Smile dynamics: modeling the smile of

volatility (of volatility)

Lorenzo Bergomi

Equity Derivatives Quantitative Research

Société Générale
lorenzo.bergomi@sgcib.com

Roma, Global Derivatives 2009

 Outline

ƒ Rationale for new type of model

ƒ A simple framework for the dynamics of fwd variances
ƒ Generating smiles for fwd variances
ƒ Using the model
ƒ Calibration to VIX smiles
ƒ Application to swaptions & options on realized variance
ƒ The vanilla skew
ƒ The Skew Stickiness Ratio
ƒ Conclusion

Lorenzo Bergomi 1
 Motivation - 1

Nov Dec Jan VIX smiles - Sept 25 2008

ƒ New listed volatility instruments 100%

90%

ƒ VIX futures & options 80%

70%
60%

50%

ƒ New OTC options on variance 40%

30%

ƒ Swaptions 20% 25% 30% 35% 40%

ƒ Options on realized variance, spot or forward-starting

ƒ Multi-asset options on realized variance t T1 T2

ƒ Hybrid equity / realized variance payoffs

2
1 T ⎛ Si +1 ⎞ 2

S = VTT T1
1 2
S= ∑ ln⎜⎜ ⎟⎟
T2 − T1 i =T ⎝ Si ⎠ 1

¸ Require modeling of
- joint dynamics of spot / variance curve
- smile of forward variances

Lorenzo Bergomi 2
 Motivation - 2

ƒ1-factor stoch. vol models generate poor smile dynamics

1 1
- One single time scale: for T >> 1/k, σ Vol ∝ , Skew ATM ∝
T T

ƒ Popular smile models have other structural problems

ƒ Heston model:
- short ATM vol is normal
- short ATM skew inv. prop. to level of ATM vol
- simulation of dynamics of variance curve uselessly arduous

ƒ Jump / Lévy models:

- forward skew but no dynamics of vol / skew
- process for producing skew generates unrealistic skewness for short-term returns
¸ pbs for very path-dependent options

ƒ Stoch. Vol Lévy

- dynamics is very constrained: for short maturities dσˆ 1
∝
d ln K F σˆ F

Lorenzo Bergomi 3
 A simple framework - 1

ƒ Start with simple lognormal dynamics for forward variances driven by OU factors:
ω2
T
ξt = ξ0 eT
T
ω xt −
2
E [ ( x tT ) 2 ]
ξ T
0 =
d
dT
T ˆ
σ 2
t = 0 (T ) ( )
xtT = (1 − θ ) e − k1 (T − t ) X t + θ e − k 2 (T − t )Yt
X
dX = − k1 Xdt + dW
Y
dY = − k 2Ydt + dW
X
Xt, Yt are Gaussian: correlation of W and W Y is ρ XY

ƒ Dynamics is Markovian:
1
Vt T T (t , X t , Yt ) =
T
∫
2 T
- Fwd variances are functions of just 2 easy-to-simulate factors: 1 2
ξ t dT
T1 T2
T2 − T1 T1

dV
T1 T2
= αTX1 T2 dW X + αYT1 T2 dW Y
V
T2 T2
ξtT e − k1 (T −t )dT ξtT e − k 2 (T −t )dT
αTX1 T2 = (1 − θ )
∫T1
, αYT1 T2 = θ
∫
T1
T2 T2

Lorenzo Bergomi
∫ T1
ξtT dT
∫T1
ξtT dT
4
 A simple framework - 2

ƒ We can use same dynamics for discrete forward variances VtT T

i i+1

- Allows for independent control of short fwd skew

Ti Ti+1

ƒ 2-factor model is minimal setup that provides handle on:

- term-structure of volatilities of volatilities
- term-structure of skew

80% -0.2
-2.5 -1.5 -0.5 0.5 1.5
70% SP500 -0.4
SP500
60% STOXX STOXX
-0.6

ln( Vol of Vol )

NIKKEI NIKKEI
Vol of vol

50%
-0.8
40%
Slope ~ -0.4 -1
30%

20% -1.2

10% -1.4
0%
-1.6
0 0.5 1 1.5 2 2.5 3
Mat ln(Mat)

