Deep Learning in Finance

option pricing

Damien Challet
[email protected]

4th December 2023

 Hard problems in Finance

1. Modelling and calibration

2. Solving high-dimensional equations

3. Prediction

4. Portfolio construction
 Options

Insurance / bet: price/volatility change

option  right to buy (call) or to sell (put) an asset at price K

(strike).

 European option: single exercise time T (maturity)

 American option: any time until expiry
 barrier options
 exotic options
 Payoff shape

Strike K = 100

from [link]
 Options variables

 asset price pt

 strike price K

 time to maturity T

 interest rate r

! option price at time t = 0 ?

 European option price

 Hyp: call option, strike K , time to maturity T

 Option price C = fair price

 Naively,
C = E [max(pT K ; 0)]
 Interest rate?
C can be invested in a risk-free asset

=
Z1
C (1 + r )T = E [max(pT K ; 0)]

P(pT )(pT K )dpT

K

! model of price needed

 Bachelier’s model

 Price evolution

pt+1 = pt + t
t  N (0; B2 )
 Normal diffusion
(pT p0 )2

P(pT jp0 ) = p
1 2T  2
e B
2 T B

 Additive world
 Black-Scholes model

 Price evolution

log pt+1 = log pt + t

t  N (0; BS
2
)

 Log-normal diffusion

P(log pT j log p0 ) = p 1 (log pT log p0 )2

e 2 T
2 T 

 N.B.: no jumps, constant volatility

 Fair price in discrete time with hedging

 t 2 f0;


; T g
 vt : number of shares invested in underlying asset

Xh i
C (1 + r )T = E [max(K pT ; 0)]
t 1
E (1 + r )T k 1
vt (pt+1 pt rpt )
t=0

 vt : hedging (reduce or eliminate risk)

 Optimal hedging

 Capital evolution of option underwriter:

Xh i
∆W = C (1 + r )T [max(K pT ; 0)]
t 1
+ (1 + r )T k 1
vt (pt+1 pt rpt )
t=0

 Residual risk = minfvt g var(∆W )

 Continuous time & Gaussian increments

! Black-Scholes miracle:
zero residual risk with optimal hedging
 Black-Scholes option price

Continuous-time limit: T = N  , ; 1=N ! 0,

call price

CBS (p0 ; K ; T ; r ;  ) = p0 P> [y ] Ke rT

P> [y+ ]

p
1 + erf(x = 2)
P> (x ) =
2
log(p0 =K
e rT )   2 T =2
y = p
 T
 Price scaling

CBS (p0 ; K ; T ; r ;  ) = p0 P> [y ] Ke rT

P> [y+ ]

p
1 + erf(x = 2)
P> (x ) =
2
log(p0 =Ke rT )   2 T =2
y = p
 T
Ke rT
Setting k = p0

CBS (k ; T ; r ;  )
= P> [y ] kP> [y+ ]
p0
 Payoff shape

Strike K = 100

from [link]
 Black-Scholes option price

Continuous-time limit: T = N  , ; 1=N ! 0,

Put price

PBS (p0 ; K ; T ; r ;  ) = p0 P> [ y ] + Ke rT

P> [ y+ ]

p
1 + erf(x = 2)
P> (x ) =
2
log(p0 =K
e rT )   2 T =2
y = p
 T
 Parameters of Black and Scholes

 Known: p0 ; K , T

 Possibly known: r

 Optimised: hedging fvt g

 Estimated: 

 Assumed: continuous time, Gaussian returns

 Black-Scholes vs reality: returns

Heavy-tailed skewed price returns+stochastic volatility

! implicit volatility:
| {z } | {z }
CBS (p0 ; K ; T ; r ; impl ) = option price(p0 ; K ; T ; r )
BS world real world

 Heavy tails ! smile

 Skewed returns, leverage effect ! skewed smile
 Non-constant volatility

! stochastic volatility model

! non-explicit option price formulae (e.g. Heston, etc.)
N.B.: additional risk: volatility of volatility
 Deep learning and option pricing

1. Model-driven speed up calibration

2. Data-driven approach Ito who?

3. Model + Data improve on 1. and 2.

 1. Deep learning option prices from model

1. NN fitted to reproduce expensive model output

1.1 Monte-Carlo
1.2 Heston
1.3 Rough Bergomi [link]

2. Find best parameters of the NN-fitted model

2.1 with standard optimisation
2.2 add no-arbitrage conditions
2.3 NN [link]
 2. Data-driven option price

1. Fixed parameters (K ; p0 ; r ; T )

2. Add estimated parameters ;skew, tail exponent

3. Fit option prices with DL ! option pricer

3. Deep learning option prices: prices from model and data
Example: Jang et al. (2020)
[https://doi.org/10.1016/j.inffus.2020.12.010]
 3. Deep learning option prices from model

Improvement:
1. Give what you know as inputs

2. Learn residuals

Example: Funahashi (2021) [link]

 Learn C3 Cdata , where C3 = third order approximation of

CBS
 Smaller training time
 More robust and stable
 Faster calculations
 Solving complex option pricing equations

Wang et al (2022) [link]

 Physics informed NN (PINNS) (2019) [link] (7200 citations)

 Solving complex option pricing equations

Physics informed NN (PINNS) (2019) [link] (7200 citations)

 Solving complex option pricing equations
Wang et al (2022) [link]

Non-linear BS equation (with transaction costs)

Two-asset case
Deep learning: regression + no-arbitrage constraints

Ackerer, Takasovska and Vatter (2020),

Deep Smoothing of the Implied Vol Surface [link]

 Implied volatility ansatz (residuals)

 = NN prior
! (k ;  ) =  2 

where prior comes from a model and  : time to maturity

 6 shape and no-arbitrage constraints
 DL: regression + constraints

C1, C2, C3 are guaranteed by activation function

Constraints help generalization abilities

DL: regression + constraints
 Nonarbitrage constraint: put-call parity

 C : call price

 P : put price

 St : spot (current) price

 K : strike price

 T : time to maturity

 r : interest rate
rT
C P=S e K