Evnine-Vaughan Associates, Inc.
A multi-timescale statistical feedback model of volatility:
Stylized facts
and
implications for option pricing
Lisa Borland
October, 2005Acknowledgements: Jeremy Evnine
Roberto Osorio Jean-Philippe Bouchaud
Layout
• Stylized facts of markets
- Why we need a new model • The non-Gaussian model
-Properties
-Applications: Options and Credit • The multi-time scale model
-Capturing the stylized facts • Work in progress and conclusions
Properties of Financial Time-Series
• Power Law distributions, persistent over
very many timescales: minutes to weeks
Cumulative distribution power law tail -3
Properties of Financial Time-Series
• Power Law distributions, persistent
• Slow decay to Gaussian, as
• Volatility clustering and correlation
• Volatility relaxation (Omori law)
• Close-to log-normal distribution of volatility
• Returns normally diffusive over time-scales
• Leverage effect (Skew: Negative returns Æ higher volatility)
• Time-Reversal asymmetry
25 . 0 −τ
o Empirical --- Gaussian
Consequences
• Risk control
:
under-estimate rare events
• Derivative markets
:
(options, credit)
wrong model of underlying leads to wrong
pricing, wrong hedging
Challenge
• A model that can reproduce the stylized facts
• A model that can reproduce option prices, credit etc.
• A model that captures the correct dynamical features
Desirable
• Intuition
• Parsimony
• Stochastic volatility
(Heston 1993)• Levy noise
• GARCH
• Multifractal models
(Bacry,Delour,Muzy, 2001)Popular models
Problems
• Typically converge too quickly to Gaussian
• Less parsimonious
The Standard Stock Price Model
S
Y ln
=
) ' ( ) ' ( 0 t t t (t)d d dω − >= < = > < δ ω ω ty volatili : return of rateσ
µ
:ω
σ
µ
dt
d
dY
=
+
The Standard Stock Price Model
) ' ( ) ' ( 0 t t t (t) ω − >= < = > < δ ω ω ty volatili : return of rateσ
µ
: Gaussian Distribution Fokker-Planck Equation 2 22
1
ω
d
P
d
dt
dP =
)
2
exp(
2
1
)
(
2t
t
P
ω
π
ω
=
−
ω
σ
µ
dt
d
dY
=
+
S
Y ln
=
Ω
+
=
dt
d
dY
µ
σ
The Generalized Returns Model
Borland L, Phys. Rev.Lett 89 (2002) Borland L, Quantitative Finance 2 (2002)ω
d
P
d
q 2 1)
(
−Ω
=
Ω
Ω
+
=
dt
d
dY
µ
σ
The Generalized Returns Model
The Generalized Returns Model
Borland L, Phys. Rev.Lett 89 (2002) Borland L, Quantitative Finance 2 (2002)The Generalized Returns Model
The Generalized Returns Model
Borland L, Phys. Rev.Lett 89 (2002) Borland L, Quantitative Finance 2 (2002)Tsallis Distribution Nonlinear Fokker-Planck 2 2 2
2
1
Ω
=
−d
P
d
dt
dP
q q t q t Z P = − − Ω 1− 1 2) ) ( ) 1 ( 1 ( ) ( 1β
ω
d
P
d
q 2 1)
(
−Ω
=
Ω
Ω
+
=
dt
d
dY
µ
σ
In other words:
State dependent deterministic model
t t t t t
a
q
b
d
d
2ω
1 2]
)
1
(
[
+
−
Ω
=
Ω
Work with)
(S
Ω
=
Ω
Extensions to Model: i q i q i
P
d
dY
σ
2ω
1)]
(
[
−Ω
−
Ω
=
0
=
Ω
Current Model: eg moving average More realistic model:Extensions to Model: i q i q i
P
d
dY
σ
2ω
1)]
