position S at time t and ﬁnishing at position z at time t + u given by10 log 1 exp − f (S.CHAPTER 9. V (S. t. Now by deﬁnition V (ST . z. The next step is to take expectations giving s EQ e−rs V (St+s . t. z. T ) − where we have used the deﬁnition of the Hfunction. t + u) du. T ) = G(ST . We now consider the expectation under the integral sign which can be expressed as EQ H(z. T ) − T −t 0 e−ru EQ L V (St+u .
At this stage we can see that using the gain function of the standard American put option. t. immediately follows. t). b(t + u).33). GA (S.t = EQ e−r(T −t) (K − S)+ .11).e. b(t+u) = 0 with H(S. 11 . t) = G(b(t). t). i. i.37) in order to determine the large T −t behaviour of the free boundary.t P = VE (S. much in the same way as we did for the function Sh (t). b(t + u). hence P G(b(t). t + u)dz. H(z.b). (9. T ) . t + u)I z ≤ b(t + u) f (S. As such this shall be left as the topic of future research. t) = (K − S)+ will yield H(z. THE BRITISH OPTION 203 where we have deﬁned J(S. The equations involved are clearly much more complicated and so it will not be a trivial matter to extract such asymptotic behaviour. t) − T −t 0 e−ru L(b(t). z. t. t) as deﬁned by equation (9. t) = EQ e−r(T −t) G(ST .36) at the free boundary S = b(t) where we know that V (b(t). S.e. t + u) = −rK and the earlyexercise premium representation for the American put option. t + u)du. t.11 Now to determine the free boundary we can evaluate equation (9. (9. S. the corresponding European put option value and L(S. z. t + u) = 0 ∞ H(z. t + u)f (S. t) = VE (b(t). We can attempt to utilise the nonlinear integral equation (9.CHAPTER 9.37) For the proof of the uniqueness of the above representation see Peskir and Samee (2008a. t + u)dz. t.
It is concluded that there is insuﬃcient ﬁnancial modelling to describe the true price dynamics in such situations. However. involving the option gamma. this analysis also gives guidance on how to tackle these problems numerically at times away from expiry (the full problem). It is clear that the period close to expiry is the most critical for optionpricing models 204 . the powerful tool of asymptotic analysis has been used to extract important information about the behaviour of such models close to expiry. A feature common to a number of these models is that the overall dispersion term. Here. This is clearly a somewhat undesirable feature. invariably these solution features lead to completely spurious solutions if standard numerical procedures are adopted.Chapter 10 Conclusions In this thesis we have investigated a number of models which have been proposed to incorporate ﬁnite liquidity of the underlying asset into the classical BlackScholesMerton option pricing framework. instead. incorporating the appropriate discontinuities into the numerical scheme. Indeed. which is exacerbated by the possibility of negative option values for puts. we must allow solutions with discontinuous deltas. the vanishing of the denominator in the dispersion term can also be a serious issue. The upshot of this is that models of this general class cannot exhibit fully diﬀerentiable solutions at times prior to expiry. diminishes in magnitude as the gamma increases in magnitude (as indeed it must as standard payoﬀ conditions are approached). Allied to this.
CHAPTER 10. CONCLUSIONS
205
and any model that successfully treats this regime should also successfully replicate the option value dynamics for all time. The approach detailed in this thesis should give guidance for the development of models incorporating ﬁnite liquidity without the undesirable features observed in a number of the existing models. Several models in the past have circumvented these diﬃculties close to expiry but generally using ad hoc, rather than intuitively justiﬁable arguments. The hope is that the analysis presented in this thesis will help in this respect; in addition, below we discuss brieﬂy some preliminary ideas about how this may be achieved in future research. One problem with the modelling framework introduced in chapter 2, from a ﬁnancial viewpoint, is that the change in price dS becomes unbounded when the ‘forcing’ term df becomes unbounded, and for the case in which f = ∆ this will happen when d∆ becomes unbounded. Unfortunately, for option contracts with nonsmooth payoﬀ proﬁles, the unbounded nature of d∆ is unavoidable and so if we are to incorporate these (common) situations into such modelling frameworks then it is suggested that a nonlinear response of df to d∆ may well overcome such diﬃculties. We could choose to incorporate such nonlinearity into our deﬁnition of the forcing term f . i.e. instead of setting f = ∆ we could set the forcing term to be some function of ∆, i.e. f = g(∆). However it is not the function f which ultimately aﬀects the price, but rather its inﬁnitesimal change df , hence it is the term λdf which we require to remain bounded (irrespective of the trading strategy ∆). Furthermore, when considering option pricing, the only term in df that ﬁlters through into the option price is the therefore it is desirable to bound
∂f ∂S ∂f dS ∂S
term,
rather than f . This can be done if we model
the derivative of the forcing term (f ) with respect to the asset price as a bounded (nonlinear) function of the derivative of the trading strategy (∆), i.e. ∂f =g ∂S ∂∆ ∂S (10.1)
where g(·) is bounded above. An example of such a function is the ratio of two
CHAPTER 10. CONCLUSIONS nth order polynomials such as g(x) = αn xn + αn−1 xn−1 + . . . + α0 , βn xn + βn−1 xn−1 + . . . + β0
206
which in the limit x → ∞ is bounded above by the ratio αn /βn . If the aim is to prevent the denominator from vanishing then we require this ratio to be 1/λ or below, hence the function g(x) = x λx + β
would suﬃce, where β can be chosen to obtain the desired shape of the response curve. Note also that since the response function g(x) is concave, this model is consistent with the empirical evidence of (nonlinear) price impact discussed in section 8.2. The functional form of this dependence, however, has been chosen arbitrarily and so again this extension to the model can be seen as merely an ‘ad hoc’ ﬁx to the diﬃculties associated with the vanishing of the denominator. It is possible, however, that the ideas presented above could be suﬃciently formalised with further research. Another area of future research could be to exploit the links of the existing models with the theory of linear and nonlinear elasticity (of solids).1 Indeed, in some sense, the models formulated in chapter 2 are analogous to Hooke’s law in linear elasticity, the more we push the market the more it will move, and in a linear fashion. It may be that this force/extension relationship can be approximated as linear, however it is likely that this will only be the case provided the force remains within reasonable limits. Hence, just as current models are analogous to Hooke’s law for elastic media, there may also be an analogy to the ‘elastic limit’ of a material, i.e. the elastic limit of the market, beyond which the market will respond in a nonlinear fashion. This limit point might correspond to the current market depth, or a point much further away. However, ultimately the aﬀect on the price must be bounded, since there are only a ﬁnite number of shares (and hence a ﬁnite force) available. It is hoped that if the above considerations are incorporated into the current modelling framework then the problems associated with the vanishing of the denominator may be overcome.
