Danske Bank

Bruno Dupire –
Shaping derivatives markets for 30 years
Antoine Savine
 Danske Bank

Introduction: Pillars of Modern Finance

From Black and Scholes (1973) to Dupire (1992-1996)
 Danske Bank

Defining works of modern finance

• Black and Scholes (1973)
Condition for absence of arbitrage

C  2C 2 2 
=−  St or t = − t  2 St 2
t 2S 2
2

• Heath-Jarrow-Morton (1992)
Condition for respect of initial yield curve

df ( t , T ) =  ( t , T ) (   (t, u ) du − dW )
T

t
Q

• Dupire (1992-1996)
Condition for respect of initial call prices

2C  2C 2C T
E  ST  T
2 2
ST = K  = = = “2 calendar spreads over 1 butterfly”
T K 2 CKK = forward variance (in normal terms)
= conditional variance tradable through European options
 Danske Bank

Black and Scholes (1973)

• Ground breaking paradigm: replication: options are hedged with trading strategies
Using ”Greek” notations   C S ,    2C S 2 ,  C t in a simplified world where rates = dividends = 0

 
 ( )
From Ito’s lemma: dC ( St , t ) = dSt +   −  t  dt after delta-hedge: dC ( St , t ) − dSt =   −  ( t )  dt = 0
  
2 2
dS dS
 2 dt   2 dt  AF
 
  2 St 2 

dSt
 = 0 dt + dWt
• Black & Scholes arbitrage-free PDE: t = − t  t 2 St 2
2
and by Feynman-Kac: Ct =E St
CT St 

• (Rather simplistic) model: volatility is known

Ct = E ( ST − K ) St  = St N ( d1 ) − KN ( d 2 )
+
The solution is analytic:  
Replication startegy: hold N ( d1 ) stocks, borrow KN ( d 2 ) dollars
Hence value: St N ( d1 ) − KN ( d 2 )
 Danske Bank

Dupire (1992-1996)
• Ground breaking paradigm: calibration: models must respect market prices of calls
2C
− Necessary and sufficient condition (Dupire, Unified Theory of Volatility, 1996): E  ST 2 T 2 ST = K  = T
CKK
− Applies to a wide a class of diffusion models - demonstration:
1
By Tanaka’s formula: d ( ST − K ) = 1S
+
T  K
dST +  ST − K ST 2 T 2 dT
2

1
Taking (RN) expectations on both sides: dE Q ( ST − K )  = 0 + E Q  S
+
−K ST 2 T 2  dT
2 T

dE Q ( ST − K ) 
+
  = E Q  2CT
ST − K ST  T  = E  ST − K E  ST  T ST   = q ( K , T ) E  ST  T ST = K   E  ST  T ST = K  =
So: 2 2 2
 Q
 Q
 2 2  Q
 2 2  Q
 2 2 
dT  CKK
=CKK
CT

2C T
− Define forward variance:  f 2 ( K , T )  then the condition for calibration is written: E  T 2 ST = K  =  f 2 ( K , T )
K 2CKK

• Local volatility model: simple case where volatility is a known function of S:  t   ( St , t )

2C
Then the condition becomes (Dupire, Pricing with a Smile, 1992):  ( K , T ) =  ( K , T )  K C 2
f
2
2
T

KK
 Danske Bank

Calibration or estimation?
• Like Black-Scholes, Dupire’s work is about the approach, not the model or the formula
• Black and Scholes established the key concept of replication
But left open the question of the parameters
• Practical application of a derivatives model
(with modern lingo borrowed from Machine Learning):
− First, learn model parameters from market data
− Then, apply the model to a more complicated risk management problem: OTM option, exotic option, CVA, ...

• Calibration approach: imply parameters from market prices of Europeans

− Natural approach for every derivatives professional

• Estimation aproach: statistically estimate parameters from historical data

− Natural approach for everyone else

• Why is calibration the natural choice on capital markets?

 Danske Bank

Calibration and Risk Management

From the Fundamental Theorem of Derivatives Trading to Dupire’s Sigma-Zero formula
 Danske Bank

The Fundamental Theorem of Derivatives Trading

• Address pricing problem from hedging and risk management perspective
• Assuming diffusion price process (ignoring jumps)
• Buy C(K,T) and risk manage with Black and Scholes and some implied volatility −hat
• What are the residual risks?

