# Page 1 of 30

**Modeling the Volatility Smile
**

Emanuel Derman Columbia University

October 27, 2006

Stanford.Smile.fm

October 21, 2006

Page 2 of 30

The Implied Volatility Smile/Surface

Black-Scholes

S&P 1995

B way & 116th, 2004

• • • •

Black-Scholes implied volatilities for equity indices: Σ ( S, t, K, T ) Term structure of strike and expiration, which change with time and market level Always a negative slope w.r.t strike for equity index options What model replaces Black-Scholes?

Stanford.Smile.fm

October 21, 2006

Page 3 of 30

The Smile is a Post-Crash Phenomenon

Stanford.Smile.fm

October 21, 2006

or Δ σ τ
it steepens with maturity
Foresi & Wu: Journal of Derivatives.fm
October 21.Smile.Page 4 of 30
Current Equity Index Smile
as a function of K/S
ln ( K ⁄ S ) as a function of moneyness -------------------. 2006
. Winter 2005
Stanford.

4 –0.4 0.c1)
IT
IS
IL
LE
0.7 –0.5
2
2.3
TO
RE
EXHIBIT 5
PR
0.2
GA
L
–0.5
1
1.Page 5 of 30
Negative Correlation Between Level and Slope
Correlation decreases with expiration
Cross Correlations between Volatility Level and Smirk Slope
0 –0.2 –0.6 –0.9
Corr(Δ c0. 2006
.8
Corr(c0.5
4
4.5
4
4.5
3
3.2 –0.Δ c1)
0
–0.Smile. The left panel measures the correlation based on daily estimates.5
5
Maturity in Years
Maturity in Years
Lines denote the cross-correlation estimates between the volatility level proxy (c0) and the volatility smirk slope proxy (c1).
Stanford.5 –0.8 –0.5
3
3.5
5
0. the right panel measures the correlation based on daily changes of the estimates.1 –0.5
2
2.5
1
1.6 –0.fm
October 21.4 –0.

This suggests mean reversion or stationarity for the instantaneous volatility of the index.
Stanford.Smile.Page 6 of 30
Volatility of Implied Volatility Decreases with Expiration
Short-term implied volatilities are more volatile. 2006
.fm
October 21.

5 0.fm
October 21. 2006
.75 1 Strike/Spot 1.Page 7 of 30
Other Markets
VOD Imp Vol (12/27/01)
70% 60% Imp Vol 50% 40% 30% 0.90 9.25 1.5 1 Mon 3 Mon 6 Mon 12 Mon
single stock implied vols
USD/EUR
MXN/USD
JPY/USD
currency implied vols
ATM Strike = 0.67
Stanford.Smile.85 123.

∼ -------.fm
October 21.-------.∼ 400 ∂Σ 2π ∂Σ ∂Σ 0.∼ 0.8 points
1 δS 1 δS The P&L earned from perfect hedging is Γ × -------. T.25 points .1.0002 ∂ S ∂ K 100
Eq. What’s the right way to hedge index options? Estimate of hedging error for an S&P index option with S ~ 1000 and T = 1 year
dC BS ( S.= ∂Σ∂S dS ∂S ∂C S T ∼ ---------. 2006
. If the valuation formula is wrong. t.
2
2
Stanford.Smile.Page 8 of 30
The Smile is a Problem for Hedging
Black-Scholes is being used as a quoting mechanism for vanilla options (cf. Σ ) ∂C BS ∂C ∂Σ + Δ = --------------------------------------------. therefore the price doesn’t correspond to the correct replication formula. then the hedge ratio is wrong.( 50 ) = 0.08 ∂Σ∂S
For a typical move δS ≈ 10 S&P points the hedging error = 0.∼ -------------.02 ∼ ∼ --------.0002 = 0. yield to maturity for quoting bond prices) but is not the correct valuation/hedging model. 2 2 200 SΣ T
which is dwarfed by the error.1
∂C ∂Σ ∼ 400 × 0. K.