This is touched upon in the paper by Sch¨nbucher and Wilmott (2000), in which it was suggested o that the problems of the vanishing denominator may disappear if some ‘elasticity’ is incorporated into the response of the market price to large trades.
1
CHAPTER 10. CONCLUSIONS
207
Finally, it may be possible to preclude the denominator from vanishing if we utilise the techniques outlined in Soner and Touzi (2007), who extended the ideas originally introduced in Broadie et al. (1998). These works deal with cases in which the gamma of the replicating portfolio is bounded above (or below) by some trading constraint. In such situations, not all options can be perfectly replicated due to the inability of the replicating portfolio to replicate the option value in regions of large gamma. However, the minimal superreplicating price can be deﬁned as the cheapest replicating strategy that dominates the option value in the entire domain. For the classical BlackScholesMerton framework it can be shown that such a minimal superreplicating price corresponds to the perfectreplicating price (i.e. with no constraints) of the same option, but with a suitably ‘facelifted’ payoﬀ proﬁle. Such a facelifted payoﬀ proﬁle corresponds ostensibly to a suﬃciently smoothed payoﬀ proﬁle. It is thought that applying the constraint VSS < 1/λ to the nonlinear PDE (4.1) may help to regularise the solutions. However, it is not obvious that these techniques can be extended to the illiquid situation, since the results rely on the stochastic representation of the option value (cf. equation (1.4)), a representation that does not exist for the fully nonlinear equation (4.1). In addition, the superreplicating price is only one possible paradigm for option pricing in incomplete markets, and so it is not clear cut that this is the paradigm to use. Furthermore, it is a generally held consensus that the premium paid to superreplicate (i.e. remain entirely riskfree) is, in practise, too high. If it transpires that the problems associated with the vanishing of the denominator cannot be remedied by the above suggestions, or by some other means, then it is asserted that, of the three models that were shown to admit wellposed solutions close to expiry, the Bakstein and Howison (2003) model is one that would be the most desirable alternative model; the reason for this is twofold. Firstly, whilst remaining wellposed close to expiry the option price behaviour also remains suﬃciently diﬀerent from that of the corresponding BlackScholes (liquid) option. Secondly, in the limit of no price slippage, this model reduces to the model of Cetin et al. (2004) which has
CHAPTER 10. CONCLUSIONS
208
become a popular model for liquidation costs in recent years. An alternative would be to specify an entirely new framework for incorporating price impact onto option pricing theory. The aims of such a model would be: (i) to ﬁx the problem with the denominator vanishing, resulting in wellposed problems for standard payoﬀ proﬁles, and (if possible) (ii) to incorporate the empirical evidence of nonlinear price impact. Criteria such as consistency with empirical data, ﬂexibility in application and also computational aspects (such as the regimes close to expiry considered in this thesis) would be crucial to the success of such a model. Also in this thesis, and on a related theme, we have investigated the properties of the newly introduced British option;2 a new nonstandard class of early exercise options. Such options help to mediate the eﬀects of a ﬁnitely liquid market since the contract does not require the holder to enter the market and hence incur liquidation costs. Here, once again, the powerful tool of asymptotic analysis, coupled with advanced numerical methods have been used to shed light on the behaviour of the earlyexercise boundary for both large and small times from expiry. Furthermore, a pleasingly simple variable transform was found that helped to reduce the associated freeboundary problem to that of the standard American option (with dividends) in the regime close to expiry. Finally, most researchers in quantitative ﬁnance have an opinion on the direction of future research in the ﬁeld, some more outspoken than others. Wilmott and Rasmussen (2002) hypothesise that future models will move away from the simplicity of traditional stochastic models and their assumptions about probabilistic behaviour. They also suggest that future models will inevitably draw from a wider range of mathematical tools. Lipton (2001) goes further to suggest that future research needs to pay much more attention to the issue of determining the spot price and to predict its short term evolution, in other words, to provide a suﬃciently formal framework in which to study the market microstructure including supply and demand and liquidity
2
See Peskir and Samee (2008a,b).
In addition.wilmott. eminent physicist turned ‘quant’ Emanual Derman states in his blog3 that hopefully future work will aim to narrow the gap between the invisible microstructure of markets and the observable macroscopic properties such as market prices. CONCLUSIONS 209 eﬀects.CHAPTER 10. It is hoped that this thesis is at least a step in that direction.com/blogs/eman/. 3 See http://www. . a gap which at present is particularly large.
Hauptmann. M. D. July. 2:73–88. 3:5–40. H.Bibliography Abramowitz. Almgren.. Avellaneda. A marketinduced mechanism for stock pinning. Avellaneda. Journal of Risk. (1986). A. E. Pricing and hedging derivative securities in markets with uncertain volatilities. 17:223–249. and Chriss. (2001).. (2003). Dover. and Li. 3:417–425. R. pages 57–62.. and Howison. C.. Avellaneda. 210 . G. Levy. (1995). working paper. R. Asset pricing and the bidask spread. I. and Stegun. (2003). Bakstein. (2007). E. and Lipkin. working paper. M. H. Quantitative Finance. (2005). Equity market impact. and Lipkin. Risk Magazine. M. M. S. (1968). Amihud. working paper. (2001). M. Applied Mathematical Finance. Thum. Handbook of Mathematical Functions. New York. and Mendelson. and Andergassen. R. Kasyan. N. A. and Paras. Agliardi. Powerlaw priceimpact models and stock pinning near option expiration dates.. Almgren. Y. M... Journal of Financial Economics. Feedback eﬀects of dynamic hedging strategies in the presence of transaction costs. Optimal execution of portfolio transactions. A nonarbitrage liquidity model with observable parameters.