ˆ
• Apply Ito to Black and Scholes’s valuation formula: dCˆ ( S , t ) = ˆ dS + ˆ dt + 2 ( dS )t t t t
t
t
2
,  t : realized volatility
 ˆ   t St dt
2 2

• After delta-hedge: dCˆ ( St , t ) − ˆ t dSt =  ˆt + t  t 2 St 2  dt

 2 
• Applying Black and Scholes PDE:
ˆ t 2 2 ˆ
ˆt = − ˆ St so dCˆ ( St , t ) − ˆ t dSt = t ( t 2 − ˆ 2 ) St 2 dt
2 2

• In English:
mis-replication ~ realised – implied variance
 Danske Bank

The Fundamental Theorem of Derivatives Trading

Cˆ ( ST , T ) = Cˆ ( S0 , 0 ) +  ˆ t dSt +  ˆ t ( t 2 − ˆ 2 ) St 2 dt
T 1 T
• Integrating over hedge period [0,T]: payoff premium
0 2 0
delta − hedge mis − replication
textbook − replication

• In English:
− Delta-hedging is only risk-free when model correctly predicts realized volatility
− PnL has an additional term: weighted sum (weighted by gamma) of realized – predicted (normal) variance
− Replication error = finite variation process => slowly ”bleeds” but does not ”explode”

• This was called ”Robustness of Black-Scholes” (El-Karoui and al, 1998)

and ”Fundamental Theorem of Derivatives Trading” (Poulsen and al, 2015)
• This is a central formula for understanding and managing derivatives risk
• Bruno Dupire heavily relied on this approach since the early 1990s
 Danske Bank

The Sigma-Zero formula (Dupire, 1996)

Cˆ ( ST , T ) = Cˆ ( S0 , 0 ) +  ˆ t dSt +  ˆ t ( t 2 − ˆ 2 ) St 2 dt
1 T
• Consider the FTDT: T

0 2 0

• Applying expectations (under Q = risk neutral real world probability):

 
 t   ˆ
 
E Q C ( ST , T )  = C ( S0 , 0 ) + 0 + E Q   (  t 2 − ˆ 2 ) St 2 dt   E Q Cˆ ( ST , T )  − Cˆ ( S0 , 0 ) = E Q   t ( t 2 − ˆ 2 ) St 2 dt 
T T

 0 2   2
0

price under BS (ˆ )
correct price under Q
 mis − replication 

• In English: mis-pricing is the expected mis-replication

• More generally: E Q1 CT  − E Q0 CT 

− If M1 (Q1) and M0 (Q0) are two arbitrage-free diffusion models
(where M1 is a general diffusion
 T  0t
= E Q1  
 0 2
( ) 
 t 2 −  0 ( St , t ) St 2 dt 
2


and M0 is a non-stochastic volatility model)
− Then the difference of price between the two models
 T  0t Q1
= E Q1  
 0 2
( 2
) 
E  t 2 St  −  0 ( St , t ) St 2 dt 

is the expected mis-hedge from delta-heding with M0
in a world driven by M1
 Danske Bank

The implied volatility formula (Dupire, 1996)

• Consider Sigma-Zero with M0 = Black-Scholes(sigma-hat) and M1(Q) general:
 T ˆ 
( )
E Q CT  − E BS ( ) CT  = E Q   t E Q  t 2 St  − ˆ 2 St 2 dt 
ˆ

 2 
0

• The model and Black-Scholes agree if and only if:

E Q CT  − E BS ( ) CT  = 0
ˆ

= E Q   ˆ tˆ 2 St 2 dt  = E Q   ˆ t E Q  t 2 St  St 2 dt 

T T

 0   0 
E Q  t 2 St  q ( St , t ) ˆ ( St , t ) St 2 dSt dt
T

= ˆ =
2  0 

  q ( St , t ) ˆ ( St , t ) St dSt dt
T
2
0

• In English: we have a ”reciprocal” Dupire formula

Expressing the implied variance as an average of local variances in spot and time
Weighted by probability times gamma (times spot^2 in log-normal terms)
 Danske Bank

The IV formula
Density maximum
around the spot today
 w ( x, t ) E
T
ˆ 2 =  Q
 t 2 St = x dxdt
0

q ( x, t ) ˆ ( x, t ) x 2
w ( x, t ) =
  q ( x, t ) ˆ ( x, t ) x dxdt
T
2
0

Gamma maximum
around the strike at maturity

Weights a suspended bridge

between the spot today and
the strike at maturity
 Danske Bank

The IV formula
• Does not lend itself to simple or efficient implementation
(contrarily to the local volatility formula the other way around)

