A fast adjointbased quasilikelihood parameter estimation method for diffusion processes
Research output: Contribution to journal › Published meeting abstract
Details
Authors  

Organisations  
Research areas and keywords  Subject classification (UKÄ) – MANDATORY

Original language  English 

Journal  [Publication information missing] 
Publication status  Published  2014 
Publication category  Research 
Peerreviewed  Yes 
Event  8th World Congress of the Bachelier Finance Society  Brussels Duration: 2014 Jun 2 → 2014 Jun 6 
Bibliographic note
A fast adjointbased quasilikelihood parameter estimation method for diffusion processes
Josef Höök (Uppsala University, Sweden)
Joint work with Erik Lindström
Tuesday June 3, 16:0016:30  session P2  Poster session  room lobby
Likelihood based parameter estimation for diffusion processes is an important topic in many areas of mathematical finance. For a general irreducible diffusion model it is common to approximate the transition density using either Monte Carlo based methods or by finite difference discretization of the FokkerPlanck equation. These methods require the evaluation of an approximate probability density between each observation, which quickly becomes very time consuming as the number of observations increase. Instead of approximating the transition density explicitly in the construction of the likelihood a simple strategy is to replace the exact, but unknown density by an approximate density with exact moments. This technique is known as the quasilikelihood method and many previous studies have been on diffusion models where one can obtain analytical expression for the moments. Monte Carlo estimation is the standard method of choice when analytical moments are unavailable.
Instead of using Monte Carlo based methods we here suggest to estimate the moments for the quasilikelihood from the approximate solution of the Kolmogorovbackward equation using finite differences. The Kolmogorovbackward equation is the adjoint to the FokkerPlanck equation. The immediate advantage of this is that we need only to solve one backward equation for any number of observations, which is a dramatic reduction in computational complexity. Another nice property of the backward equation is the wellbehaved initial condition in terms of moments, which should be contrasted to the initial condition of the FokkerPlanck equation given by a Dirac measure.
The quasilikelihood method together with approximate moments from the discrete backward equation is tested on common models e.g. CIR, GEN1 and the low computational complexity and high performance is demonstrated.