Bruno Dupire

Bloomberg LP
Lecture 9
Volatility

Forward Equations (1)

BWD Equation:
price of one option C (K 0 ,T0 ) for different (S, t )
FWD Equation:
price of all options C (K , T ) for current (S 0 ,t0 )
Advantage of FWD equation:
If local volatilities known, fast computation of implied
volatility surface,
If current implied volatility surface known, extraction of
local volatilities,
Understanding of forward volatilities and how to lock
them.
Bruno Dupire

Forward Equations (2)

Several ways to obtain them:
Fokker-Planck equation:
Integrate twice Kolmogorov Forward Equation

Tanaka formula:
Expectation of local time

Replication
Replication portfolio gives a much more financial
insight

Bruno Dupire

Fokker-Planck
If dx = b( x, t )dW
1 2 (b 2 )
Fokker-Planck Equation:
=
t 2 x 2

2C
Where is the Risk Neutral density. As =
K 2
2
2

C
C
2
2
2 C

b

2
2
t = K = 1 K
x 2
t
x 2
2

Integrating twice w.r.t. x: C b2 2C

=
2
t 2 K
Bruno Dupire

FWD Equation: dS/S = (S,t) dW

T

Define CS K ,T

C K ,T + T C K ,T

T
CK ,T +T

CK ,T

ST

CS K T,T at T

dT
dT/2

2 (K , T )

T 0

ST

ST

K 2 K ,T

2
2
(
)
C
K
,
T
C

2
Equating prices at t0:
K
=
T
2
K 2

Bruno Dupire

FWD Equation: dS/S = r dt + (S,t) dW

TV

CS K T,T at T = Time Value + Intrinsic Value

S T Ke r T

IV

ST K

(Strike Convexity)

(Interest on Strike)

CK ,T
CK ,T +T

KerT

ST

2 (K , T )
2
TV

KerT K

IV

T 0

rK

rKDigK ,T
ST

Equating prices at t0:

Bruno Dupire

K 2 K ,T

ST

C 2 (K , T ) 2 2C
C
K
=
rK
2
T
2
K
K
6

FWD Equation: dS/S = (r-d) dt + (S,t) dW

ST KerT

TVIV

CS KT,T at T =

ST e dT Ke rT

CK,T+T

(dr)T

Ke

CK,T

TV + Interests on K
Dividends on S

ST

2 (K , T )
2

T 0

(r d )K
Ke(d r )T K

d CK ,T

T
Bruno Dupire

(r d ) K DigK ,T

ST

Equating prices at t0: C

K 2 K ,T

2 (K , T )
2

ST

d CK ,T

2C
C
(r d )K
K
d C
2
K
K
2

Stripping Formula
C 2 (K , T )K 2 2C
C
=
(r d )K
d C
2
T
K
K
2
If (K , T ) known, quick computation of all CK ,T (S 0 ,t0 )
today,
If all CK ,T (S0 ,t0 ) known:
(K , T ) =

C
C
+ (r d )K
+ dC
T
K
2
2 C
K
K 2

Local volatilities extracted from vanilla prices

and used to price exotics.
Bruno Dupire

Smile dynamics: Local Vol Model (1)

Consider, for one maturity, the smiles associated
to 3 initial spot values
Skew case

Local vols

Smile S
Smile S 0
Smile S +

S S0 S +

ATM short term implied follows the local vols

Similar skews
Bruno Dupire

Smile dynamics: Local Vol Model (2)

Pure Smile case
Local vols

Smile S +
Smile S

Smile S 0

S0

S+

ATM short term implied follows the local vols

Skew can change sign
Bruno Dupire

10

Summary of LVM Properties

0 is the initial volatility surface
(S,t ) compatible with 0 =

( ) compatible with E[
0

k,T

Bruno Dupire

local vol

ST = K ] = (local vol)

deterministic function of (S,t)

future smile = FWD smile from local vol

11

Volatility Replication

Volatility Replication
dS
= t dW
S

Apply Ito to f(S,t).