(
[
−Ω
−
Ω
=
0
=
Ω
Current Model: eg moving average More realistic model:Ω
Or: i i j q j i q iP
d
dY
σ
∑
−ω
= −Ω
Ω
=
1 0 2 1)]
|
(
[
Not a perfect model of returns:
Well-defined starting price and time
Nevertheless:
Reproduces fat-tails and volatility clustering Closed form option-pricing formulae
Example European Call Q rT S T K e c = − max[ ( )− ,0] Stock Price
⎭
⎬
⎫
⎩
⎨
⎧
Ω
2
−
+
Ω
=
2∫
T −q t TrT
P
dt
S
T
S
0 1)
(
exp
)
0
(
)
(
σ
Example European Call Q rT S T K e c = − max[ ( )− ,0] Stock Price
⎭
⎬
⎫
⎩
⎨
⎧
Ω
2
−
+
Ω
=
2∫
T −q t TrT
P
dt
S
T
S
0 1)
(
exp
)
0
(
)
(
σ
Integrate using generalized Feynman-Kac 2
)
(
TT
Ω
∝
γ
Example European Call Q rT S T K e c = − max[ ( )− ,0] Stock Price
⎭
⎬
⎫
⎩
⎨
⎧
Ω
−
−
2
−
+
Ω
=
(
0
)
exp
2(
)
(
1
)
(
)
2)
(
T TrT
T
q
g
T
S
T
S
σ
γ
2 1d
d
≤
Ω
T≤
}
)
(
{
S
T
>
K
Payoff ifÆ
Example European Call Q rT S T K e c = − max[ ( )− ,0]
∫
Τ −−
Ω
Ω
=
2 1)
(
)
)
(
(
d d T q rTS
T
K
P
d
e
c
σ σ , ,)
0
(
M
qe
rTKN
qS
−
−=
K=50, T=0.4, sigma=0.3, r=.06
q=1.5
Volatility Smiles
o Empirical implied vols __ q=1.43 implied vols
72 76 80 84 Strike 9 10 11 12 13 14 Vo l
72 76 80 84 Strike 9 10 11 12 13 14 Vo l
72 76 80 84 Strike 9 10 11 12 13 14 Vo l
72 76 80 84 Strike 9 10 11 12 13 14 Vo l
72 76 80 84 Strike 9 10 11 12 13 14 Vo l
Example Currency Futures:
(500 options)1. 0.16 1.4 0.008
q Mean square relative pricing error
Benefits of a more parsimonious model:
1) Better pricing - arbitrage opportunities 2) Better hedging
The Generalized Model with Skew
1 −∝
S
αS
dS
Ω
+
=
Sdt
S
d
dS
µ
σ
αω
d
P
d
q 2 1)
(
−Ω
=
Ω
Volatility Leverage Correlation0 1 2 3 4 Time [Years] -1.5 -1.0 -0.5 0.0 0.5 1.0 1.5 Return [Hourl y ]
Example European Call Q rT S T K e c = − max[ ( )− ,0]
∫
Τ −−
Ω
Ω
=
2 1)
(
)
)
(
(
d d T q rTS
T
K
P
d
e
σ α σ α, , , ,)
0
(
M
qe
rTKN
qS
c
=
−
−5
.
0
−
=
α
q=1.5
Strike BS Implied Volatility
Strike K Strike K T=.03 T=0.1 T=0.2 T=0.3 T=0.55 SP500 OX q=1.5, alpha = -1.
4OCT30A 4OCT35A 4OCT40A 4OCT45A 4OCT50A Series 0 5 10 15 cB id
Call Bid/Ask Theoretical value
4NOV35A 4NOV40A 4NOV45A 4NOV50A V12 0 2 4 6 8 10 12 cB id
Call Bid/Ask Theoretical value
0 5 10 15 20 cB id
4DEC25A 4DEC30A 4DEC35A 4DEC40A 4DEC45A 4DEC50A 4DEC50A 4DEC50A V56
Call Bid/Ask Theoretical value
5JAN10A 5JAN15A 5JAN20A 5JAN25A 5JAN30A 5JAN35A 5JAN40A 5JAN45A 5JAN50A V23 0 10 20 30 cB id
Call Bid/Ask Theoretical value
5MAR30A 5MAR35A 5MAR40A 5MAR45A 5MAR50A V34 0 5 10 15 cB id
Call Bid/Ask Theoretical value
6JAN15A 6JAN20A 6JAN25A 6JAN30A 6JAN35A 6JAN40A 6JAN45A 6JAN50A V45 1 11 21 31 cBid
Call Bid/Ask Theoretical value
)
,
,
(
2σ
α
q
V
S
dS
V
⎟
=
⎠
⎞
⎜
⎝
⎛
=
Volatility
)
,
,
(
2σ
α
q
V
S
dS
V
⎟
=
⎠
⎞
⎜
⎝
⎛
=
Volatility
q-alpha-sigma Volatility vs. VIX
0.00 10.00 20.00 30.00 40.00 50.00 60.00 70.00 10/28/1995 3/11/1997 7/24/1998 12/6/1999 4/19/2001 9/1/2002 1/14/2004 Sqa ISD VIX Close
Options look good ….
