Geodesic strikes for composite, basket, Asian, and spread options

Peter Jäckel∗

First version: 28th July 2012

This version: 11th February 2017

Abstract In practice, it is not uncommon for X to be the mar-

ket price of a standard investment asset such as an
We discuss simple methodologies for the selection of equity share, denominated in its domestic currency
most relevant or effective strikes for the assessment DOM, and for Y to be an FX rate that converts the
of appropriate implied volatilities used for the valu- final investment asset value to a target currency TAR.
ation of composite, basket, Asian, and spread options As a consequence of the natural direction of the FX
following the spirit of geodesic strikes [ABOBF02]. rate, i.e., DOMTAR, meaning the value of one DOM
currency unit expressed in units of TAR, the valu-
ation of a composite option may also incur a quanto
Introduction effect since the asset X and the FX rate DOMTAR
In the context of vanilla, or near-vanilla, derivatives cannot be martingales in the same measure, and this
trading, we often also encounter composite, basket, has to be taken into account. For the purpose of this
Asian, and spread options. Whilst their respective article, however, we shall assume that all involved
payoff rules do in principle warrant treating them underlyings are martingales in the same measure. In
as genuinely exotic options, it is often desirable to practice, this may mean that we have to determine
use a simple approximation for the sake of tract- an effective quanto forward for the asset X in the tar-
ability. A popular approach is to use a relatively get currency TAR by other means of approximation
simple valuation methodology based on multivariate prior to being able to commence with our effective
geometric Brownian motion, the well-known Black- geodesic strike procedure. We shall return to this
Scholes-Merton framework, and to find a suitable im- point at the end of section 3.
plied volatility (or term structure thereof) for each Basket options are, conventionally, derivatives
underlying selected by the concept of a most relev- with a payoff of the form
ant, or effective, strike for each observation date. For P
(θ · [ i wi Xi (T ) − K])+ . (2.2)
a plain vanilla option, clearly, the volatility must be
taken from the implied volatility surface at the op- This is a vanilla option on a linearly weighted average
tion’s expiry date and at the option’s strike. For com- of a number of underlyings, whence it is also referred
posite, basket, Asian, and spread options, the most to as an arithmetic basket option. In contrast, whilst
suitable strike for the looking up of implied volatility rarely traded, there is also the geometric basket op-
is not necessarily as obvious. In this document, we tion
suggest a systematic procedure that addresses this (θ · [ i Xiwi (T ) − K])+ .
Q
(2.3)
question.
Asian options are in the framework of multivariate
geometric Brownian motion merely a special case of
Composite, Basket, Asian, and arithmetic basket options in that the fixings that con-
spread options tribute to the average are from the same underlying,
but for different observation times:
A composite option is a contract of European style P
(θ · [ i wi X (Ti ) − K])+ . (2.4)
which pays at a future payment date Tpay the payoff
Spread options usually simply pay according to
(θ · [X (T ) · Y (T ) − K])+ , (2.1)
(θ · [(X(T ) − Y (T )) − K])+ (2.5)
with θ = ±1 for calls and puts, based on two underly-
ings X and Y observed on the expiry date T ≤ Tpay . for two underlyings X and Y .
∗
Deputy head of Quantitative Research, VTB Capital
By allowing the subscript i on an observation index
Key words and phrases. effective strikes, geodesic, compos- X i to indicate a specific underlying as well as fixing
ite options, basket options, Asian options, spread options. time, and by additionally permitting weights to be

1
negative as well as positive, it is clear that all of bas- relating to the dispersion matrix A according to
ket, Asian, and spread options take on the form of
an arithmetic average option C = A · A> . (3.5)
P
(θ · [ i wi Xi − K])+ . (2.6) The approximation of effective geodesic strikes is now
to find a set of logarithmic shift coefficients ξ ∗ , which
Equally, it is evident that both composite and geo-
relates to the effective strike for underlying #i as
metric basket options appear as a geometric average
∗
option Q wi Ki∗ = X̂i · eξi , (3.6)
(θ · [ i Xi − K])+ . (2.7)
Subsequently, we will therefore concentrate on these such that the multivariate probability density of ξ
two generic cases: geometric and arithmetic average under the joint normal law with
options. D E
ξ · ξ> = C (3.7)