df = f S dS + f t dt + 12 f SS t2 S 2 dt
T

T
T

f SS ( St , t ) S dt = 2 f ( ST , T ) f ( S 0 ,0) f t ( St , t )dt f S ( St , t )dSt

0
0

2
t

European PF

2
To replicate g ( S , t ) dt ,find f : g ( S , t ) = f SS ( S , t ) S
2
t

Bruno Dupire

-hedge

f =

g
S2
13

Examples
Variance Swap
Corridor
Variance Swap
FWD Variance
Swap
Absolute
Variance Swap
Local Time at
level K
Bruno Dupire

g (S , t ) = 1
g ( S , t ) = 1[ a ,b ] ( St )

g ( S , t ) = 1[T1 ,T2 ] (t )

S
f ( S , t ) = ln( )
S0

S
) on [a,b]
S0
+ linear extrapolation

f ( S , t ) = ln(

S
f (S, t) = ln( )1[T1,T2 ] (t)
S0

g (S , t ) = S 2

(S S0 ) 2
f (S , t ) =
2

g (S , t ) = K (S )

(S K )+
f (S , t ) =
K2
14

Conditional Instantaneous FWD

Variance
From local time:

Differentiating wrt T:

T 2

C(K ,T )
E t K ( S )dt = 2
K2
0

] [

E T2 K ( ST ) = E T2 ST = K E [ K ( ST )] =

2 C

(K ,T )
2
T
K

2C
E [ K ( ST )] =
(K ,T )
2
K

And, as:

C
(K ,T )
2
2
= loc
(K ,T )
E T2 ST = K = 2 2T
C
K
(K ,T )
K 2

Bruno Dupire

15

Deterministic future smiles

It is not possible to prescribe just any future
smile
If deterministic, one must have
C K ,T (S 0 , t 0 ) = (S 0 , t 0 , S , T1 ) C K ,T (S , T1 )dS
2

Not satisfied in general

K
S0

t0

Bruno Dupire

T1

T2

16

Det. Fut. smiles & no jumps

=> = FWD smile
2
2
If (S , t , K , T ) / VK ,T (S , t ) (K , T ) lim imp (K , T , K + K , T + T )

K 0
T 0

stripped from SmileS.t

Then, there exists a 2 step arbitrage:

Define
2C
2

( (K , T ) V (S , t )) K (S , t , K , T )

PL t

K ,T

At t0 : Sell PL t Dig S ,t Dig S + ,t

At t: if S [S , S + ]
t

buy

S0

t0

2
CS K, T , sell
2
K

t
2

(K , T ) K ,T

gives a premium = PLt at t, no loss at T

Conclusion: VK ,T (S , t ) independent of
from initial smile
Bruno Dupire

(S , t ) = VK ,T (S0 , t0 ) = 2 (K , T )
17

Consequence of det. future smiles

Sticky Strike assumption: Each (K,T) has a fixed impl ( K , T )
independent of (S,t)
Sticky Delta assumption: impl ( K , T ) depends only on
moneyness and residual maturity

In the absence of jumps,

Sticky Strike is arbitrageable
Sticky is (even more) arbitrageable

Bruno Dupire

18

Example of arbitrage with Sticky Strike

Each CK,T lives in its Black-Scholes ( impl ( K , T ) )world

C1 C K 1 ,T1

assume 1 > 2

C 2 C K 2 ,T2

P&L of Delta hedge position over dt:

PL (C 1 ) =

1
2

PL (C 2 ) =

1
2

(( S ) S t )
(( S ) S t )
2

1C 2

12 2 2
PL (1C 2 2 C 1 ) =
S 1 22 t > 0
2
(no , free )

!
Bruno Dupire

2 C 1

S t1

S t + t

If no jump

19

Arbitraging Skew Dynamics

In the absence of jumps, Sticky-K is arbitrageable and Sticky- even more so.
However, it seems that quiet trending market (no jumps!) are Sticky-.
In trending markets, buy Calls, sell Puts and -hedge.
Example:
K1

PF C K 2 PK1
S

K2

1 , 2
VegaK > Vega K
2

1 , 2
VegaK < Vega K
2

Bruno Dupire

St

PF
1

PF

-hedged PF gains
from S induced
volatility moves.