Option pricing with transaction costs and a nonlinear BlackScholes equation. Cambridge Texts in Applied Mathematics. Hedging and portfolio optimization in ﬁnancial markets with a large trader.. (1990). Bj¨rk. H. Pricing and hedging in illiquid ﬁnancial markets. A. A. P. and Scholes. 45:673–689. Finance and Stochastics. and Soner. Presentation at the 4th world congress of the Bachelier ﬁnance society. F. Basel. D. (1973). Z. I. M. B. Bick. (1987). and Wiener. (2004). Bank for International Settlements (1999). Basel (March). A discussion note by the Committee on the Global Financial System. G. Bank. Bank for International Settlements (2001). 14. 81:637–659.BIBLIOGRAPHY 211 Bank. . M. Committee on the Global Financial System. Market liquidity: Research ﬁndings and selected policy implications. 22:259–275. and Baum. 14:1–18. General properties of option prices. 2:369–397. 51:1573–1610. On the consistency of the BlackScholes model with a general equilibrium framework.. Journal of Political Economy. (1996). (1998). Z. Black. Structural aspects of market liquidity from a ﬁnancial stability perspective. The Journal of Finance. G. Oxford University Press. o Second Edition. D. Journal of Financial Quantitative Analysis. Journal of Finance. Selfsimilarity. Cambridge University Press. Barles. Bick. and Intermediate Asymptotics. The pricing of options and corporate liabilities. Barenblatt. Mathematical Finance. (2004). (2006). On viable diﬀusion price processes of the market portfolio. Grundy. Arbitrage Theory in Continuous Time. T. Y. Bergman. Scaling. Cambridge. P. (1996).
(2002). A. Y. Portfolio insurance and ﬁnancial market equilibrium. H. Modern Portfolio Theory and Applications. R. U. Bordag. G... E. Journal of Financial Economics. M. 62:455–476. Finance and Stochastics. 2:78–106. Modeling liquidity eﬀects in discrete time. (1998). L. and Kubota. Liquidity risk and arbitrage pricing theory.. (2007). and Potters. Mathematical Finance. M. C. J... 8:311–341. R. Expected return and the bidask spread.. 10:1–21. and Chmakova. N. Chalmers. Carr. working paper. 11:59–79. 1014 April. International Journal of Theoretical and Applied Finance. M´zard. The Review of Financial Studies. Jarrow. (1989). Cetin. arXiv:0708.. M. M. 2:251–256. 17:15–29. and Kadlec. R. K. Statistical properties of stock e order books: Empirical results and models. Gakujutsu Shuppan Center. . and Myneni.BIBLIOGRAPHY 212 Bordag. and Soner. A. Mathematical Finance. (2007).. (1998).. Quantitative Finance. R.1568v1 [math. A. 48:159–188. R.. (1992). A. 2007. Osaka. On the valuation of a fair price in case of an non perfectly liquid market. editors. P. P. Cetin. G. and Kan. J. Sawaki. (2007). Bouchaud. L. Explicit solutions for a nonlinear model of ﬁnancial derivatives.AP]. K. (2007). J. and Schwartz. P. and Rogers. and Frey. A. M. Journal of Business. Brennan. Cvitani´. An empirical examination of the amortized spread. B. (1995). L. M. Lisbon. Nonlinear option pricing models for illiquid markets: Scaling properties and explicit solutions. Jarrow. and Protter. Optimal replication of contingent c claims under portfolio constraints. U. Proceedings of the Workshop on Mathematical Control Theory and Finance. (2004). Alternative characterizations of American put options. Bordag. Chen. L. Broadie. S.. In Saitou.
Cole. and Etheridge. Datar. and Yang. T. P. 6:19–26. J. V. (1998). J. 6:370–398.. Perturbation Methods in Applied Mathematics. The Review of Financial Studies. M. Optimal consumption choices for a ‘large’ investor. and Finucane. R. Chadam. Convexity of the exercise boundary of the american put option on a zero dividend asset. (1980). 18:185–197. and Cvitanic. X. A.. (1998). J. and Ma. 6:765–797. 10:220–231. J. Duﬃe. A family of embedded RungeKutta formulae. P. Dormand. L. (1996). and Zheng. Princeton University Press. (2006).. Newton. J. F. (1996). J. D. C. (1993). P. Duck. W. Princeton University Press.. Mathematical Finance. Naik. The Annals of Applied Probability. Math. J. Diz. to appear.. Widdicks. Cvitanic. Blaisdell Publ.. Mechanics Appl. D. Crank. R. (1968). Jiang. Mass. Journal of Economic Dynamics & Control. Cuoco. 22:401–436. Q. and Prince. The rationality of early exercise decisions: Evidence from the S&P100 index options market. Walttham. Louis Bachelier’s Theory of Speculation: The Origins of Modern Finance. 1:203–220. (2008). M. Journal of Financial Maarkets. 2nd Edition. and Radcliﬀe. J. Mathematical Finance. Liquidity and stock returns: An alternative test. Dynamic Asset Pricing Theory.. N.. Two methods for the numerical solution of movingboundary problems in diﬀusion and heat ﬂow.BIBLIOGRAPHY 213 Chen. Singular perturbation techniques applied to multiasset option pricing. Hedging options for a large investor and forwardbackwards SDEs.. (2008). J. D. Journal of Computational and Applied Mathematics. J.. (1957). Co. . D. W. Davis.