- Adding more factors generates no additional complexity:

1 2
ω x tT − ω E [( x tT ) 2 ]
ξ tT = ξ 0T e 2 xtT = ∑w e
n
n
− k n (T − t )
X tn

Lorenzo Bergomi 5
 Smile for fwd variances – continuous 1

ƒ Consider instantaneous fwd variances ξT

ƒ Would like to keep model Markovian while relaxing lognormality:
- Define
xtT = α θ ((1 − θ ) e − k1 (T − t )
X t + θ e − k 2 (T − t )Yt ) with α θ = 1 / (1 − θ ) 2 + θ 2 + 2 ρ XY θ (1 − θ )

- Write ξ tT = ξ 0T f T ( xtT , t )
T 2 2 T
- Conditions on fT: - ξ T
is driftless: df λ ( T − t ) d f
+ = 0
dt 2 dx 2
with
[
λ2 (τ ) = α θ2 (1 − θ ) 2 e − 2 k1 τ + θ 2 e − 2 k 2 τ + 2 ρ XY θ (1 − θ ) e − (k1 + k 2 )τ ]
- f T (0, 0 ) = 1
- f T monotonic in x

ω2 T
ωx − ∫T −t λ2 (τ )dτ
- Ex: lognormal dynamics given by f T ( x, t ) = e 2
Lorenzo Bergomi 6
 Smile for fwd variances – continuous 2

ƒ Analogous to ’’Markov-functional’’ models of fixed income

- Is in fact a local vol model:

∂ ln f T
σ (ξ , t ) = η(T − t )
T

∂x x= f T
−1
( ξ ,t )

- f Tuniquely determined by terminal profile: f T (x, t = T )

1 Ti+1
ƒ No market smiles for individual ξ T, rather for discrete fwd variances: Vt T T = i i +1

Ti +1 − Ti ∫
Ti
ξtT dT
- (1) Assume f (x,Ti ) are identical for T ∈ [Ti ,Ti+1[: exact calibration on VIX smiles
T
T
- (2) Use analytical ansatz for f : representation on basis of space-time harmonic functions
+∞ ω2 T
∫T −t λ(τ ) dτ
2
ωx −
f T (x, t ) =
∫ dµ(ω)
0
e 2

Using just 2 exponentials:

ωT2 T
λ(τ ) dτ ( βT ωT )x − ( βT ωT )2 T
∫T −t ∫T −t λ(τ ) dτ
2 2
ωT x − 1
f T
(x, t ) = (1 − γT ) e 2 + γT e 2 with ωT = (2νζT )
1 − γT + γT βT

Initially vol of very short vol is ν : dξtT=T = ( 2νζ T ) ξ T dxT

Lorenzo Bergomi 7
 Smile for fwd variances – continuous 3

ƒ Process for spot

dS = ( r − q ) Sdt + ξ 0t f t ( xtt , t ) S dWt

xtt = αθ ((1 − θ ) X t + θYt )

ƒ Vanilla skew is set by correlations ρSX , ρSY

ƒ Pricing is painless
• Processes X t , Yt are simulated (1) exactly, (2) with no time-stepping
• Brownian motion for spot process

ƒ Before we turn to examples: discrete version

Lorenzo Bergomi 8
 Smile for fwd variances – discrete 1

ƒ Assume a tenor structure of expiries, for example those of VIX futures & options
V TiTi+1
ƒ Model discrete fwd variances V
TiTi+1
: Ti Ti +1
VtTi Ti +1 = V 0Ti Ti +1 f i ( xti , t )
df i λ2 (Ti − t ) d 2 f i
(
xti = α θ (1 − θ ) e − k1 (Ti − t ) X t + θ e − k 2 (Ti − t )Yt ) dt
+
2 dx 2
=0

ƒ Determine terminal profile f ( x,t = Ti ) so as to exactly recover market smile forV i i+1
i TT