20

Skew from Historical Prices

Theoretical Skew from Prices

?

Problem : How to compute option prices on an underlying without options?

For instance : compute 3 month 5% OTM Call from price history only.
1) Discounted average of the historical Intrinsic Values.
Bad : depends on bull/bear, no call/put parity.
2) Generate paths by sampling 1 day return recentered histogram.
Problem : CLT converges quickly to same volatility for all strike/maturity;
breaks autocorrelation and vol/spot dependency.
Bruno Dupire

22

Theoretical Skew from Prices (2)

3) Discounted average of the Intrinsic Value from recentered 3 month
histogram.
4) -Hedging : compute the implied volatility which makes the -hedging a
fair game.

Bruno Dupire

23

Theoretical Skew
from historical prices
How to get a theoretical Skew just from spot price
history?
S
K
Example:
ST
3 month daily data
t
T2
T1
1 strike K = k ST1
a) price and delta hedge for a given within Black-Scholes
1

model
b) compute the associated final Profit & Loss: PL( )
c) solve for k / PL k = 0
d) repeat a) b) c) for general time period and average
e) repeat a) b) c) and d) to get the theorical Skew

Bruno Dupire

( )

( ( ))

24

IV. Volatility Expansion

Introduction
This talk aims at providing a better
understanding of:
How local volatilities contribute to the value of
an option
How P&L is impacted when volatility is
misspecified
Link between implied and local volatility
Smile dynamics
Vega/gamma hedging relationship
Bruno Dupire

26

Framework & definitions

In the following, we specify the dynamics
of the spot in absolute convention (as
opposed to proportional in Black-Scholes)
and assume no rates:

dSt = t dWt

: local (instantaneous) volatility

(possibly stochastic)
Implied volatility will be denoted by $
Bruno Dupire

27

P&L of a delta hedged option

Option Value

P&L

Break-even
points
Delta
hedge

Ct + t

Ct

St

Bruno Dupire

S t + t

St

28

P&L of a delta hedged option (2)

Correct

Volatility higher than

1st

1st
1st

St

1st

S t + t

P&L

St

S t + t

P&L
Expected P&L = 0

Expected P&L > 0

Ito: When t 0 , spot dependency disappears

Bruno Dupire

29

Black-Scholes PDE
P&L is a balance between gain from and
P& L( t ,t +dt )

2
= 0 + 0 dt

From Black 02
0
Scholes PDE: 0 =

=> discrepancy if different from

1 2
gain over dt = ( 02 )0dt
2
> 0: Profit
Magnified by 0
< 0: Loss
Bruno Dupire

30

P&L over a path

Total P&L over a path
= Sum of P&L over all small time intervals

Gamma

P& L =
1 T ( 2 2 ) dt
0
0
2 0
No assumption is made
on volatility so far

Time
Spot

Bruno Dupire

31

General case
Terminal wealth on each path is:
T
1
wealth T = X ( 0 ) + ( 2 02 ) 0 dt
2 0
( X ( ) is the initial price of the option)
0

Taking the expectation, we get:

T
1
E wealthT = X (0 ) + 2 E[0 ( 2 02 )| S] dSdt
0 0

The probability density may correspond to the density of

a NON risk-neutral process (with some drift) with volatility .
Bruno Dupire

32

Non Risk-Neutral world

In a complete model (like Black-Scholes), the drift does
not affect option prices but alternative hedging strategies
lead to different expectations
Example: mean reverting process
towards L with high volatility around L
We then want to choose K (close to L)
T and 0 (small) to take advantage of it.
L

In summary: gamma is a volatility

collector and it can be shaped by:
a choice of strike and maturity,

Profile of a call (L,T) for

different vol assumptions

a choice of 0 , our hedging volatility.