Geodesic strikes is maximal at ξ = ξ ∗ subject to the constraint

The formal derivation of the procedure for the calcu- f (K ∗ ) = K . (3.8)

lation geodesic strikes in [ABOBF02] involves con-
cepts of projection onto an effective local volat- In other words, we seek
ility representation of a basket process for large   
deviations, as well as Varadhan’s geodesic the- ∗ ∗
ξ = arg max ψ(ξ) f (K ) = K (3.9)

ξ
orem [Var67], and is rather technical. In this doc-
ument, we shall attempt to obtain a similar result with
with a somewhat less rigorous, though tractable, ar-
1 > −1
gument, bearing in mind that the sole purpose of e− 2 ξ C ξ
ψ(ξ) := p (3.10)
the exercise is to arrive at a set of suitably chosen (2π)n · |C|
effective strikes for the lookup of implied volatilit-
ies from the underlyings’ implied volatility smiles. and K ∗ being given elementwise in equation (3.6).
These volatilities are then to be used in whatever The log-bilinear form of (3.10) allows us to sim-
near-vanilla approximation that is chosen for the re- plify (3.9) to
spective target product. We emphasize that it is clear   
that this process cannot possibly arrive at a sophist- ∗ > ∗
z = arg min z · z  f (K ) = K (3.11)

icated exotic product pricing framework. Instead, it
merely is intended to give a procedure that suffices
for the simplistic, but very fast, valuation of some
near-vanilla products whilst preserving some sensible
consistency conditions.
The starting point is that all of the involved
stochastic financial observables X are governed by
a joint log-normal law, and that there is a critical
level K for a function f (·) of the financial observ-
ables, identified by the fact that the payout is of the
form
(θ · (f (X) − K))+ . (3.1)
As is well known, without loss of generality, we can
transform the vector of financial variables X to a
vector of independent standard Gaussian variables z
according to
1 P
Xi = X̂i · e− 2 cii + j aij zj
with
X̂ = hXi ,
the matrix A being the dispersion or factor loading
matrix, and the log-covariance matrix C whose ele-
ments are
cij = hln Xi , ln Xj i (3.4)
z
with
ξ∗ = A · z ∗ . (3.12)
As we shall see below, it turns out that the effect-
ive strikes themselves depend on (implied) volatilit-
ies. This makes the task of effective geodesic strike
calculation ultimately an implicit problem, since we
need the effective strikes to be able to look up the
implied volatilities in the first place. In practice, we
resolve this by the approximation that all volatilit-
ies that show up in the effective strike formulæ are
to be taken as at-the-forward implied volatilities. In
this context, we recall that we mentioned at the end
of the first paragraph in section 2 that, when some
of the underlyings require translation into the target
(3.2) valuation measure, an approximation for this trans-
lation of forward and implied volatility smile has to
be employed separately and prior to the invocation of
(3.3) the effective geodesic strike procedure. When choos-
ing the quanto-translation procedure, it is useful to
bear in mind that the subsequent geodesic strike se-
lection is only an approximation for the sake of ana-
lytic tractability for a range of near-vanilla products,
2
and thus judge the required level of sophistication As for the effective geodesic strikes, equation (3.6)
for the quanto translation in line with the overall gives us
level of the chain of approximations. We emphasize,
however, that whilst the geodesic strike procedure is Ki∗ = X̂i · eξi (3.23)
only an approximation, its purpose is to be consistent
with
and accurate in a certain asymptotic sense, namely
P
that of the local volatility projection for large de- κ· j cij wj
viations, i.e, for out-of-the-money options, and that ξi := (3.24)
w> · C · w
of the geodesic distance asymptotics of [ABOBF02]
and [Var67]. for geometric average options. For composite options
as defined in (2.1), we have wi ≡ 1 and
  
Geodesic strikes for geometric average op-

K σX ·(σX +ρXY σY )
ln ·
KX∗ = X̂ · e
2 +2σ ρ
σX 2
X̂ Ŷ X XY σY +σY
tions (3.25)
   