RI. C. Eleswarapu. Superreplication of options on several o underlying assets. 2:115–142.. S. 8:93–126. pages 125–136. M. Journal of Applied Probability. Market illiquidity as a source of model risk in dynamic hedging. (2004). S. and Patie. L. 12:219–237. and Moench. JeanblancPicqu´. Janson. American Mathematical Society. B. N. Mathematical Finance. R. Fama. (2003). Kuske. (1996). V. Robustness of the Black e and Scholes formula. Springer. Providence. and Keller. K.BIBLIOGRAPHY 214 Duﬀy. A critique of the Crank Nicolson scheme strengths and weaknesses for ﬁnancial instrument pricing. (2002). and French. F¨llmer. J. American options on assets with dividends near expiry. (1998). El Karoui. A. Modeling feedback eﬀects with stochastic liquidity. (2000). Model Risk. (2005). and Shreve. D. A microeconomic approach to diﬀusion models o for stock prices. J. pages 137–159. R. Ekstr¨m. (2002). 3:1–23. 34:373–386. Frey. and Tysk. Mathematical Finance. Finance and Stochastics. M. Journal of Financial Economics.. E. (1998). (1998). E. M. Evans. and Reinganum. J.. Wilmott Magazine. H. Frey. Evans. In: K Sandmann and P Sch¨nbucher (Eds. Esser. and Schweizer... B.. (1993). Mathematical Finance. R. Advances in ﬁnance o and stochastics. J. R. Journal of Finance. Frey. working paper. . The seasonal behavior of liquidity premium in asset pricing. 42:27–38. Risk management for derivatives in illiquid markets: A simulation study. Partial Diﬀerential Equations. D. 51:55–84. 4:68–76. Multifactor explanations of asset pricing anomalies. P.). (1993). Perfect option replication for a large trader.
R. Matched asymptotic expansions in ﬁnancial engineering. and Stremme. H. Friedrichs. P. V. 15:313– 316. Journal of Financial Economics. S. Hodges. J. Geske. 11:215–260. R. (1981). and Selby. M. Special Topics in Analysis. Journal of Financial and Quantitative Analysis. 7:351–374. Plerou. HenryLabord´re. and Kreps. M. Management Science. S. E.BIBLIOGRAPHY 215 Frey. (1954). e working paper. R. 33:347–355. Journal of Engineering Mathematics. Mathematical Finance. Market volatility and feedback eﬀects from dynamic hedging. Harrison. S. Martingales and stochastic integrals in the theory of continuous trading. S. S. and Pliska. Stochastic Processes and Their Applications. and Stanley. San Diego. J. and Pliska. On the evaluation of compound options. M.. 20:381–408. Harrison. The valuation of compound options. (1979). The valuation of corporate liabilities as compound options. (1997). 7:63–81. D. . (1983). Nature. P. O. CA: Academic. arXiv:condmat/0404520v1 [condmat. X. Gopikrishnan. Howison. Diﬀerential Equations. and Smale. (2005). and Linear Algebra. W. New York University. Martingales and arbitrage in multiperiod securities markets. Geske. R. M. Stochastic Processes and Their Applications. (1979). Hirsch. Harrison.. 53:385–406. (2003). (1974). D. (2004). (1977). A. M. A stochastic calculus model of continuous trading: Complete markets. 12:541–552. K. R.. Gabaix. J. P. (1987). M.other]. Dynamical Systems. The feedback eﬀect of hedging in portfolio optimization. J. 423:267–270. Journal of Economic Theory. A theory of power law distributions in ﬁnancial market ﬂuctuations.
R. Jarrow. Second Edition. P.. 13:890–913.. (2005). Journal of Applied Mathematics. (2004). Jacklin. (1992). Journal of Financial Quantitative Analysis. D. Iori. and Tysk. IMS Lecture Notes. and Samuel. Jarrow. V. (1992). Market manipulation. S. Discussion paper. Janson. Derivative securities markets. Underestimation of portfolio insurance and the crash of october 1987. London. A. 1:1–14. Kleidon. Jarrow. J. Mathematical Finance. R. 45:75–91. Jeannin. G. C. Improved Numerical Techniques for Occupationtime Derivatives and Other Complex Financial Instruments. Probability Essentials. SpringerVerlag Berlin Heidelberg New York. Springer. (2003). In The Herman Rubin Festschrift. . thesis. 18801970. A. J. and Pﬂeiderer. and Protter. and Protter. P.. Optimal stopping and the American put.. (1991). Jacod. market manipulation and option pricing theory. 29:241–261. R. (2007). New York. A. A short history of stochastic integration and mathematical ﬁnance: The early years. On the riskadjusted pricingmethodologyc c based valuation of vanilla options and explanation of volatility smile. (2004). D. Journal of Financial Quantitative Analysis. P. ˇ e Jandaˇka. Johnson. M. City University. M.D. (2006). D. (2003). S. UK. (1994). A. Ph. Modeling stock pinning. Annals of Applied Probability.BIBLIOGRAPHY 216 Jacka. 5:35–63. Johnson. University of Manchester. S. P. Volatility time and properties of option. bubbles. 3:252–258. Review of Financial Studies. 27:311–336. and Sevˇcoviˇ. Singular Perturbation Theory: Mathematical and Analytical Techniques with Applications to Engineering. R. corners and short squeezes.
pages 17–20. Numerical methods in boundarylayer theory. Optimal exercise boundary for an American put option. 2:29–48. The Review of Financial Studies. D. W. 5:107–116. Fluid Mech. (1978). Lecture Notes in Physics. D. December. I. and Korn. (2001). Oxford University Press. Rev. J. S. Technical Report 496. and Keller. J. 3:547–572. Kreps. and Nelken. A. J. Ann. O. H. and Smith. R. and Lagersrom. B. Journal of Financial Markets. and Keppo. Mech. Math. 495:171 – 205. Keller. S. P. Asymptotic expansions of the NavierStokes solutions for small reynolds numbers. (1990). Kruskal. Market depth and order size. Applied Mathematical Finance. A. SpringerVerlag.BIBLIOGRAPHY 217 Jonsson.. (1997). pages 585–593. A. Methods of Mathematical Finance. Kaplun. New York. (2002). Kuske.. I. Nonlinear Ordinary Diﬀerential Equations. The analytic valuation of American options. D.. (1999). P. and Shreve. Fluid Mechanics and Singular Perturbations. Analytic and asymptotic methods for nonlinear singularity analysis: A review and extensions of tests for the Painleve property. S. Kempf. . N. (1967). I. M. 9:262–272. (1957). Option pricing for large agents. J. (1998). Academic Press. R. Applied Mathematical Finance. H. Graduate School of Business. (1998). 3rd Edition. Karatzas. Krishnan. (1979). Multiperiod securities and the eﬃecient allocation of risk: A comment on the BlackScholes option pricing model. New York.. Jordan. The eﬀect of stock pinning upon option prices. M. Stanford University. Kaplun. Risk Magazine. Joshi. 10:417–433. (1999). and Halburd. Kim.