Bruno Dupire

33

Average P&L
From now on, will designate the risk neutral density
associated with dSt = dWt .
In this case, E[wealthT] is also X ( ) and we have:
T
1
X ( ) = X ( 0 ) + 2 E [ 0 ( 2 02 )| S ] dS dt
0 0

Path dependent option & deterministic vol:

X ( ) = X ( 0 ) + 21 ( 2 02 ) E [ 0 | S ] dS dt

European option & stochastic vol:

C ( ) = C ( 0 ) + 21 ( E [ 2 | S ] 02 ) 0 dS dt
Bruno Dupire

34

Quiz
Buy a European option at 20% implied vol
Realised historical vol is 25%
Have you made money ?
Not necessarily!
High vol with low gamma, low vol with high gamma

Bruno Dupire

35

Expansion in volatility
An important case is a European option with
deterministic vol:

C ( ) = C ( 0 ) + 21 ( 2 02 ) 0 dS dt
The corrective term is a weighted average of the volatility
differences
This double integral can be approximated numerically

Bruno Dupire

36

P&L: Stop Loss Start Gain

Extreme case: 0 = 0 0 = K
T
2
1
C ( ) = ( S 0 K ) + 2 ( K , t ) ( K , t )dt
0
+

This is known as Tanakas formula

S
Delta = 100%

K
Delta = 0%

t
Bruno Dupire

37

Local / Implied volatility relationship

Differentiation
Local volatility

Implied volatility
31

23

27

21

23

19

19

17

15

15

11

13

strike

maturity

spot

time

Aggregation
Bruno Dupire

38

Smile stripping: from implied to local

Stripping local vols from implied vols is the
inverse operation:

C
T
2
(S ,T ) = 2 2
C
K2

(Dupire 93)

Involves differentiations

Bruno Dupire

39

From local to implied: a simple case

Let us assume that local volatility is a
deterministic function of time only:

dSt = ( t ) dWt
In this model, we know how to combine local
vols to compute implied vol:
T

$ ( T ) =

0 ( t )dt

Question: can we get a formula with

Bruno Dupire

(S , t ) ?
40

From local to implied volatility

When 0 = implied vol

1
2
2
(

0 )0 dSdt = 0

depends on 0

0 dSdt

2
0 =
0 dSdt

solve by iterations

Implied Vol is a weighted average of Local Vols

(as a swap rate is a weighted average of FRA)

Bruno Dupire

41

Weighting scheme
Weighting Scheme: proportional to 0

At the
money
case:

Out of the
money
case:
0

S
S0=100
K=100
Bruno Dupire

S0=100
K=110

42

Weighting scheme (2)

Weighting scheme is roughly proportional to
the brownian bridge density

Brownian bridge density:

BB K ,T ( x, t ) = P St = x ST = K

Bruno Dupire

43

Time homogeneous case

2
2
$
= ( S ) ( S )dS

ATM (K=S0)
(S)

dt

( S) =
dSdt
0

OTM (K>S0)

(S)

S0

S0

small
S

T
large

(S)

(S)

S0

S
Bruno Dupire

S0

S
44

Link with smile

$ K

and

$ K

K2

are

averages of the same local

S0

vols with different weighting

K1

schemes

=> New approach gives us a direct expression for

the smile from the knowledge of local volatilities
But can we say something about its dynamics?
Bruno Dupire

45

Smile dynamics
Weighting scheme imposes
some dynamics of the smile for
a move of the spot:
For a given strike K,

S1
S0

S $ K

(we average lower volatilities)

Smile today (Spot St)
&
Smile tomorrow (Spot St+dt)
in sticky strike model

t
26

25.5
25

Smile tomorrow (Spot St+dt)

if ATM=constant

24.5

Smile tomorrow (Spot St+dt)

in the smile model

23.5

Bruno Dupire

24

St+dt

St
46

Sticky strike model

A sticky strike model ( $ K ( t ) = $ K ) is arbitrageable.
Let us consider two strikes K1 < K2