K σY ·(σY +ρXY σX )
ln ·
We consider the payoff f (·) being given by KY∗ = Ŷ · e X̂ Ŷ 2 +2σ ρ
σX X XY σY +σY
2
.
Y
f (X) = Xiwi . (3.13) It is worth reflecting on equation (3.25) with respect
geometric
average i to a few benchmark cases. First, let us consider the
case when σY → 0. In that case, we have
The constraint (3.8) becomes
K
w>·A·z ∗ lim KX∗ = (3.26)
K = Ĝ · e (3.14) σY →0 Ŷ

with which is consistent with the exact plain vanilla op-

(3.15) tion we arrive at in this limit for a composite option,
Y
Ĝ := X̂iwi
i
and we obtain of course the symmetric equivalent for
σX → 0. When correlation is zero, the log-moneyness
which we write as of the composite option translates to log-moneyness
of the underlyings as a function of the log-volatility-
w> · A · z ∗ − κ = 0 (3.16)
ratio
with ν := 2 ln(σX /σY ) (3.27)
κ := ln(K/Ĝ) . (3.17) according to the logistic functions
In order to minimize the L2 -norm of z ∗ subject to the κ κ
ξX∗ = −ν
and ξY∗ = (3.28)
constraint (3.16), we solve the Lagrange multiplier 1+e 1 + eν
problem
as shown in figure 1. The graph illustrates the sym-
h  i
z ∗ = arg min 12 · z > · z − λ · w> · A · z − κ . κ
z
(3.18)
ξ*X
From ξ*Y
κ/2
   
1 > >
∇z 2 · z · z − λ · w · A · z − κ  =0
z=z ∗
(3.19) ν
0
-4 -3 -2 -1 0 1 2 3 4
we have
Figure 1: Underlying log-moneyness for composite options with
z ∗ = λ · A> · w (3.20) zero correlation according to equation (3.28).

which we substitute into (3.16) yielding metry ξX∗ (ν) + ξY∗ (ν) = 1 which is intuitively ap-
pealing. Also, when volatilities are equal, the log-
κ
λ= > . (3.21) moneyness κ is equally distributed over both under-
w ·C ·w lyings, which is again what one would intuitively ex-
Upon resubstitution into (3.20), we arrive at pect from a sensible effective strike approximation
formula. In the limit of ρXY = 1, we obtain
>
κ·A ·w
z∗ = > . (3.22) κ κ
w ·C ·w ξX∗ = − ν2
and ξY∗ = ν (3.29)
1+e 1 + e2

3
 κ
√
k = K , and the contribution of the continuous
part will be largely centered near k and tail off rap-
*
ξX idly as k 0 → ∞ due to the rapid decay of out-of-the-
*
ξY money options value B(X̂, k 0 , σ̂(k 0 )), assuming that
κ/2
σ̂(k 0 ) rises only moderately such as would be consist-
ent with finite second and higher moments. Hence,
while the √ selection of an effective implied volatility
ν at k = K for the underlying does not exactly re-
0
-4 -3 -2 -1 0 1 2 3 4 produce the smile dependence one can obtain from
Figure 2: Underlying log-moneyness for composite options with the continuous integration over all strikes to infin-
perfect correlation according to equation (3.29). ity (3.33), we at least capture the fact that the dom-
inant contribution does √ indeed come from the implied
volatility near k = K , as intended.
which we display in figure 2. We see in figure 2 that
with diverging volatilities, i.e., as ν diverges from 0,
Geodesic strikes for arithmetic average op-
the underlyings’ log-moneyness is less rapidly shifted
from an equal split at ν = 0 to just one of the un-
tions
derlyings, as indicated by the lesser slope near zero, Here, we have
when comparing with the case ρXY = 0 in figure 1. X
This behaviour is also intuitively desirable. f (X) = wi Xi . (3.34)
arithmetic
Finally, we consider the case of perfect correlation average i
given by the option under consideration actually be- The constraint (3.8) is
ing a square option
g(z ∗ ) = K (3.35)
θ · X2 − K + .



(3.30)
with
In comparison, by a standard argument of continuous
X P
g(z) := wi X̂i ·e j aij zj . (3.36)
replication, we can value this option based on the i
Taylor expansion with complete remainder for any
From
smooth function h(x) around k  h i 
1 >
Z x ∇z 2 · z · z − λ · g(z) − K  = 0 (3.37)
h(x) = h(k) + h0 (k)(x − k) + h00 (z)(x − z) dz