- Use procedure outlined by J. Kennedy, P. Hunt, A. Pelsser

T Ti+1
- VIX smiles provide market prices for digitals on VTT T
i
i i+1
hence on VT i
i

L L
V TiTi+1 DL = P (VTTiTi+1 < L ) = P ( xTi i < x (L )) with f i (x ( L ), Ti ) =
0 i
V0TiTi+1

1
DL
f(x)

N(x)
x
x (L)
Lorenzo Bergomi 9
 Smile for fwd variances – discrete 2

ƒ Process for spot over interval [Ti , Ti+1[

[ TT
ƒ Solution (1): variance over interval Ti , Ti +1 stays constant, equal to V i i+1
Ti
[ = V0TiTi+1 f i ( xTi i ,T )
TiTi+1
dS = ( r − q)Sdt + VT S dWt
i
¸ both vanilla skew and fwd skew set by correlations ρSX, ρSY
⎛ S TT ⎞
ƒ Solution (2): control fwd skew by choosing ’’local vol’’ σi (S ) ≡ φ⎜⎜ ,VTi i i+1 ⎟⎟ such that:
• VS variance for maturity Ti +1 is VTTiTi+1
⎝ STi ⎠
i
• Desired level of fwd skew over [T , T [ is obtained
i i +1

• Desired dependence of level fwd skew on level of vol is achieved

ƒ For example, use CEV form: φ( x ) = σ 0 x

1− β

σˆ95% − σˆ105%
= 0.25
σˆ95% −σˆ105% = 5% σˆ100%
Cst skew Prop skew
25 60% 7 60%

50% 6 50%
20
β 5
β
σ0 40% 40%
σ0
15
β4 σ0
β 30% 30%
10
σ0 3
20% 20%
2
5
10% 1 10%

0 0% 0 0%
0% 10% 20% 30% 40% 50% 60% 0% 10% 20% 30% 40% 50% 60%
σVS σVS
Lorenzo Bergomi 10
 Smile for fwd variances – discrete 3

ƒ Benefits of using discrete version:

ƒ Control on smile of fwd vols – VIX smiles easily calibrated

ƒ Control on short fwd skew and its dependence on level of fwd vol
ƒ Correlations ρSX, ρSY still at our disposal to control vanilla smile independently

ƒ Again pricing is painless:

• Processes Xt , Yt are simulated (1) exactly, (2) with no time-stepping
• Brownian motion for spot process
• Mapping functions f
i
(x, t) are generated once and for all

Lorenzo Bergomi 11
 Using the model – calibration to VIX smiles - 1

ƒ Here we use continuous version

ƒ Calibration on VIX smiles:

• Choose values for ν, θ, k1, k2 , ρXY : sets general dynamics of fwd vols in model
• Use piecewise constant parameters γi , βi , ζi on each interval [Ti , Ti+1[
- γ, β control skew / ζ controls vol level – of fwd vol

• Use VS term-structure computed on SP500 smile as input

• Calibrate both
- VIX futures themselves F = E[ VTT T ]
i
i i +1
30%
Sept 25,2008

- VIX options C K = E[ ( VTTiTi+1 − K )+ ] 28%

i
26%

• Is VIX market consistent with SP500 option’s market ? 24%

VIX futs
SP VS
- 1-month fwd variance can be replicated with VIX futures & options 22% VIX VS

F ∞
∫ ∫
TiTi+1 2 20%

V = F + 2 PK dK + 2 CK dK

9
9
8

9
XZ

XG

XM
XJ
XH

XK
XF
XV

XX

VI

VI
VI
VI

VI

VI

VI

VI

VI
0 F

Lorenzo Bergomi 12
 Using the model – calibration to VIX smiles – 2

• Pricing of options on fwd vol/var easy: simple Gaussian integration

•March 18, 2008

28% 90%
γi βi ζi ν 130.0%
Model
27%
Market
Market 16-Apr-08 87% 21% 105% θ 28%
80%
Model
21-May-08 36% 11% 94% k1 8.0
26%
70% 18-Jun-08 35% 0% 95% k2 0.35
25% 16-Jul-08 30% 0% 99% ρXY 0%
60% 20-Aug-08 24% 0% 100%
24%