• Essential analysis tool for valuation and risk

• Leads to many useful results (including Dupire’s formulas of 1992 and 1996)
• See Blacher (RiO 2018) for an application to calibration of complex hybrid models

• Generally taught in programs in Finance, including

− NYU (by Bruno Dupire himself)
− Baruch (by Jim Gatheral, whose book Volatility Surface discusses applications of sigma-zero in deep detail)
− Copenhagen University (where I teach volatility)
 Danske Bank

Back of the envelope calculation: spot risk vs volatility risk

• Very roughly:
− Normal (Bachelier) approximation
− First order analysis

• S&P ATM call

− Delta ~ 0.5, vega ~ 0.4 * sqrt(T)
− VIX (S&P volatility) generally between 10 and 20 outside stress periods
− With volatility 15, maximum loss from ”forgetting” delta-hedge with 97.5% confidence ~ 2 standard devs
= delta * 2 std = 0.5 * 2 * 15 sqrt(T) = 15 sqrt(T)
− Expected loss from mis-predicting volatility by a maximum of 5 points (bounds of normal range)
= 5 vegas = 2 sqrt(T)
− volatility risk ~ spot risk / 7.5 times smaller but of same order (”similar”)
 Danske Bank

Options are hedged with options

• If we don’t hedge volatility risk
− Estimate volatility, delta-hedge
− Run risk that volatility realizes differently than predicted
− Risk similar to not hedging in the first place

• If we do hedge volatility, we must trade options

− European options become hedge instruments, like additional underlying assets
− ”Options are hedged with options”

• Our model must respect the market price of hedge assets

− Calibrate (not estimate) volatility
− (Should also reasonably represent their price dynamics => stochastic volatility models)
 Danske Bank

Superbuckets
• Valuation pipeline
market calibration model pricing
e.g. local volatility V0
C ( K , T )  ˆ ( K , T )
E  T 2 ST = K  =
2C T  ( St , t ) Monte-Carlo
K 2CKK FDM
…
− Ultimately: V0 = h (ˆ ) where ˆ : two-dimensional surface

V0 V0
• Superbucket = two-dimensional collection of differentials ˆ ( K , T )

C ( K , T )

V0
− C K , T = how many calls (K, T) should I sell to hedge exposure in ˆ ( K , T )
( )

− (Evidently) European options have a single non-zero superbucket since they hedge themselves
− But the vega of exotic transactions is split over strikes and maturities and hedgeable by trading Europeans
 Danske Bank

Superbuckets in practice
• Superbuckets are famously hard in practice,
especially with Monte-Carlo
− Unstable differentiation of Monte-Carlo result
− Unstable differentiation through calibration
− Discontinuous profiles like barriers are not differentiable
− Slow differentiation by finite differences
− ...

• See discussion and elements of solution

leveraging automatic adjoint differentiation (AAD),
parallel Monte-Carlo
and modern C++ code in the chapter 13 of ➔
 Danske Bank

Bruno Dupire’s legacy to derivatives risk management

dSt 2CT
• Goes well beyond the local volatility model =  ( St , t ) dW ,  ( K , T ) =
St K 2CKK
− Rather simplistic and criticized (not least by Bruno Dupire himself) for unrealistic dynamics
− Yet still implemented as standard in all in-house and external derivatives systems 25 years after publication

• Helped establish the universal practice of calibration

− Approaching derivatives from the hedge persepctive and applying the FTDT
− Leading to the idea of ”hedging options with options”
− Itself leading to the principle of calibration: predict is good, hedge is better

• Provided a practical necesary and sufficient condition for calibration to options:

 dS 2  2C
E  T  ST = K  =  f 2 ( K , T )  T
 ST   CKK

• And an ”inverse” sigma-0 formula to analyze option prices across models

and understand/manage model risk
Q1 Q0  T  0t Q1
Price difference in different models E CT  − E CT  = E  0
 T 2
Q1
( 2
)
E  t 2 St  −  0 ( St , t ) St 2 dt 
 q ( x, t ) ˆ ( x, t ) x 2
Implied volatility = weighted average of local volatility ˆ =   w ( x, t ) E Q  t 2 St = x dxdt , w ( x, t ) =
2

  q ( x, t ) ˆ ( x, t ) x dxdt
0 T
2
0
 Danske Bank

Variance swaps and the VIX index

From Arbitrage Pricing with Stochastic Volatility (Dupire, 1992) to the VIX index
 Danske Bank

VIX index
• One month S&P volatility index published by CBOE
• Highly succesful barometer of capital markets nervousness
− Example: Financial Times, May 22, 2018
”Heightened volatility struck US equities on Tuesday as developments in Italy’s political sphere fueled
investor anxiety. The CBOE’s VIX index, a widely tracked measure of volatility, rose 4.4 points to 17.64.”
− VIX futures and options trade on the CBOE in extremely large volumes

• How is the VIX calculated?