z=z ∗
k
(3.31) we obtain
whence
alj zj∗
X P
zi∗ − λ · ali wl X̂l ·e
= 0 j(3.38)
h(x) · 1{x>k} = h(k) · 1{x>k} + h0 (k)(x − k)+ l
Z ∞
Equations (3.35) and (3.38) can be solved for λ and
+ h00 (z)(x − z)+ dz . (3.32)
k the elements of the vector z ∗ by the aid of a nonlinear
√ root finding algorithm such as NL2SOL [DGW81].
Choosing h(X) = X 2 − K and k := K , and taking The effective geodesic strikes are then
the expectation over X, we obtain for the call option
h
i Ki∗ = X̂i · eξi (3.39)
2
E X −K + (3.33)
Z ∞ with
= 2kB(X̂, k, σ̂(k)) + 2 B(X̂, k 0 , σ̂(k 0 )) dk 0
k ξ = A · z∗ . (3.40)
wherein σ̂(k 0 ) is the implied volatility of the asset for As for an initial guess for z ∗ in equations (3.35)
strike k 0 , and B(X̂, K, σ) is the Black vanilla call op- and (3.38), we can either use zero, or proceed to find
tion function for forward X̂, strike K, and implied an expansion as follows.
volatility σ. Equation (3.33) has two parts. √ First, Defining
there is a vanilla call option struck at k = K with X
absolute weight 2k. Second, we have a continuum of X̄ := wi X̂i and κ := ln(K/X̄) , (3.41)
vanilla call options for all strikes above k with dens- i
ity 2. For significantly out-of-the-money options, as
we rewrite (3.35) as
stipulated by the large deviations asymptotics at the
heart of the geodesic argument, the dominant part ∗
X P
wi X̂i ·e j aij zj = X̄ · eκ . (3.42)
will be the discretely weighted call option struck at i

4
We now seek an expansion given by whence
!
λ = 0 + κ · λ(1) + κ2 · λ(2) + · · · (3.43) (2) 1 A> · Ω · A ω> · C · Ω · C · ω
z = 1+2· > −3· 2 ·1
2 ω ·C ·ω (ω > · C · ω)
∗ (1) 2 (2)
z =0+κ·z +κ ·z + ··· . (3.44)
A> · ω
Substituting (3.44) into (3.42) and expanding in κ · > (3.56)
ω ·C ·ω
around 0 gives
" with 1 denoting the identity matrix. This finally

(1) (2)
 gives us the second order expansion
ωi · 1 + aij κzj + κ2 zj
P P
(3.45)
i j A> · ω
z∗ = κ · (3.57)
ω> ·
#
C ·ω
(1) 2
P 
1
+ 2 aij κzj !
j 1 2 A> · Ω · A ω> · C · Ω · C · ω
+κ · 1+2· > −3· 2 ·1
2 ω ·C ·ω (ω > · C · ω)
= 1 + κ + 12 κ2 + O κ3


A> · ω
+ O κ3 .


· >
where we have used the normalized effective weights ω ·C ·ω
As an example, we show in figure 3 the first and
wi X̂i
ωi := .
(3.46) second order expansion solutions for a monthly ob-
X̄
served Asian option of one year maturity in com-
Matching coefficients up to order O κ2 , equa- parison with a numerical solution for an arbitrarily

tion (3.45) gives chosen term structure of arbitrage-free implied volat-

P ility.
i ωi = 1 (3.47)
0 35%
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
> (1) T
ω ·A·z = 1 (3.48) 30%

-0.5
(1) >
ω > · A · z (2) = 1
2 − 1
2 z · A> · Ω · A · z (1) , (3.49) 25%

-1

with the matrix Ω defined to be diagonal and its ele- 20%

ξ σ̂(T )
ments being equal to those of ω. Expanding (3.38) 15%
-1.5

in κ yields Lognormal expansion, first order

10%
Lognormal expansion, second order
(1) (2)
κ· zi + κ2 · zi (3.50) -2
NL2SOL
5%
  implied volatility [right axis]

(1) 2 (2)

P P (1)
− κ·λ +κ ·λ · wl X̂l ali · 1 + alj κzj -2.5 0%

l j
Figure 3: An example for the effective strike adjustments (3.40)
= O κ3


of first and second order as given in (3.57) in comparison with
a numerical solution by the aid of NL2SOL of equations (3.35)
from which we derive and (3.38) for a monthly observed Asian option of one year
maturity with an arbitrarily chosen term structure of arbitrage-
z (1) = λ(1) · X̄ · A> · ω (3.51) free implied volatility. All forwards were equal to 1, and the
strike was at 0.3.
z (2) = λ(1) · X̄ · A> · Ω · A · z (1) λ(2) · X̄ · A> · ω
When any of the weights are negative, it is pos-
(3.52)
by matching coefficients of κ up to second order. sible for X̄ to be zero. As a consequence, ln(K/X̄)
Combining (3.48) and (3.51) results in is undefined, and none of the above equations can be
evaluated. In that case, it may be better to use an
1 1
λ(1) = · (3.53) expansion in
X̄ ω > · C · ω κ̃ := K − X̄ , (3.58)
and hence which gives us, to first order