50%
23%

22% 40%
Apr May Jun Jul Aug 20% 25% 30% 35%

•Sept 12, 2008

28% 90%
Model
γi βi ζi
Market
27% Market 22-Oct-08 35% 16% 87%
Model 80%
19-Nov-08 26% 5% 89%
26%
70%
17-Dec-08 23% 5% 95%
25% 21-Jan-09 39% 26% 91%
60% 18-Feb-09 23% 9% 90%
24%

23% 50%

22% 40%
Oct Nov Dec Jan Feb 20% 25% 30% 35% 40%

Lorenzo Bergomi 13
 Using the model – calibration to VIX smiles – 3
•Sept 25, 2008
28%
100%
Market
γi βi ζi 30%
Sept 25,2008

Market
27%
90% Model 22-Oct-08 76% 21% 114%
Model 28%
80% 19-Nov-08 70% 28% 109%
26% 17-Dec-08 76% 27% 104% 26%
70%
25% 21-Jan-09 41% 29% 96%
60% 24% VIX futs
18-Feb-09 30% 21% 94% SP VS
24%
50% 22% VIX VS

23%
40% 20%
22%

9
9
8

9
30%

XZ

XG

XM
XJ
XH

XK
XF
XV

XX

VI

VI
VI
VI

VI

VI

VI

VI

VI
Oct Nov Dec Jan Feb 20% 25% 30% 35% 40%

ƒ Feb 25, 2009

49%
130%
Model
γi βi ζi 49%
Feb 25,2009

Market VIX futs

47% Market 18-Mar-09 70% 20% 122% 47% SP VS
Model
110% 15-Apr-09 59% 26% 123% 45%
VIX VS
45%
20-May-09 26% 0% 119%
43%
43%
90% 17-Jun-09 60% 37% 123%
41%
41% 22-Jul-09 89% 39% 125%
39%
39%
70%
37%
37%
35%
35%
50%

9
9

9
9

9
XQ
XM
XJ
XH

XK

XN

XU

XV

XX
March April May June July

VI
VI

VI

VI

VI

VI

VI

VI
VI
30% 50% 70%

ƒ March 13, 2009

49%
110%
Model
γi βi ζi March 13, 2009
Market 44%
47% Market 15-Apr-09 76% 23% 109% VIX futs
Model 42%
SP VS
20-May-09 54% 21% 109% VIX VS
45%
90%
17-Jun-09 62% 28% 114% 40%
43%
22-Jul-09 50% 30% 116% 38%

41% 19-Aug-09 40% 31% 116% 36%

39% 70% 34%

37% 32%

30%
35%
50%

9
9

9
9

9
XQ
XM
XJ

XK

XN

XU

XV

XX
April May June July August

VI

VI

VI

VI

VI

VI

VI
30% 50% 70%

VI
Lorenzo Bergomi 14
 Using the model – options on realized variance – swaptions
+
⎛ 1 T ⎛ S ⎞2 ⎞
• Option on realized variance pays⎜ ∑ ln⎜⎜ i +1 ⎟⎟ − K ⎟, swaption pays
⎜ T 0 ⎝ Si ⎠ ⎟
(V T1T2
T1 − K)
+
⎝ ⎠

• Popular approximation:
tσ rt2 + (T − t )σˆt2,T −t
- Define underlying effective underlying Y: Y = 0.2
log / log

- Y hedged by trading VS of maturity T: Y is driftless

T 0
0 1 2 3 4
-0.2

- Dynamics for Y depends on dynamics of σˆt ,T −t -0.4

slope ~ -0.35
-0.6

¸ Price of option on variance only depends on dynamics of VS vol of residual maturity. -0.8