− From CBOE’s white paper

K i +1 − K i −1
− Where F = 1m forward, Q = price of call (Ki>S0) or put (Ki<s0) of quoted strike Ki, maturity 1m and Ki =
2
− So the VIX (square) is a the price of a portfolio of European options, not an estimation of volatility

• How is this possible?

− The answer is given in Dupire’s 1992 founding paper Arbitrage Pricing with Stochastic Volatility
− Established theoretical grounds for Variance Swaps and a forward looking volatility index (VIX)
 Danske Bank

Variance swaps
2
 dSt 
• (Apparently) exotic option that delivers realized variance at maturity: VT = 
T

T
 = 0  t dt
2

 t 
S
0

• How do we replicate it and what is its value?

ˆ
T  S
2
ˆ 2
ˆ t St 2
− Recall the FTDT: CT = C0 + 0 ˆ t dSt + 0 ˆ t ( t 2 − ˆ 2 ) St 2 dt = C0 + 0 ˆ t dSt + 0 t t  t 2 dt −
T 1 T T T

2 2 2 
0 2
dt

ˆ t St 2 2
− The term in red looks similar to the payoff, as long as = 1  ˆ t = 2
2 St

ˆ t St 2
• We buy ~some~ option C and delta hedge it; we get: (  t 2 − ˆ 2 ) dt
T T
CT − C0 −   t dSt = 
0 0 2

2
− If C is such that ˆ t =
T T
2 then CT − C0 −   t dSt =   t dtdt −  T
2
ˆ2
St 0 0
constant
buy and delta-hedge desired payoff

− Note: we generally don’t like the error term but for variance swaps, we use it to produce the desired payoff
 Danske Bank

Log-contracts
2 2
• Elementary integration bears: ˆ t = 2  ˆ t = − +   Cˆ t = −2 log ( St ) +  St + 
St St

• Must hold at maturity T so the payoff must be: CT = −2 log ( ST ) +  ST + 

 S − S0 S 
For instance: CT = 2  T − log T  ”ATM delta-neutral log-contract”
 S0 S0 

St − S 0
• Conversely: Cˆ = E
t
BS (ˆ )
CT St  = 2
S0
S
− 2 log t + ˆ 2 (T − t )
S0

• So, effectively, to buy the log-contract and delta-hedge it replicates the desired payoff
CT − Cˆ 0 −  ˆ t dSt =  ( t 2 − ˆ 2 ) dt =   t 2 dt − Cˆ 0    t 2 dt = CT −  ˆ t dSt
T T T T T

0 0 0 0 0
 Danske Bank

Variance swaps in practice

• To replicate a variance swap, buy a log-contract and delta-hedge it with Black-Scholes
It follows that the premium of the variance swap is the market price of the LC

• Log-contracts are European options  S − S0

CT = 2  T
S 
− log T 
that (obviously) don’t trade directly  S0 S0 

What does trade are European calls and puts of given strikes Ki

• Like any European payoff

LCs are approximated
as piece-wise linear payoffs,
with a combination of calls and puts
See e.g. Carr-Madan (1999)

• We can now reveal what VIX really is

 Danske Bank

What really is VIX?

• VIX (squared) is the price of a combination of traded calls and puts of expiry 1m

• When rates=dividends=0 and ATM option trades, F=S0=K0 and the formula simplifies
VIX 2T K −K
=  wi C ( K i ) +  wi P ( K i ), wi = i +1 2 i −1
2 Ki  S0 Ki  S0 Ki

• We chart the payoff of this combination on calls and puts:

LC
• Clearly:
− VIX portfolio = traded calls and puts that best replicate LC VIX
− VIX^2 is the value of a 1m variance swap on S&P
 Danske Bank

Bruno Dupire on variance swaps and the VIX

• The VIX is (the square root of) the price of a variance swap

• This explains why VIX is so succesful:

− Not an estimation but a forward looking, tradable consensus extracted from quoted option prices
− VIX^2 = necessary amount for delivering future realized variance modulo simple trading strategy

• Bruno Dupire laid the foundation of the multi-billion variance swap business
(for which he is generally credited)