1 dt

K 1)
1C(

2 dt

1C
(

2)

K 1)
1C(

1 /
2 *
C(K
2)

The model assumes constant vols 1 > 2 for example

2 dt

1 dt

By combining K1 and K2 options, we build a position with no gamma and

positive theta (sell 1 K1 call, buy 1/2 K2 calls)
Bruno Dupire

47

Vega analysis
If

& 0 are constant

1 2
C( ) = C( 0 ) + ( 02 ) 0 dSdt
2

= 02 +

1
C( + ) = C( ) + 0dSdt
2
{
2
0

2
0

2C
2
C C 2 C
=
=
Vega =
2
2 2

Bruno Dupire

48

Gamma hedging vs Vega hedging

Hedge in
insensitive to realised
historical vol
If =0 everywhere, no sensitivity to
historical vol => no need to Vega hedge
Problem: impossible to cancel now for
the future
Need to roll option hedge
How to lock this future cost?
Answer: by vega hedging
Bruno Dupire

49

Superbuckets: local change in local vol

For any option, in the deterministic vol case:
X ( ) = X ( 0 ) + 21 ( 2 02 ) E[0 | S ] dS dt
For a small shift in local variance around (S,t), we have:

X ( ) = X ( 0 ) + 21 E[ 0 | S ]

dX

d (2S ,t )

= 21 E[ 0 ( S , t )| S ] ( S , t )

For a european option:

Bruno Dupire

dC

( )

d (2S ,t )

= 21 0 ( S, t )( S, t )
50

Superbuckets: local change in implied vol

Local change of implied volatility is obtained by combining
local changes in local volatility according a certain weighting

dC
dC d ( 2 )
=
2
d ($ )
d ( 2 ) d ($ 2 )
sensitivity in
local vol

Thus:
<=>
<=>
Bruno Dupire

weighting obtain
using stripping
formula

cancel sensitivity to any move of implied vol

cancel sensitivity to any move of local vol
cancel all future gamma in expectation
51

Conclusion
This analysis shows that option prices are
based on how they capture local volatility
It reveals the link between local vol and
implied vol
It sheds some light on the equivalence
between full Vega hedge (superbuckets)
and average future gamma hedge
Bruno Dupire

52

Delta Hedging
We assume no interest rates, no dividends, and absolute
(as opposed to proportional) definition of volatility
Extend f(x) to f(x,v) as the Bachelier (normal BS) price of
f for start price x and variance v:
with f(x,0) = f(x) f ( x, v) E x ,v [ f ( X )] 1
2v
Then,

f ( y )e

( y x)2
2v

dy

1
We explore
f v ( xvarious
, v) = f xxdelta
( x, v) hedging strategies
2

Bruno Dupire

53

Calendar Time Delta Hedging

Delta hedging with constant vol: P&L depends on the
path of the volatility and on the path of the spot price.
Calendar time delta hedge: replication cost of
f ( X t , 2 .(T t ))
1 t
2
f
X

T
+
f
(
dQV

ducost
)
(
,
.
)
0
0 ,u
In particular, for
sigma2 =
of
0 0,xx replication
2

f (Xt )
1 t
f ( X 0 ) + f xx dQV0,u
2 0

Bruno Dupire

54

Business Time Delta Hedging

Delta hedging according to the quadratic variation:
P&L that depends only on quadratic variation and
spot price
1
df ( X t , L QV0,t ) = f x dX t f v dQV0,t +

f xx dQV0,t = f x dX t

QV0,T L,

Hence, for

f ( X t , L QV0,t ) = f ( X 0 , L) + f x ( X u , L QV0,u ) dX t
0

f ( X t , L QV0,t )

f ( X 0 , L)

f (And
X 0 , Lthe
) replicating cost of

is
: QV0, = L
finances exactly the replication of f until

Bruno Dupire

55

Daily P&L Variation

Bruno Dupire

56

Tracking Error Comparison

Bruno Dupire

57

V.