-1
-1.2

• Price option with model

Volatility of VS vol
Set1 Set 2 Set 3 120%
ν 130.0% 137.0% 125.0% Set 1
θ 28% 29% 32%
k1 8.0 12.0 4.5 100% Set 2
k2 0.35 0.30 0.60 Set 3
ρXY 0% 90% -70% 80%
Spot-starting realised
Set1 Set 2 Set 3 + 60%
⎛ 1 T ⎛ S ⎞2 ⎞
6 months 2.03% 2.03% 2.03% ⎜ ⎟ −K⎟
⎜T ∑
ln⎜⎜ i +1
1 year 2.22% 2.21% 2.24% ⎝ Si ⎟⎠ ⎟ 40%
⎝ 0
⎠
+
6 months realized in 6 months ⎛ 1 T ⎛ S ⎞2 ⎞
⎜ ⎟
2

⎜ T2 − T1 ∑
ln⎜⎜ i +1
⎟ −K 20%
3.04% 2.89% 3.25%
⎝ Si ⎟⎠ ⎟
⎝ T1
⎠

(V − K)
+
6 months in 6 months swaption 0%
T1T2
2.28% 2.10% 2.57% T1 0 6 12 18 24
Maturity (months)

¸ Pop. approx ok for spot-starting options on realized variance

Lorenzo Bergomi 15
 Using the model – dynamics of fwd vols
Set1 Set 2 Set 3
ν 130.0% 137.0% 125.0%
θ 28% 29% 32%
k1 8.0 12.0 4.5
k2 0.35 0.30 0.60
ρXY 0% 90% -70%

Correlation of spot-starting 1-month VS vol with 1- Volatility of VS vol

120%
month fwd vols
100 Set 1
100% Set 2
Set1
80 Set 3
Set 2
80%
Set 3
60
60%
40
40%
20
20%
0
0 1 2 3 4 5 6 7 8 9 10 11 0%
-20 0 6 12 18 24
maturity (months) Maturity (months)

¸ term-structure of vols of spot-starting vols does not uniquely determine vols & correls of fwd vols
Spot-starting variance = basket of fwd variances. Same basket vol can be achieved with:
• low vols / high correls
• high vols / low correls

For flat term-structure of VS vols, inst. vol of VS vol given by:

2 2
⎛ 1 − e − k1T ⎞ ⎛ − k 2T ⎞ ⎛ − k1T ⎞⎛ 1 − e − k 2T ⎞
σTvol = ναθ θ2⎜ ⎟ + (1 − θ ) 2 ⎜ 1 − e ⎟ + 2 ρ XY θ (1 − θ )⎜ 1 − e ⎟⎜ ⎟
⎜ k1 ⎟ ⎜ k2 ⎟ ⎜ k1 ⎟⎜ k2 ⎟
⎝ ⎠ ⎝ ⎠ ⎝ ⎠⎝ ⎠
Lorenzo Bergomi 16
 Using the model – options on realized variance – smile

• Use parameters calibrated on VIX smiles of march 18, 2008

+
⎛ 1 T ⎛ S ⎞2 ⎞
⎜ ln⎜⎜ i +1 ⎟⎟ − K ⎟
• Price option on realized variance – back out smile: ∑
⎜ T 0 ⎝ Si ⎠ ⎟
⎝ ⎠

Smile of realized variance

100%

80%

60%

40%

20%

0%
60% 80% 100% 120% 140% 160% 180%
Vol moneyness

Lorenzo Bergomi 17
 Using the model – the vanilla skew – 1
• Vanilla skew in continuous version is generated by spot / vol correlations ρSX , ρSY

• Approximate expression for vanilla skew:

⎛S ⎞ dσˆ ΣT
- ATM skew is related to skewness Σ T of x = ln⎜⎜ T ⎟⎟ through: =
⎝ S0 ⎠ d ln K F 6 T
- is given at 1st order in vol-of-vol by (double) integral of spot/vol correlation function:
T
t dS τ
∫ ∫
3
( x − < x >) = 6σ 0 dt σt
0 Sτ
0
¸ yields expression of ATMF skew at order 1 in vol of vol:
95% -105% skew
5%
dσˆ ⎡ k T − (1 − e − k1T
) k2T − (1 − e − k 2T
)⎤
= ναθ ⎢(1 − θ )ρSX 1 + θρSY ⎥
d ln K F ⎢⎣ k12T 2 k22T 2 ⎥⎦ 4% Actual
Approximate

