Pricing and calibration in local volatility models via fast quantization
Lucio Fiorin
Universit`a degli studi di Padova
Parma, 29
thJanuary 2015.
Joint work with Giorgia Callegaro and Martino Grasselli
Quantization: a brief history
Birth: back to the 50’s, due to the necessity to optimize signal transmission, by appropriate discretization procedures;
Applications: information theory, cluster analysis, pattern and speech recognition, numerical integration and numerical probability (90’s);
Idea: approximating a signal admitting a continuum of possible values, by a signal that takes values in a discrete set;
Two types (probability):
Vector quantization random variables;
Functional quantization stochastic processes;
How: numerical procedures mostly based on stochastic optimization algorithms very time consuming.
Lucio Fiorin Universit`a degli studi di Padova
Quantization: technical introduction
Given an R
d-valued random variable X on (Ω, A, P), X ∈ L
r, quantizing X on a grid Γ = (x
1, . . . , x
N) consists in projecting X on Γ, following the closest neighbor rule.
A grid Γ
?minimizing the L
r−mean quantization error
||X − Proj
Γ(X )||
r= || min
1≤i ≤N|X − x
i| ||
rover all the grids with size at most N is the L
r-optimal quantizer.
The projection of X on Γ
?, Proj
Γ?(X ) is called the quantization of X :
Proj
Γ?(X ) =
N
X
i =1
x
i11
Ci(Γ?)(X )
where C
i(Γ
?) ⊂ {ξ ∈ R
d: ||ξ − x
i|| = min
1≤j ≤N||ξ − x
j||}, is
called the Voronoi partition, or tessellation induced by Γ
?.
Figure: Voronoi diagram for the optimal grid of N (0, I
2).
The L
r−mean quantization error goes to zero as the grid size N → +∞ and the convergence rate is rules by the so-called Zador Theorem.
From a numerical point of view, finding an optimal quantizer may be a very challenging and time consuming task. This motivates the introduction of sub-optimal criteria: stationary quantizers.
Lucio Fiorin Universit`a degli studi di Padova
Stationary quantizers
Definition: An N-quantizer Γ = {x
1, · · · , x
N} inducing the quantization Proj
Γ(X ) of X is said to be stationary if
E [X |Proj
Γ(X )] = Proj
Γ(X ).
Optimal quantizers are stationary;
Stationary quantizers ˜ Γ are critical points of the distortion function associated with Γ:
D(Γ) :=
N
X
i =1
Z
Ci(Γ)
|z − x
i|
2d P
X(z).
Stationary quantizers are interesting insofar they can be found through zero search recursive procedures like Newton’s
algorithm or fixed point procedures.
“Step by step marginal quantization”: ideas
A new quantization approach recently introduced by Pag` es and Sagna, 2014.
Consider a continuous-time Markov process Y
dY
t= b(t, Y
t)dt + a(t, Y
t)dW
t, Y
0= y
0> 0, where W is a standard Brownian motion and a and b satisfy the usual conditions ensuring the existence of a (strong) solution to the SDE;
Given T > 0 and {t
0, t
1, . . . , t
M}, with t
0= 0 and t
M= T ,
∆ = t
k− t
k−1, k ≥ 1, the Euler scheme for the process Y is Y e
tk= e Y
tk−1+ b(t
k−1, e Y
tk−1)∆ + a(t
k−1, e Y
tk−1)∆W Y e
t0= e Y
0= y
0where t
k= k∆ and ∆W := (W
tk− W
tk−1) ∼ N (0, ∆);
Lucio Fiorin Universit`a degli studi di Padova
KEY REMARK: for every k = 1, . . . , M L
Y e
tk| Y e
tk−1= x
∼ N m
k−1(x), σ
k−12(x)
(1)
where
m
k−1(x ) = x + b(t
k−1, x )∆
σ
2k−1(x ) = [a(t
k−1, x )]
2∆.
IDEA: quantize recursively, using vector quantization, every marginal random variable e Y
tk, given its (Gaussian) conditional distribution given e Y
tk−1;
It can be seen (see Pag` es and Sagna, 2014) that the error
made by quantizing the Euler scheme can be controlled, under
some mild regularity assumptions on the process.
“Step by step marginal quantization”: stationary quantizers The distortion function at time t
k+1, relative to e Y
tk+1, is
D
k+1(x) =
N
X
i =1
Z
Ci(x)
(y
k+1− x
i)
2P
Y e
tk+1∈ dy
k+1(2)
where N is the (fixed) size of the quantizer x = {x
1, x
2, . . . , x
N}.
We are looking for x ∈ R
Nsuch that
∇D
k+1(x) = 0.
QUESTION: Is it possible to apply Newton-Raphson now?
ANSWER: NO! We do NOT know the distribution of e Y
tk+1!
Lucio Fiorin Universit`a degli studi di Padova
Using the conditional distribution in (1) we have
P( e Y
tk+1∈ dy
k+1) = Z
R
φ
mk(yk),σk(yk)(y
k+1) P( e Y
tk∈ dy
k)dy
k+1where φ
m,σis the density function of a N (m, σ
2).