Stochastic Volatility Models

Hull & White

Stochastic volatility model Hull&White (87)
dS t
= rdt + t dW
St
d

= dt + dZ

P
t
P
t

Implied volatility

Implied volatility

Incomplete model, depends on risk premium

Does not fit market smile

Strike price

Z ,W = 0
Bruno Dupire

Strike price

Z ,W < 0
59

Role of parameters
Correlation gives the short term skew
Mean reversion level determines the long term
value of volatility
Mean reversion strength
Determine the term structure of volatility
Dampens the skew for longer maturities

Volvol gives convexity to implied vol

Functional dependency on S has a similar effect
to correlation
Bruno Dupire

60

Heston Model
dS
S = dt + v dW

dv = (v v )dt + v dZ

dW , dZ = dt

Solved by Fourier transform:

FWD
x ln
=T t
K
C K ,T ( x, v, ) = e x P1 ( x, v, ) P0 ( x, v, )

Bruno Dupire

61

Spot dependency
2 ways to generate skew in a stochastic vol model
1) t = xt f (S , t ), (W , Z ) = 0
2)

(W , Z ) 0

S0

ST

ST
S0

-Mostly equivalent: similar (St,t ) patterns, similar

future
evolutions
-1) more flexible (and arbitrary!) than 2)
-For short horizons: stoch vol model local vol model
+ independent noise on vol.
Bruno Dupire

62

Convexity Bias
dS = t dW
2
d t = dZ
(W , Z ) = 0

E t2 | S t = K = 02 ?

[ ]

NO! only E t2 = 02

E t2 | St = K

02
S0

t likely to be high if St >> S0 or St << S0

Bruno Dupire

63

Impact on Models
Risk Neutral drift for instantaneous forward
variance
Markov Model:
dS
= f (S , t ) t dW fits initial smile with local vols (S , t )
S
2 (S , t )
f (S , t ) =
E [ t2 | S t = S ]

Bruno Dupire

64

Smile dynamics: Stoch Vol Model (1)

Local vols

Skew case (r<0)

Smile S
Smile S 0
Smile S +

S+

S S0

- ATM short term implied still follows the local vols

(E [

S T = K = 2 (K , T )

- Similar skews as local vol model for short horizons

- Common mistake when computing the smile for another
spot: just change S0 forgetting the conditioning on :
if S : S0 S+ where is the new ?
Bruno Dupire

65

Smile dynamics: Stoch Vol Model (2)

Pure smile case (r=0)

Local vols

Smile S

Smile S

Smile S 0

S0

S+

ATM short term implied follows the local vols

Future skews quite flat, different from local vol
model
Again, do not forget conditioning of vol by S
Bruno Dupire

66

Forward Skew

Forward Skews
In the absence of jump :
model fits market

K , T

2
E[ T2 ST = K ] = loc
(K ,T )

This constrains
a) the sensitivity of the ATM short term volatility wrt S;
b) the average level of the volatility conditioned to ST=K.
a) tells that the sensitivity and the hedge ratio of vanillas depend on the
calibration to the vanilla, not on local volatility/ stochastic volatility.
To change them, jumps are needed.
But b) does not say anything on the conditional forward skews.

Bruno Dupire

68

Sensitivity of ATM volatility / S

At t, short term ATM implied volatility ~ t.
As t is random, the sensitivity

Et [

2
t +t

2
S

is defined only in average:

2
loc
(S , t )
St = St + S ] = ( St + S , t + t ) ( St t )
dS
S
2
t

2
loc

2
loc

2
2
In average, ATM
follows loc
.

Optimal hedge of vanilla under calibrated stochastic volatility corresponds to

perfect hedge ratio under LVM.

Bruno Dupire

69