Smile. lookbacks … which effectively involve multiple “strikes” and therefore multiple implied volatilities. averages.fm
October 21. 2006
. we don’t know how to value all sorts of exotic averages: barriers.Page 9 of 30
The Smile is a Problem for Valuing Exotics
Since Black-Scholes is inappropriate. Which one do you use in a Black-Scholes world? How can we fix/replace/alter/correct Black-Scholes?
Stanford.

t ) . CEV models – Stochastic volatility – Jumps/crashes
Stanford. so do realized volatilities – Expectation of changes in volatility over time – Support/resistance levels at various strikes – Dealers’ books tend to be filled by the demand for zero-cost collars
•
Structural Causes: Violations of Black-Scholes assumptions – Inability to hedge continuously – Transactions costs – Local volatility σ ( S. 2006
.Smile.Page 10 of 30
What Could Cause the Smile?
•
Behavioral Causes – Crash protection/ Fear of crashes) [Katrina] Out-o.fm
October 21. leverage effect.money puts rise relative to at-the-money.

virtually no assumptions. then markets are incomplete and you need a utility function. finance is mostly about expectation of the future. liquidity assumptions and a big investment in technology. – dynamic replication is next best – but needs uncertain replication models. calibration.Page 11 of 30
The Principles of Modeling
• • •
Physics is about the future.
•
A model is only a model: capture the important features. – If you can’t do either of those. find the market price of something similar and liquid. – Similar: same payoffs under all (?) circumstances.
Stanford. Absolute valuation – Newton in physics. no such thing in finance Relative valuation: The only reliable law of finance – If you want to know the value of something (illiquid). – static replication is best – needs no systems.Smile. not all of them. 2006
.fm
October 21.

-.Smile. 2006
.fm
October 21. puts and calls with a range of strikes. For terminal payoffs:
S
S
• • •
Any of these payoffs can be exactly duplicated by a linear combination of forwards.
Stanford. and hence exactly replicated at a cost known if we know the volatility smile. volatility. bonds.only problem is counterparty risk.Page 12 of 30
Static Replication of Exotic Options
•
The best way to handle valuation in the presence of a smile is to use no model at all. etc. Try to do approximately the same for other exotic options. Replication will hold irrespective of jumps.

t )dt + σS dZ but they are parametric and cannot match the implied volatility surface
β
Stanford. t. does the local volatility describe the asset? Why would local volatility make sense: – Leverage effect: as assets approach liabilities. Can still replicate options. t )dt + σ ( S.1. t ) is a deterministic function of a stochastic variable S.= μ ( S. equity volatility increases – Fear: increase in implied volatility as equity declines – CEV models: dS = μ ( S.2
• σ ( S.
Eq.fm
October 21. t ) ↔ Σ ( S. still do • •
Calibration: σ ( S. t )dZ S
risk-neutral pricing. 2006
. T ) ? Even if there is a solution. K.Smile.Page 13 of 30
Dynamic Replication: Local Volatility Models
dS ----.

fm
October 21. K ) ≈ -----------------------------2
∂Σ ∂S
∼
S=K
∂Σ ∂K
1 ∼ -.Page 14 of 30
The Smile in Local Volatility Models: First Look
• •
Interest rates: long rates are averages of expected future short rates between today and maturity In local volatility models implied volatilities are “averages” of expected future volatilities between spot and strike. 2006
. between today and expiration.Smile. K ) = average vol between S and K
strike K local volatility σ(S.t) spot S
T
σ(S) + σ(K) Σ ( S.
Stanford.
Σ ( S.σ ( S ) 2 S=K
Local volatilities predict \ the change of implied volatility with stock price. and that local volatility varies twice as fast with spot as implied volatility varies with strike.

Smile.
Σ ( 80. K )
Σ ( 80. 80 )
Σ ( 100.Page 15 of 30
Local Volatility Models …
Σ ≈ Σ(K + S)
Let’s draw this for a roughly linear skew in the following figure.fm
October 21. 100 )
S = 80 S = 100 K
80
100
Smile moves up when stock price moves down. 80 )
Σ ( S. 100 ) Σ ( 100. 2006
.
Stanford.