Springer Verlag. and Ogawa. International Journal of Theoretical and Applied Finance. 24:679–702. Markowitz. N. and Shiryaev. 53:1315– 1335. Solonnikov. (2004). N. RI. Option pricing with an illiquid underlying asset market. J. and Mantegna. Lyukov. 2nd Edition. Liptser. Econometrica. Ladyzenskaja. (1959).. World Scientiﬁc: Singapore. Heat conduction in a melting solid. Proceedings of the 2003 Symposium. World Scientiﬁc. L. A. R.. Journal of Economic Dynamics and Control. N. A. and Yong. Liu. B. (2005). S. E. Single curve collapse of the price impact function for the New York Stock Exchange. (2000). A liquidityaugmented capital asset pricing model. Statistics of Random Processes II: Applications. S. Portfolio Selection: Eﬃcient Diversiﬁcation of Investments. 7:757–768. S.. Continuous auctions and insider trading. American Mathematical Society: Providence. New York. H. Agentbased computational ﬁnance: Suggested reading and early research. .BIBLIOGRAPHY 218 Kyle. A. 8:81–94. 421:129–130. O. G. A. Journal of Economic Dynamics & Control. Mancino. Linear and Quasilinear Equations of Parabolic Type. (1985). M. Nature. (2006). H. Landau. D. Nonlinear feedback eﬀects by hedging strategies. A. (2003). Lillo. V. 29:2125–2156. and Ural’ceva. A. W. In Stochastic Processes and Applications to Mathematical Finance. Journal of Financial Economics. F. (1950). J Wiley and Sons. H. (2001). (1968). (2001). Mathematical Methods for Foreign Exchange. (2003). LeBaron.. 82:631–671. New York. Option pricing with feedback eﬀects. Liu. Ritsumeikan University. R. N. Quarterly of Applied Mathematics.. Lipton. Farmer.
Local timespace calculus and extensions of Itˆ’s formula. Platen.BIBLIOGRAPHY 219 Merton. and Samee. Peskir. A. Generic properties of ordinary diﬀerential equations. Probab. Peskir. Market Microstructure Theory. Ni. M. M. F. R. Research Report No. Optimal Stopping and Freeboundary Problems. Blackwell. A changeofvariable formula with local time on curves. On the American option problem. O’Hara. M. (1998). X. (2008a). Statist. Journal of Financial Economics. (2005b). Manag. and Schweizer. G. D. American Law & Economics Review. (1995). (1997). Probab. Birkhauser. 6:28–71. 15:169– 181. E. Statist. Stock price clustering on option expiration dates. C. Pearson. 18:499–535. Group Manchester. N. Research Report No. MAA Studies of Mathematics. Birkh¨user. G. and Samee. 2. Peixoto. (2008b). Theory of rational option pricing. Bell J. M. M. (1973). Econ. The British call option. Mathematical Finance. A strong maximum principle for parabolic equations. . On feedback eﬀects from hedging derivatives. S. a Peskir. Peskir. and Shiryaev. Detecting manipulation in futures markets: The Ferruzzi soybean episode. 78:49–87. and Poteshman. Peskir. Sci. 6:167–177. 14:52–92. Basel. Group Manchester.. (1953). Mathematical Finance. F. Pirrong. G. Peskir. 8:67–84. G. 55:177–192. C. (2003). Progress o in Probability. G.. The British put option. Journal of Theoretical Probability. L. Communications on Pure and Applied Mathematics. (2006). 4:141–183. 1. A. (2005). (2004). G. (2005a).. Nirenberg.
H. (2002). S. (2002). Uber ﬂ¨ ssigkeitsbewegung bei sehr kleiner reibung. Math.fenews. (2006). M. J. III. P. 1905. G. Springer. Samuelson. ¨ Prandtl. Acta Math.html. u Intern. (2003). Rosenerans. and Stanley.com/fen49/ stochasticsstreet/stochasticsstreet. (1886). Protter. P. Poincar´. SpringerVerlag. Potters. T. New York. and Singh. Corrected reprint of the 1967 original. 32:147– 149. and Weinberger. Application of Mathematics series. Lwipzig. A. 6:13–32. X. Protter. (1904). (1965). and Bouchaud. Maximum Principles in Diﬀerential Equations. More statistical properties of order books and price impact. Fi nancial Engineering News. S. pages 484–491. e e e e e e 8:259–344. Calibration of a nonlinear feedback option pricing model. Sanfelici. Available at http://www. Physical Review E 027104.. Proceedings of the American Mathematical Society. Quantifying stock price response to demand ﬂuctuations. Modelling liquidity and its eﬀects on price. International Monetary Fund working paper WP/02/232. H. (2006). Sarr.P. Heidelberg. Gopikrishnan. C. 324:133–140. . 49. Sur le int´grales irr´guli`res des ´quations lin´aires. V.. Berlin. L.. I. Stochastics on the street: Models of liquidity risk. Measuring liquidity in ﬁnancial markets. Protter. (2007). 7:95–110. Kongr. (1972). Verh. Quantitative Finance. Industrial Management Review. working paper. Derivation of the HopfCole solution to Burgers’ equation by stochastic integrals. L. Stochastic integration and diﬀerential equations: A new approach.. Rogers. H. P. P. E. S. M. and Lybek. 66. Physica A.. (1984). Gabaix.BIBLIOGRAPHY 220 Plerou. Rational theory of warrant pricing. (1990). A. Teubner.