What about pricing credit?
Borland L, Evnine J, Pochart B, cond-mat/0505359 (2005)
Chirayathumadom R, et al, Investment Practice Report Project, Stanford University (2004)
Merton Model (1974)
• Equity is a call option on
underlying assets of firm
Assets = Debt + Equity
:
D
A
T<
Bond holders receiveA
T Stock holders receive 0:
D
A
T>
Stock holders receiveA
D
T
−
• Key assumptions
- Underlying assets follow stochastic log normal process
- Debt in terms of single zero coupon bond
- Black-scholes valuation for European call option
• Asset Process:
• Key assumptions
- Underlying assets follow stochastic log normal process
- Debt in terms of single zero coupon bond
- Black-scholes valuation for European call option
• Asset Process:
dA = µA dt + σAdz
• Generalized Process:
Merton Model and Credit Spread
>
−
=<
max[
,
0
]
0A
D
S
T Equity 0 0 0A
S
D
=
−
A S A dA dS S0σ
= 0σ
DebtMerton Model and Credit Spread
>
−
=<
max[
,
0
]
0A
D
S
T Equity 0 0 0A
S
D
=
−
A S A dA dS S0σ
= 0σ
Debt 0D
De
−yT=
⎟
⎠
⎞
⎜
⎝
⎛
−
=
−
−rTDe
D
T
r
y
1
log
0Merton Model and Credit Spread
>
−
=<
max[
,
0
]
0A
D
S
T Equity 0 0 0A
S
D
=
−
A S A dA dS S0σ
= 0σ
α α , , 0 0 q rT qe
DN
M
A
S
=
−
− Debt 0D
De
−yT=
⎟
⎠
⎞
⎜
⎝
⎛
−
=
−
−rTDe
D
T
r
y
1
log
0Analysis
Sectors 1 through 7 are Aerospace, Communication, Construction, Energy, High tech equipment, Financial services and Retail
Q 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 aer os pac e, aut o m a nuf ac tu ri ng, a ir lines C o m m un ic at ion, I T e lec tr ic al , c ons tr uc ti on m a c h iner y ener gy E qui pm en ts (m edi c a l and el ec tr oni c ) F inan c ial s e rv ic es re ta il , per s onal pr oduc ts , f ood pr oc e s s in g
Q values across Industry sectors
1 1.1 1.2 1.3 1.4 1.5 1.6 1 Companies Q q across industries q
0.4 0.6 0.8 1.0 1.2 1.4 1.6 d 0.12 0.17 0.22 0.27 0.32 0.37 q=1, alpha=0 q=1.2, alpha=1 q=1.4, alpha=1 q=1.4, alpha=0.5 q=1.4, alpha=0 Standard model “Reality”
Credit Implied volatility
Results
- q mainly in range 1.2 – 1.5 and α = 0.3
Summary
Non-Gaussian model well describes many features of:
Stock Markets Option Markets
Debt and Credit Markets
Now :
A multi-time scale non-Gaussian model of stock returns [Borland L., cond-mat/0412526 2004] i i j q j i q ij
P
y
y
d
w
W
dy
σ
∑
−ω
−∞ = −=
1[
(
|
)]
1)
)
(
)
1
(
1
(
1
2 1 j i ij ij q qq
y
y
Z
P
−=
−
−
β
−
A multi-time scale non-Gaussian model of stock returns [Borland L., cond-mat/0412526 2004] i i j q j i q ij
P
y
y
d
w
W
dy
σ
∑
−ω
−∞ = −=
1[
(
|
)]
1)
)
(
)
1
(
1
(
1
2 1 j i ij ij q qq
y
y
Z
P
−=
−
−
β
−
Motivation: Traders act on all different time horizons
0
j ij
w
=
δ
More General
A multi-timescale model for volatility [Borland and Bouchaud,(2005)]