2006
.fm
October 21.Page 16 of 30
The Relation Between Local and Implied Vols
Think of implied volatilities as a complex probe/X-ray of the interior of the local volatility surface.
Stanford. Can one deduce local volatilities from implied?
This is an inverse scattering problem.Smile.

2006
.fm
October 21.Page 17 of 30
Inverse Scattering: Schematic Calibration
Derman-Kani binomial method
Stanford.Smile.

T ))
K
We can go further than this.Smile. the local volatility at S T = K can be determined from the derivatives of option prices:
∂ C ( S. T ) --------------------.= -----------------------------------2 2 2∂ C K 2 ∂K
2
calendar spread ~ variance x prob butterfly spread ~ variance
Stanford. K. T ) ∂T σ ( K. T ) = e
r(T – t)
∂
2 2
∂K
(C ( S. t. 2006
. t ) from market prices of options. t. and actually find the local volatility σ ( S.Page 18 of 30
Dupire Continuous-Time Equation
Recall the Breeden-Litzenberger formula: risk-neutral probability/Arrow-Debreu price is a butterfly spread:
p ( S. K. t.fm
October 21. For zero interest rates and dividends. K.

fm
October 21.05
local vol vs.25 0.Smile.1 0.15 0.2 0. strike K
spot or strike
60 80 100 120 140
Stanford.Page 19 of 30
Example: Local Volatility Extraction
Local volatilities are the analogs of forward/sideways rates
local vol vs. strike
Implied and Local Vols at 1 years 0. 2006
. spot S implied vol vs. spot S
implied vol vs.

we don’t have a continuous implied volatility surface so we cannot use these relations blindly.Page 20 of 30
Local Volatility …
• •
In practice.Smile.∫ ---------. Nice approximate harmonic average relation for short expirations due to Roger Lee:
x = ln ( S ⁄ K ) 1 1 1 ---------.dy x σ(y) Σ(x)
o x
Stanford.= -. Needs more sophisticated methods.fm
October 21. 2006
.

V. V.fm
October 21. V.Smile.difficult! Can one really know the stochastic PDE for volatility? A “tall order.Page 21 of 30
Stochastic Volatility Models
dS = μ S ( S. Option prices are risk-neutral. t )dZ t dV = μ V ( S. t )dt + σ V ( S. t )dW t V = σ E [ dWdZ ] = ρdt
2
• • • • •
Can replicate only if you can trade “volatility”/options as well as stock. 2006
.
Stanford. t )dt + σ S ( S. “volatility” is not -. You must hedge with stock and options . V.need market price of risk for volatility.” (Rebonato). Correlation is even more stochastic than volatility.

Page 22 of 30
The Smile in Stochastic Volatility Models
Poor man’s point of view: Black-Scholes with the volatility in one of two regimes: Take expectation of prices:
1 C SV = -. K. Σ ) ⎝ S⎠
So
K Σ = g ⎛ ---⎞ ⎝ S⎠
∂Σ ∂Σ at the money ≈– ∂S ∂K
Just the inverse of the relation in local volatility models. σ H ) + C BS ( S. K. we can write
K C SV = Sf ⎛ ---⎞ ≡ SC BS ( S.fm
October 21. σ L ) ] 2
σH σL
Because the Black-Scholes equation is homogeneous in S and K. 2006
.[ C BS ( S.
Stanford.Smile. K.

if volatility doesn’t change. 100 )
S = 100 S = 80 120 K
80
100
Stanford.
Σ ( S. 120 ) Σ ( 80. K )
Σ ( 100.fm
October 21. 80 ) Σ ( 100. 80 )
Σ ( 100. 2006
.Smile. 100 ) Σ ( 80.Page 23 of 30
The Smile in Stochastic Volatility Models …
Σ ≈ Σ(S – K)
Smile moves up when stock moves up.

t ) – rV = 0 σ S + -q + 2 ∂ σ2 ∂ S ∂σ ∂S ∂σ ∂ t 2 ∂ S2
Risk neutrality: volatility drift or market price of risk φ is determined by options prices being riskneutral.σd
2
2
Heston.
Stanford. etc.σu Su.Page 24 of 30
Stochastic Volatility PDE
Volatility is mean-reverting: dσ
2
= α ( m – σ )dt + ξ σ dW r
Su.fm October 21. 2006
2
2
2
. σ.σd
∂V ∂V ∂V 1 ∂ V 2 2 1 ∂ V 2 ∂ V + -σqSρ + rS + φ ( S.Smile.σu Sd.σ
Sd.
S.