Cambridge University Press. Examination. (2005). R. P.. Oxford. D. S. Smith. Academic Press. D. and Newton. J. Howison. Numerical Solutions of Partial Diﬀerential Equations: Finite Diﬀerence Methods. 10:29–58. M. P. and Papanicolaou. (1975). Whaley. Manipulation on Trail: Economic Analysis and the Hunt Silver Case. The feedback eﬀect of hedging in illiquid o markets. Wilmott.. C. Risk Magazine. thesis. Extension and Creation of Methods for Pricing Options with Early Exercise Features..D. (1982). Applied Mathematical Finance.BIBLIOGRAPHY 221 Sch¨nbucher. E. The BlackScholes equation revisited: Asymptotic expansions and singular perturbation. SIAM Journal of Applied Mathematics. UK. K. . Second Edition. The Heat Equation. D. V. Widder. Academic Press. and Wilmott. General BlackScholes models accounting for increased market volatility from hedging strategies. van Dyke. A. 6:59–66. P. P. Journal of Financial Economics. Oxford University Press. and Wilmott. New York. H. (1964). Duck. Cambridge University Press. M. 5:45–82. (1998). G. Widdicks. G. M. Williams. (1995).. Perturbation Methods in Fluid Dynamics. and Touzi. D. (2002). (1993). Widdicks. Hedging under gamma constraints by optimal stopping and facelifting. P. (1978). E. W. D. University of Manchester. Soner. Sircar. Ph. 15:373–391. Counting the costs.. Mathematical Finance. M. Valuation of American call options on dividendpaying stocks. (2007). P. 17:59–79. The Mathematics of Financial Derivatives. 61:232–272. and Dewynne. (1995). N. A. Whalley. Andricopoulos. New York. (2000). J. Mathematical Finance.
O. . New Directions in Mathematical Finance. H. Wiley. (2002). and Rasmussen.BIBLIOGRAPHY 222 Wilmott. P.
for example. Nirenberg (1953) states (and proves) the maximum principles applied to equations that can be written in the form1 L(V ) + cV = D where L(V ) = n m n m aij VSi Sj + i. this reduces to L(V ) = a11 VSS + b11 Vτ τ + a1 VS + b1 Vτ .β=1 bαβ Vτα τβ + i=1 a i VS i + α=1 b α Vτ α .j=1 α. Evans (1998) and Protter and Weinberger (1984). τ ). such as monotonicity in parameters.Appendix A Maximum Principles Maximum principles are extremely useful tools to investigate the properties of the solutions to partial diﬀerential equations. and hence we can apply the maximum principles to equations of the form AVSS + BVτ τ + aVS + bVτ + cV = D(S.e.1) Also see. 1 (A. 223 . when m = n = 1. If we restrict ourselves to the one dimensional case. uniqueness and convexity. i. These principles date back to as early as 1839 and for a readable overview of their history and a more indepth exposition see Protter and Weinberger (1984).
3) which we now wish to apply the maximum principle to. τ ). Before we state the maximum principle we can formally deﬁne the solution domain as Ω ∈ R+ and assume that it is open. τ ) = D(S. τ ). In the case m = n = 1 this reduces to 2 a11 η1 > 0 ⇒ a11 > 0 ⇒ A > 0 and also 2 b11 ξ1 ≥ 0 ⇒ b11 ≥ 0 ⇒ B ≥ 0.2) Furthermore we make the assumption that c = 0 in the above. Let ΩT = Ω × (0. i. . connected and bounded. Hence we can apply the maximum principle to equations of the form (A.β=1 bαβ ξα ξβ ≥ 0 hence positive semideﬁnite for any real vectors η.1) provided A > 0 and B ≥ 0 with no restriction on the sign of a and b.APPENDIX A. This last simpliﬁcation may seem a little restrictive but it should be noted that it is possible to reduce equation (A. (A. In what follows (for simplicity) we shall assume that B = 0 and b = −1 to obtain a forward parabolic (diﬀusion) equation of the form AVSS + aVS − Vτ + cV = D(S. MAXIMUM PRINCIPLES 224 In order to do this we require the diﬀerential operator L to be elliptic in the Svariables and parabolic in the τ variables. T ]. where T > 0 and also deﬁne ∂ ∗ ΩT = ∂Ω\Ω × {T }. (A. Therefore we require n aij ηi ηj > 0. A > 0.j=1 hence positive deﬁnite and m α.2) (for any c) to one in which c = 0 by making the transformation V = ecτ u to arrive at an equation of the form ˆ L(u) = AuSS + auS − uτ = e−cτ D(S. ξ = 0.
3). τ ). u applying the maximum (minimum) principle to the equation ˆ L(u) = D(S. τ ). Furthermore if D = 0 then both the maximum and minimum must occur on the boundary.APPENDIX A. τ ) ≥ 0 and A > 0 everywhere in the closure of the domain. and alternatively if D ≤ 0 then the minimum must occur on the .1 in the interior of ΩT ) must occur on the boundary ∂ ∗ ΩT . i. ΩT . u u u ˆ ˆ ⇒ AˆSS + aˆS − uτ = L(ˆ) = −D(S. ˆ The maximum principle states that if D(S. sup u = max u = max u.3) are bounded in ΩT . S = 0 and S = Smax . ∗ ΩT ∂ΩT ∂ ΩT By analogy we can ﬁnd the minimum principle by making the substitution u = −u ˆ to obtain ˆ L(−ˆ) = −AˆSS − aˆS + uτ = D(S. provided A > 0. then if D ≥ 0 then the maximum occurs on the ˆ boundary. Suppose that u(S. τ ) satisﬁes the inequality L(u) > 0 in the domain ΩT then u cannot have a (local) maximum at any interior point. MAXIMUM PRINCIPLES 225 hence for a onedimensional rectangular domain. An outline of the proof of such maximum principles is easily seen by considering the heat equation operator L(u) = uSS − uτ . since at such a point uSS ≤ 0 ˆ boundary ∂ ∗ ΩT . u u ˆ u ˆ ˆ Hence the maximum principle states that if −D ≥ 0 or D ≤ 0 then u has its maximum ˆ on the boundary and hence u = −ˆ has its minimum on the boundary. and furthermore that the coeﬃcients of equation (A. ˆ we have that. To summarise. τ ). then the maximum of the solution (which is assumed to be C 2. ∂ ∗ ΩT corresponds to the boundaries τ = 0. We are now ready to formally state the maximum principle applied to equations of the form (A.e.