• It follows that Bruno Dupire really invented the VIX

(for which the CBOE white paper ”forgets” to credit him)
 Danske Bank

Functional Ito Calculus (Dupire, 2009)

Extension to path-dependent products
 Danske Bank

Fundamental elements of risk analysis and risk management

C
• Delta-hedge: dC ( St , t ) =
S
dSt + ...dt

C  2C 2 2
• Condition for absence of arbitrage: t
= − 2 St  t
S

 t ( t 2 −  0 2 ) St 2 dt
1 T 0
• Foundamental Theorem of Derivatives Trading: T
CT = C0 +   0t dSt +
0 2 0

• Sigma-Zero formula:  0 2
(
 T  0t St 2 Q1
E Q1 CT  − E Q0 CT  = E Q1   ) 
E  t 2 St  −  0 ( St , t ) dt 
2

• All rely on first and second order sensitivities of: Ct = E CT St  = C ( St , t )

C  C  2C 2 2 
and stochastic calculus (Ito’s Lemma): dC ( St , t ) = dSt +  + S  t  dt
S  t 2S
2 t

 Danske Bank

What about path-dependent options?

• Path-dependent option (barrier, lookback, asian...): CT = h ( St , 0  t  T ) , functional of the path

• We still have risk-neutral pricing (Harrisson-Pliska, 1980): Ct = E CT Ft 

• But price is now a functional of current path: Ct = h ( Su , 0  u  t )

• Calculus involving functionals of stochastic processes did not exist

➔ No well defined delta, no theta/gamma equation,

no PnL explain or model risk decomposition (sigma-zero)
 Danske Bank

Functional Ito Calculus (Dupire, 2009)

• Did not stop market participants to estimate sensitivities with frozen history:
h (Su , 0  u  t , St +  , t ) − Ct
Ct = h (Su , 0  u  t , St , t ) ,  t  , etc.


• And apply delta-hedge, theta-gamma equation, PnL explain and model risk
Even in the absence of a mathematical definition or guarantee
• Turns out this methodology is correct
• In 2009, Bruno Dupire extended stochastic mathematics to:
− Define sensitivities of functionals of stochastic processes
− Demonstrate that Ito’s lemma, theta-gamma equation, FTDT and sigma-zero all hold for exotics

• Functional Ito Calculus ”closed the circle”

− Completed financial theory by extending mathematical principles of risk management to exotics
− Reconciliated theory and practice
− Established mathematical guarantees for market and model risk analysis and management of exotics

• See Yuri Saporito, RiO 2018, for a (excellent) introduction to FIC

 Danske Bank

Conclusion
 Danske Bank

A model-free thinker
• Bruno Dupire is best known for his local volatility model of 1992
which is a fraction of his work

• Dupire’s work always focused on hedging and replication

• His major contributions:
− Calibration condition, systematic application of FTDT, sigma-zero, VS/VIX, functional calculus
− All apply to a wide variety of models

• Bruno Dupire is not a ”one-model researcher” but a ”model-free” thinker

• Further emphasized by his less known work on
− Mapping major martingale properties to market arbitrage (1997)
− Tradable estimates and indicators (2015)
 Danske Bank

A vast contribution to the theory and practice of capital markets

• Bruno Dupire only published a tiny fraction of his work
He mainly contributed through:
− Leading research in SocGen(1991-1992), Paribas(1992-1997), Nikko(1997-1998) or Bloomberg(since 2004)
− Internal memos like his ”Notes de Recherche” from Societe Generale of 1991-1992 (French manuscript)
− Professional presentations and working papers

• Yet, his contribution is universally recognized and earned multiple prestigious awards
• Some of his major research remains less well known:
− First application of artificial neural networks (ANN) to finance in 1988, 30 years ahead of current craze
− Substantial work on numerical integration and Monte-Carlo simulations around 1995
− Model-free arbitrage outside VS/VIX

• In addition to his scientific contribution

Dupire trained and mentored many ”babies” who grew up to run derivatives markets:
− 3 former partners and a number of MDs at Goldman-Sachs,
a former Global Head of Trading and a former Global Head of Research at BNP,
a current Global Head of Interest Rate Research at BAML, and many more
 Danske Bank

Bruno Dupire
• 60 years of existence as of November 30, 2018
• 30 years of major innovations, shaping global derivatives markets
• An outstanding thought leader, admirable human being and incredible friend

• Thank you for your attention

antoinesavine.com
 Danske Bank

antoinesavine.com