Page 25 of 30
Stochastic Volatility Characteristic Solution
1 2 σ T = -. σT.Smile.∫ σ t dt T
0 2 T
average variance along a path
V =
∑ p ( σT ) × B S ( S. r. T )
σT
the Mixing Formula for ρ = 0
σH σL
Stanford. K. 2006
.fm
October 21.

fm
October 21. Δ < Δ BS when ρ < 0
Imp Vol ρ = -1 ρ > -1
ρ=0
•
Local volatility is the average of all stochastic volatilities at a definite future stock price and time. then the optimal pure stock hedge ratio looks a lot like the local volatility hedge ratio. When ρ ≠ 0 smile is asymmetric.e.Smile. If you cannot hedge with options.Page 26 of 30
Stochastic Volatility Smiles:
• • • • • •
symmetric for ρ = 0 . i.
strike
Stanford. flat for long expiration as mean reversion wins out. highly curved for short expirations. 2006
. because of fat tails. You need very high volatility of volatility to get a steep short-term smile.

S+J
λ
S 1.fm
October 21. stock prices move discontinuously. K.Page 27 of 30
Jump-Diffusion Models
In fact. Merton: Poisson distribution of unhedgeable but diversifiable jumps. σ + ---------. r + --------------------------.– λ ⎜ ⎟⎟ n! τ τ ⎜ ⎝ ⎠⎟ n=0 ⎝ ⎠
Stanford. τ. Geometric Brownian motion isn’t the right description. but more seriously.λ S–K
BS
BS
C JD
2 ⎛ J + 1 σ 2⎞ ⎛ 2 -n ∞ ⎛ J + 1σ ⎞⎞ n -⎝ ⎜ 2 J⎠ – λτ 2 nσ J ( λτ ) ⎜ e 2 J – 1⎟ ⎟ = e ∑ -------------CBS ⎜ S.Smile. not only is volatility stochastic. 2006
.

The smiles in a stochastic volatility model are more pronounced at longer expirations.
Stanford. has difficulty ln(K/S) producing a steep short-term smile because the volatility diffuses.05 0.265 0.25 0.27 0.
A higher jump frequency helps produce a steeper smile at expiration.4" "100"
σ τ overwhelms the jumps.1" "0.Page 28 of 30
Jump-Diffusion Smiles
• • •
Steep realistic short-term smiles from the instantaneous jump. and therefore doesn’t change too much initially.24 0. 2006
.fm
October 21.26 0.29
Implied volatility under jump diffusion
Long-term smiles tends to be flat as diffusion
implied volatility
0. because the jumps tends to happen more frequently and therefore are more likely to occur in the future as well.
0.245 0.28 0.1 0 0.Smile.275 0.1 Stochastic volatility.235 "0.
2
•
-0. in contrast.255 0.1 -0.285 0.

fm
October 21. 2006
.Page 29 of 30
Pragmatic Interpolation Methods
•
Practitioners think of options models as interpolating formulas that take you from known prices of liquid securities to the unknown values of illiquid securities: – options. Find a static hedge composed of a portfolio of vanilla options that approximately replicates the payoff or the vega of the exotic options.Smile.
• • •
An exotic option is a mixture of ordinary options. – convertible bonds. Then figure out the value of the portfolio of vanilla options in a skewed world.
Stanford.

The best you can do is pick a model that mimics the most important behavior of the underlyer in your market.Page 30 of 30
Summary
• • • • •
There is no “right” model.Smile. Try to be as model independent as possible.fm
October 21. Examine the value of an option in a variety of plausible models. Then add perturbations if necessary. There is lots of work to be done on modeling realistic distributions and the options prices they imply. 2006
.
Stanford.