τ
ω
σ
iy
=
∆
)]
[(
1
1 2 j i i j ijz
y
y
w
W
−
=
∑
− −∞ =σ
ARCH-like
(
2)
2 0(
y
iy
j)
j
i
z
z
z
−
−
+
=
τ
A multi-timescale model for volatility [Borland and Bouchaud,(2005)]
τ
ω
σ
iy
=
∆
)]
[(
1
1 2 j i i j ijz
y
y
w
W
−
=
∑
− −∞ =σ
ARCH-like
(
2)
2 1 0(
)
(
)
)
(
i ji
j
y
iy
jz
y
y
j
i
z
z
z
−
−
+
−
−
+
=
τ
τ
A multi-timescale model for volatility [Borland and Bouchaud,(2005)] kurtosis decay
)]
[(
1
1 2 j i i j ijz
y
y
w
W
−
=
∑
− −∞ =σ
τ
ω
σ
iy
=
∆
a iji
j
w
=
(
−
)
− tails skew elementary timescale(
2)
2 1 0(
)
(
)
)
(
i ji
j
y
iy
jz
y
y
j
i
z
z
z
−
−
+
−
−
+
=
τ
τ
ARCH-like
1 0 z z g
τ
α
2z
Parameters: controls the tails
controls the memory
the elementary time scale
base volatility
Calibration U n i v e r s a l 1 0 2 min 7 300 / 1 15 . 1 85 . 0 z z z = = = =
τ
α
controls the tails
controls the memory
the elementary time scale
base volatility
Volatility-Volatility correlation variogram
α − 2
(Build-up ) and decay of kurtosis
Elementary timescale tau =1/300 day Signature of jumps in real data?
Multifractal scaling
nl
A
x
x
l
M
n(
)
=
|
i + l−
i|
n=
n ζEvolution of
σ
2 conditioned on an initial volatilityσ
2e
2sVolatility relaxation
Also time-reversal assymmetry
∑
∞ = + − − = 1 1 2 ) ( l l i i i l y y X α Regression gives0
.
9
2≈
z
Analytic results
Borland,Bouchaud 2005• Volatility-volatility correlations: decay as • Model well-defined with power-law tails for • Volatility normally diffusive
α
− 2
l
Numerical results
• Tsallis distribution excellent description on all time-scales • Distribution of volatility
• Multifractal scaling • Volatility relaxation
• Time-reversal assymmetry
• Tested premise of model on real data
t z z y ∆ − = ∆ 2 0 2 1 ) (
1
,
1
2<
α
>
z
Summary
Implications of model to market
• Soft-calibration (due to long relaxation)
• Model operating close to an instability
• Past price changes do influence future investor
behavior
Implications for option pricing
• Price depends on past path history :
Low vol period Æ different price than high vol period
• Returns:
Tsallis-Student distributions on all time-scales
q Æ 1 in a predictable way
• Approximation:
Use single-time model with
q(T).
ασ
Conclusions
- Simple multi-time scale model
-Captures many statistical properties of real returns
-Closed form solution for single time case: options,credit
References:
Borland L, Phys. Rev.Lett 89 (2002)
Borland L, Quantitative Finance 2 (2002)
Borland L, Bouchaud J-P, Quantitative Finance(2004)
Chirayathumadom R, et al, Investment Practice Report Project, Stanford University (2004) Borland L, cond-mat/04122526 (2004)
Borland L, Evnine J, Pochart B, cond-mat/0505359 (2005) Borland L, Bouchaud J-P, Muzy J-F,
Zumbach G, Wilmott Magazine, (March 2005)