Due to the discrete nature of the quantizer, the integral in (2) becomes a finite sum;
We deduce a recursive procedure to obtain the stationary (here optimal) quantizer at time t
k+1, based on the quantizer at time t
k, k ∈ {1, . . . , M} (Y
0= y
0is not random);
The distorsion is continuously differentiable: it is possible to
compute the gradient and the Hessian matrix of the distortion
function Newton-Raphson faster computations wrt
stochastic algorithms.
The Quadratic Normal Volatility model
Much attention in the financial industry due to its analytic tractability and flexibility (Blacher, 2001; Ingersoll, 1997; Lipton, 2002; Andersen, 2011):
dY
t= (e
1Y
t2+ e
2Y
t+ e
3)dW
t, Y
0= y
0> 0, (3) for some e
1, e
2, e
3∈ R, where W is taken under
the risk neutral measure.
Includes, as special cases, the Brownian motion, the geometric Brownian motion and the inverse of a three-dimensional Bessel process. We refer to (Andersen, 2011) and (Carr, Fisher and Ruf, 2013) for technical properties of the model.
Mimicking a quadratic spot volatility gives some chances to get an implied volatility curve that reproduces the smile and skew effects using a parsimonious number of parameters.
Lucio Fiorin Universit`a degli studi di Padova
Vanilla options in the QNV model
RE-PARAMETRIZATION : dY
t= σ
qY
t+ (1 − q)x
0+ 1
2 s (Y
t− x
0)
2x
0dW (t).
PRICING:
Closed-form solutions for vanilla derivatives available (see Andersen, 2011);
Solutions depend on the roots of the polynomial in (3);
Even the implementations of closed form solutions requires some care;
CALIBRATION: must allow for the possibility to switch from the real roots case to the complex roots case without
constraints;
Numerical results: pricing of Vanilla options
Given stationary quantization grids, pricing is immediate: the price (with r = 0) of an European Vanilla Put option on Y with
maturity T and strike K , given an N-quantization grid
y = (y
1, . . . , y
N) at t = T and associated optimal quantizer b Y
T, is
E[(K − Y
T)
+] =
N
X
i =1
(K − y
i)
+P
Y b
T= y
i.
The dimension of every grid is 20 and we have 10 time steps;
Parameters as in Andersen, 2011;
Error measure: sum of the squared differences between implied volatilities (“Res-norm”) on 7 strikes, from 85% to 115% of the spot initial value (y
0= 100) and 6 maturities, from 2 months to 2 years.
Lucio Fiorin Universit`a degli studi di Padova
Results obtained using Matlab on a CPU 2.4 GHz and 8 Gb memory laptop.
Real roots Complex roots
Res. Norm. 3.5025e − 04 2.7504e − 04
Comp. Time (closed form) 0.0295 sec 9.8383 sec Comp. Time (quantization) 0.8804 sec 0.8338 sec
Figure: Quantization grids, an example.
Numerical results: calibration
We work on European Vanilla Call-Put option on the Dax Index, as of 19 June 2014;
Calibration is done via a standard non-linear least-squares optimizer that minimizes
X
n
σ
impn, market− σ
impn,model2.
Using closed form formulas, it turns out that the implied volatility smile produced by the market is fitted better when the two roots are complex.
Closed form formulas do not perform well for short maturities.
Short maturities from 2 to 5 months, long maturities from 1 to 3 years, 7 strikes.
Lucio Fiorin Universit`a degli studi di Padova
Short maturities Long maturities Closed formulas Res. Norm. - 5.6292e − 04
Comp. Time - 339.1503 sec
Quantization Res. Norm. 3.7074e − 04 4.2690e − 04 Comp. Time 141.6354 sec 168.3524 sec
Figure: Calibration via quantization: squared errors of the implied
volatilities. On the left long maturities, on the right short maturities.
Numerical results: pricing of barrier options
Same model data as before, we fix T =
13and K = 100.
We compare the price of up-and-out put options obtained via the quantization method, with 100-dimensional quantizers, with Monte Carlo, with 5 ∗ 10
5simulations. The time step equal to
3601.
The benchmark price is computed via Monte Carlo, with 10
3discretization points in the time grid and 10
6simulations.
Benchmark price Quantization price Monte Carlo price
L = 101.25 122.83447 141.06537 144.33757
L = 102.5 205.96689 216.54927 220.69586
L = 103.75 264.11461 268.67753 273.11058
L = 105 299.71130 299.93827 304.81308
L = 106.25 318.74848 316.32067 321.23110
Computational time - 13.98614 sec 50.06883 sec
Lucio Fiorin Universit`a degli studi di Padova
Future developments
Increasing dimensionality: fast quantization in stochastic volatility models, like Heston, SABR, multi-Heston, Wishart;
Increasing derivatives’s complexity: pricing of (more) exotic derivatives;
Application to more general discretization schemes?
Thank you for your attention
Lucio Fiorin Universit`a degli studi di Padova