A> · ω κ̃ A> · ω
z (1) = . (3.54) z̃ ∗ = · >
ω> · C · ω X̄ ω · C · ω
Further, combining (3.49) and (3.52) produces
i.e.,
" #
1 1 1 3 ω> · C · Ω · C · ω
λ(2) = · > · − · 2 K − X̄ X
X̄ ω · C · ω 2 2 (ω > · C · ω) z̃i∗ = aij wj X̂j (3.59)
(3.55) vX̄
j

5
with Sensitivity to forwards
X
vX̄ := wi X̂i cij wj X̂j .
(3.60) In this case, the linear system (3.68) reduces to
i,j h i
1 − λ · C · Γ · ξ (i) − C · γ · λ(i) = λ · C · Θ · ei
Analytical sensitivities of the exact solution
γ > · ξ (i) = −e> i · Θ · ei
The geodesic strikes are obtained as the solution of
equations (3.35) and (3.38). Multiplying (3.38) by A (3.69)
from the left, we reformulate this as wherein e i is the i-th unit basis vector,

ξ−λ·C ·γ = 0 Θ := diag{γ}/ diag{X̂} , (3.70)

P (3.61)
j γi − K = 0 and the superscript (·)(i) represents the derivative
with with respect to the forward X̂i .
γi := wi X̂i eξi ≡ wi Ki∗ (3.62)
for convenience of notation and brevity of the results. Sensitivity to the covariance matrix
In order to compute the sensitivity of the geodesic Using the superscript notation (·)(kl) for the deriv-
strikes Ki∗ to any input, which, for now, we indicate ative of any quantity (·) with respect to the (k, l)-
with a generic prime, i.e., Ki∗ 0 , we use (3.39) giving element ckl of the covariance matrix C, we obtain
us
Ki∗
Ki∗ 0 =
h i
· X̂i0 + Ki∗ · ξi0 (3.63) 1 − λ · C · Γ · ξ (kl) − C · γ · λ(kl) = λ · C (kl) · γ
X̂i
and from hereon focus on the computation of ξi0 . γ > · ξ (kl) = 0
Since the nonlinear system (3.61), which implicitly (3.71)
defines the geodesic strikes, is invariant to the change
of any of the input parameters, we obviously have
h i0 Acknowledgement
ξ−λ·C ·γ = 0
(3.64) The author is grateful to Charles-Henri Roubinet,
hP i0
Head of Quantitative Research at VTB Capital, for
j γi − K = 0.
pointing out the results in [ABOBF02], and for au-
We combine this with the definitions thorizing the release of the findings presented here
(originally from mid 2011, and first completed with
Γ := diag{γ} (3.65)
the second order expansion in March 2012) into the
χi = ln(X̂i ) (3.66) public domain.
and
γi0 = γi · χ0i + γi · ξi0 (3.67) References
into the generic sensitivity equation system
h i [ABOBF02] M. Avellaneda, D. Boyer-Olson, J. Busca, and P. Friz.
1 − λ · C · Γ · ξ 0 − C · γ · λ0 Reconstructing volatility. Risk, pages 91–95, Oc-
tober 2002. www.math.nyu.edu/faculty/avellane/
h i
Avellaneda.pdf.
= λ · C 0 · γ + C · Γ · χ0
[DGW81] J. E. Dennis, D. M. Gay, and R. E. Welsch. NL2SOL –
(3.68) an adaptive nonlinear least-squares algorithm. ACM
0
γ> ·ξ = −γ > · χ0 Transactions on Mathematical Software, 7:348–368,
1981. www.netlib.org/toms/573.
which is of course linear in the sensitivities ξ 0 and λ0 .
We note that this linear system has the same matrix [Var67] S. Varadhan. Diffusion processes in a small time inter-
val. Communications on Pure and Applied Mathemat-
of coefficients on the left hand side for all possible ics, 20(4):659–685, November 1967.
sensitivities that we might intend to compute: irre-
spective of whether we are interested in the sensit-
ivity to a volatility, a correlation, or a forward, the
left hand side is always the same, and it is always
a linear system of size m + 1 if m is the number of
geodesic strikes. It can thus be computationally ad-
vantageous to solve all sensitivities in one sweep, for
instance with a single QR factorization or a single
Moore-Penrose pseudo-inverse to safeguard against
singular (under-determined) dependencies.