τ. it is possible to interpret the maximum principle from a probabilistic point of view. MAXIMUM PRINCIPLES 226 and uτ = 0. uSS )uSS + a(S. see Protter and Weinberger (1984). uS )uS − uτ . Hence u(S. for a more formal proof. thereby violating L(u) > 0. where u(S. ∂Ω\Ω × {0}. i. Let us consider a process started at S and time t and furthermore let M denote the maximum value of the function on the boundary and m the minimum value. Taking expectations given that the process starts at S.3. in other words L(u) = A(S.1 Nonlinear equations The maximum principles outlined above also hold for any nonlinear equation that can be expressed in the form ˆ L(u) = D(S. τ ). and G(S. It follows immediately that m ≤ G (SγΩ . t) is the value of the function on the boundary of the domain. A. γΩ ) ≤ M. Aside As an interesting aside. τ ) will satisfy a forwards parabolic PDE and the solution to this diﬀerential equation must have it’s maximum on the boundary ∂Ω\Ω × {T } = ∂ ∗ Ω. u.APPENDIX A. i. t) ≤ M. γΩ ) St = S] ≤ M. uS . t) will also satisfy a backwards parabolic PDE via the FeynmanKac representation theorem outlined in subsection 1. where the coeﬃcients of the derivatives in the operator can now be functions of the solution and its derivatives.3. .e. u. where γΩ denotes the ﬁrst exit time of the process from the domain Ω. τ. ⇒m ≤ u(S.e. St = S we have that m ≤ E [G (SγΩ .
V VS V SS 2 1 2 1 where V 2 < V < V 1 . (A. In other words in general form ∂F (Si . VS .q=1 i.5) . τ. ∂p11 ∂uSS n A. τ.4) as a linear combination of V 1 − V 2 and its ﬁrst and second derivatives with respect to S. Let V 1 and V 2 denote two solutions of the above PDE. p1 . pi . MAXIMUM PRINCIPLES 227 As an example. consider a general nonlinear PDE of the form uτ = F (S. p.e. VS < V S < VS and VSS < V SS < VSS . uSS ) ξ1 > 0 ⇒ > 0. VSS ) − F (S. hence we have 1 1 2 2 Vτ1 − Vτ2 = F (S. pij ) ξl ξq > 0.APPENDIX A. positive deﬁnite. τ. V 1 . u. VSS ). We can apply the maximum principles provided that the function F is elliptic in all values of its arguments in ΩT . τ ). uSS ). τ.2 Uniqueness of PDEs We can use the maximum principle to prove the uniqueness of the solution to a PDE of the general form Vτ = F (S. where we assume that ellipticity has been shown. p11 ) 2 ∂F (S. For the one dimensional case (n = m = 1) we have ∂F (S. Doing so we obtain (V 1 − V 2 )τ = ∂F ∂V (V 1 − V 2 ) + ∂F ∂VS (V 1 − V 2 )S + ∂F ∂VSS (V 1 − V 2 )SS . ∂plq l. τ. (A. τ. p. uS . uS . VSS ). V. VS . Hence the equation takes the form L(V 1 − V 2 ) + c(V 1 − V 2 ) = D(S. u. τ.4) Using the the mean value theorem we can rewrite the righthandside of equation (A. VS . V 2 .
MAXIMUM PRINCIPLES 228 where A= a= c= ∂F ∂VSS ∂F ∂VS ∂F ∂V . since both solutions must satisfy the same boundary conditions. which exhibits no such degeneracy. Fortunately we can make the change of variable x = log S which reduces the equation to 1 Vτ = σ 2 Vxx + rVx − rV = F (x.1 The linear BlackScholes equation If we wish to apply the maximum principle to the BlackScholes equation we have the problem that the coeﬃcient of the diﬀusion term will become degenerate at S = 0. V SS . VS . Therefore provided A > 0 and that the coeﬃcients remain bounded. V Finally. 2 . V. Evaluating the derivatives of F .5) becomes 1 2 1 σ (V − V 2 )xx + r(V 1 − V 2 )x − (V 1 − V 2 )τ − r(V 1 − V 2 ) = 0. equation (A.2. A. hence we must have that u = 0 on the boundary. recall that this can be reduced to an equation of the form L(u) = 0 by making the transform u = e−cτ (V 1 − V 2 ) and so we can now apply the relevent maximum principle.APPENDIX A. then the maximum principle will ensure that u ≡ 0 ⇒ V 1 − V 2 ≡ 0 ∈ ΩT . 2 hence a constant coeﬃcient linear PDE. We know that on the boundary V 1 − V 2 must equal zero. hence V 1 ≡ V 2 proving uniqueness. Vx . Vxx ).
This is natural since. τ ) = 0 and so the application of the 2 maximum principle ensures that the maximum of the solution must occur on the boundary and so we must have that u ≡ 0. uxx − ux  < λ Transforming the above back to the original variables it is clear that this corresponds to VSS  < 1/λ for all (S.APPENDIX A.e. The next step is to check that the function F in (A.2. hence V 1 ≡ V 2 giving uniqueness.6) and we are now in a situation where the equation is no longer degenerate. i.6) is elliptic in all values of its argument.1) to uτ = σ 2 (uxx − ux ) 2 2 1 − λe−2x+rτ (uxx − ux ) = F (x. (A.7) which can be seen to be strictly positive (and hence elliptic) if and only if we have e2x−rτ ∈ ΩT . 2 229 ˆ Hence we can identify A = 1 σ 2 > 0 and D(S.1) but again making the same log transform as for the BlackScholes case will remove such a degeneracy.2 The nonlinear (illiquid) BlackScholes equation The problem of the degeneracy of the diﬀusion coeﬃcient at S = 0 is also present in the fully nonlinear equation (4. τ ). if the denominator is allowed to vanish then we . in other words the restriction that the denominator in equation (4. which reduces equation (4. For simplicity we shall make a further transform to the forward prices for the stock and the option. Diﬀerentiating F with respect to the second derivative gives σ 2 1 + λe−2x+rτ (uxx − ux ) ∂F = 3 ∂uxx 2 1 − λe−2x+rτ (uxx − ux ) (A. The consequence of this is that we are not able to use maximum principles in the regime where the denominator is allowed to vanish.1) cannot vanish (cf. MAXIMUM PRINCIPLES and letting u = e−rτ (V 1 − V 2 ) leads to 1 2 σ uxx + rux − uτ = 0. 1998). τ. hence to see under which situations we can apply the maximum principles. Frey. ux . A. we make the transform S = ex−rτ and V = ue−rτ . uxx ) .