3%
• Example: 95% - 105% skew with ρSX = −70 %, ρSY = − 36 %
2%
ν 130.0%
θ 28%
1% k1 8.0
¸ Parameters generating slow decay of vol of vol k2 0.35 Skew ∝
1
ρXY 0% ρSX = −70%, ρSY = − 36% T
also generate slow decay of skew 0%
0 0.5 1 1.5 2

Lorenzo Bergomi 18
 The Skew Stickiness Ratio – 1

• Natural question : how does the VS or ATMF vol move when spot moves ?
- in units of the ATMF skew ?
⎛ δ σVS
T ⎞
⎜ ⎟
⎜ δ ln S ⎟ δσVS δσVS δ ln S
¸ enter the Skew Stickiness Ratio RT = ⎝ ⎠ with =
dσˆ δ ln S (δ ln S )2
d ln K F ,T

• At 1st order in vol of vol, with flat VS term-structure:

• For short maturities RT → 2 : model-independent 2

• For (very) long maturities RT → 1

1.5

1
• General expression for any number of factors:
ν 130.0%
1 − e − k iT
∑
0.5 θ 28%
wi ρiS k1 8.0
kiT k2 0.35
RT = − k iT
ρXY 0% ρSX = −70%, ρSY = − 36%

∑w ρ
kiT − (1 − e ) 0
0 2 4 6 8 10
i iS 2 Maturity (years)
(kiT )

Lorenzo Bergomi 19
 The Skew Stickiness Ratio – 2

• Model-independent (no jumps) expression of the SSR at order 1 in the vol-of-vol (flat term-structure of vols):

1 T

RT = T ∫ f (t )dt
0
1 T t

T2 ∫ dt ∫ f (τ )dτ
0 0

1 dSt
where f (τ ) is the spot/vol correlation function: f (τ ) = σt + τ
dt St
• If f is monotonic then 1≤ R ≤ 2

• For T → 0 above expression recovers well-known model-independent result that R → 2

Lorenzo Bergomi 20
 The Skew Stickiness Ratio – 3
95% -105% skew
• Scaling of f (τ ) for large τ governs 5%

¸ The speed with which the ATMF skew decays 4% Actual

Approximate
¸ The value of the SSR for large T
3%

1
• Imagine that for large t f ( τ ) ∝ then : 2%
β
τ
1% 1
Skew ∝
dσˆ 1 T
ƒ If β > 1 ∝ and RT → 1 0%
d ln K F T 0 0.5 1 1.5 2

dσˆ 1
ƒ If β < 1 ∝ and RT → 2 − β
d ln K F Tβ 2

Exponential decay same as β > 1

1.5
¸ The decay of the ATMF skew (static feature) and the large-T
value of the SSR (dynamic feature) are related for large T through: 1

1 ν 130.0%

Skew T ∝ 0.5 θ
k1
28%
8.0

T 2− R k2
ρXY
0.35
0% ρSX = −70%, ρSY = − 36%
0
¸ Explains why in our example: 0 2 4 6 8 10
Maturity (years)
• intermediate regime where Skew T ∝ 1 / T 0.5 , RT ≈ 1.5
• eventually for long maturities Skew T ∝ 1 / T , RT ≈ 1
Lorenzo Bergomi 21
 Conclusion

• 2-factor stoch vol model allows control over dynamics of fwd vols
• Model is driven by easy-to-simulate Gaussian processes – no sweating
• Can be extended to N factors with no additional complexity
- needed for options on slope of VS curve

• Allows control on smile of fwd vols – almost for free

- can be calibrated to VIX smiles, exactly in discrete version

• Discrete version allows additional control on short fwd skew and decoupling from vanilla skew
- magnitude & dependence on level of short fwd vol

• Parameters governing dynamics of fwd vols also govern decay of vanilla skew
• Model-independent relationship at 1st order in the vol-of-vol links SSR to decay of the ATMF
skew

Lorenzo Bergomi 22