7) in which case equation (A.APPENDIX A. . A. Doing so and setting w = to the following second order (linear) PDE for w σ 2 (1 + λe−2x+rτ (uxx − ux )) 2 1 − λe−2x+rτ (uxx − ux ) 3 ∂u ∂λ leads (wxx − wx ) − wτ = 1 − λe−2x+rτ (uxx − ux ) σ 2 e−2x+rτ (uxx − ux )2 3. We can do so by diﬀerentiating (directly) the transformed equation (A. and investigate other properties. Next.1) to that proposed by Frey (1998). Note that this result still stands for any functional form of the liquidity parameter λ(S. provide we do not allow the denominator to vanish.5) becomes σ 2 (1 + λe−2x+rτ (uxx − ux )) 2 1− λe−2x+rτ (uxx − ux ) 3 (u 1 − u2 )xx − (u1 − u2 )τ = 0.3 Monotonicity in λ Having proved uniqueness we now wish to determine the dependence of the solution to equation (4. since uniqueness is preserved under the inverse transforms required to convert equation (A.1).6) with respect to λ.1) in the regime where VSS  < 1/λ everywhere within the domain. we shall proceed to prove uniqueness of the solution. Applying the maximum principle to the above equation yields the required uniqueness result. This can be done simply by using (A.6) back into equation (4. from which it can be seen that (provided the denominator does not vanish) then we will have that A > 0. τ ). which also contradicts the requirement for the applicability of such maximum principles. MAXIMUM PRINCIPLES 230 have unbounded coeﬃcients of the equation. V 1 ≡ V 2 in ΩT . In addition. smoothness of the solution can also not be determined a priori if the denominator is allowed to vanish. (A. we wish to provide an alternative uniqueness proof for equation (4.8) It is advantageous at this point to rewrite the above using the following inverse transform uxx − ux = S 2 uSS = e2x−2rτ uSS = e2x−rτ VSS = e2x−rτ Γ. However.1) on the liquidity parameter λ. of (4.
e. when the denominator is not allowed to vanish. Now applying the maximum principle. (A. gives that the solution w must have its minimum on the boundary. w ≥ 0 in the interior of the domain ΩT . MAXIMUM PRINCIPLES 231 where we have deﬁned Γ := VSS .APPENDIX A.e.9) is negative. or more speciﬁcally the minimum principle since the righthandside of (A. we must have w(x. and thus equation (A. i. 0) = 0 and more generally w = 0 on the boundary ∂ ∗ ΩT . i. 2 (1 − λΓ)3 2 (1 − λΓ)3 (1 − λΓ)3 parameter λ. ∂λ ∂λ hence the solution is an increasing function of λ. in other words ∂V ∂u ≥0⇒ ≥ 0 ∈ ΩT .9) Since the initial condition of any option contract will be independent of the liquidity .8) becomes σ 2 (1 + λΓ) σ 2 e2x−rτ Γ2 σ 2 (1 + λΓ) wxx − wx − w τ = − . It should be emphasised that this result only holds in the regime VSS  < 1/λ.
where ρ1 = 2r σ2 and ρ2 = 2(µc +D) . t=T − 2τ . which it trivially must satisfy due to the speciﬁes of the option contract. v(x. v(x. 232 . σ2 2r V (S. τ ) = ex+ρ1 τ Φ √ ∂x 2τ on x = xf . τ ) + 1 − ex . The resulting nondimensional system becomes vτ − vxx + (1 − ρ1 + ρ2 )vx = eρ1 τ (ρ2 ex − ρ1 ) .Appendix B Nondimensionalisation of the British Put The freeboundary formulation of the British put option value (9. t) = K e− σ2 τ v(x. 0) = (ex − 1)+ . σ2 In addition we have the condition that xf (0) = log r µc if µc ≥ r. τ ) = ex+ρ1 τ Φ x+τ x−τ √ − e ρ1 τ Φ √ 2τ 2τ ∂v x+τ on x = xf . τ ) = (ex − 1) eρ1 τ as x → ∞. v(x. 1 Note that the strike price K is scaled out of the problem (completely) by a simple linear scaling. (x.3) can be nondimensionalised by making the following substitution1 S = Kex−(µc −D)(T −t) .
i.e.1) where we have assumed the process was started at position St .Appendix C The Probability Density Function In order to determine the probability density function under the risk neutral measure Q we adopt the following procedure. This process has the closed form solution 1 St+u = St exp σ(Wt+u − Wt ) + r − D − σ 2 u 2 (C. Hence we have that the probability that St+u . We are given the stochastic process dSt = (r − D)St dt + σSt dWtQ which is the process under the measure Q. 2 St z 1 log St − r − D − 2 σ 2 u . ⇒ PQ [St+u ≤ z] = PQ Wt+u − Wt ≤ σ 233 . the value of the stock price at time t + u given that it started at St is under some value z is given by 1 PQ [St+u ≤ z] = PQ St exp σ(Wt+u − Wt ) + r − D − σ 2 u ≤ z 2 1 z = PQ exp σ(Wt+u − Wt ) + r − D − σ 2 u ≤ 2 St 1 z = PQ σ(Wt+u − Wt ) + r − D − σ 2 u ≤ log .
that Wt+u − Wt follows the same law as uW1 hence we have PQ [St+u ≤ z] = PQ W1 ≤ ⇒ P [St+u Q log where we have used the standard result that P [W1 ≤ y] = Φ(y). √ σ u − r − D − 1 σ2 u 2 . The transitional probability density function is given by the derivative of the above with respect to z. THE PROBABILITY DENSITY FUNCTION 234 It is known. Now directly computing the derivative of (C. y. z. hence ∂ ∂ PQ [St+u ≤ zSt = S] = ∂z ∂z z f (S. y. t. z. t + u) is the transitional probability density function of the process at time t + u and position y. t + u). given that it started at S at time t. t. 0 where f (S. t + u)dy. t. t. from the normally distributed independent increment property of the √ Wiener process Wt . y.2) PQ [St+u ≤ zSt = S] = f (S.APPENDIX C. t + u) = √ σz 2πu under the measure Q. t + u)dy 0 = f (S. t. z ≤ z] = Φ log z St − r − D − 1 σ2 u 2 .2) with respect to z yields log 1 exp − f (S. √ σ u z St (C. z S 1 − r − D − 2 σ2 u 2σ 2 u 2 . this can be seen from the deﬁnition.