Likelihood-based inference for correlated diffusions

Κωνσταντίνος Καλογερόπουλος, Πέτρος Δελλαπόρτας, Gareth O. Roberts

Abstract

We address the problem of likelihood-based inference for correlated diffusion processes using Markov chain Monte Carlo (MCMC) techniques. Such a task presents two interesting problems. First, the construction of the MCMC scheme should ensure that the correlation coefficients are updated subject to the positive definite constraints of the diffusion matrix. Second, a diffusion may only be observed at a finite set of points and the marginal likelihood for the parameters based on these observations is generally not available. We overcome the first issue by using the Cholesky factorisation on the diffusion matrix. To deal with the likelihood unavailability, we generalise the data augmentation framework of Roberts and Stramer (2001) to d-dimensional correlated diffusions including multivariate stochastic volatility models. Our methodology is illustrated through simulation-based experiments and with daily EUR/USD, GBP/USD rates together with their implied volatilities.

Introduction

Diffusion processes, defined by Stochastic Differential Equations (SDEs), are fundamental to continuous-time modeling in finance and physics. However, they present a significant challenge for statistical inference: while we usually only observe data at discrete time points, the likelihood function—the probability of observing that data—depends on the continuous path between those points. This likelihood is almost never available in a closed form, necessitating the use of sophisticated estimation techniques.

When dealing with multiple correlated diffusions, such as several interest rates or currency pairs moving in tandem, two specific hurdles arise:

  1. The Positive Definite Constraint: Any MCMC algorithm must ensure that the estimated diffusion matrix remains positive definite (mathematically valid) at every step. Updating correlations individually often violates this constraint.
  2. MCMC Degeneracy: As researchers attempt to "fill in the gaps" between observations with more simulated data points (data augmentation), the correlation between the parameters and the augmented paths can become so high that the MCMC algorithm effectively freezes, failing to explore the parameter space.

Methodology

This paper proposes a two-pronged solution to these challenges, focusing on mathematical consistency and algorithmic efficiency.

Cholesky Factorization

Instead of updating the diffusion matrix $A$ directly, the authors work with its Cholesky factor $C$, where $A = CC'$. This approach allows for component-wise updates of variances and correlations while automatically satisfying the positive definite constraint. This provides a natural and flexible way to handle high-dimensional matrices in a Bayesian setting, allowing researchers to update specific correlation coefficients without risking the validity of the entire matrix.

Generalized Data Augmentation and Reducibility

The authors extend the data augmentation framework of Roberts and Stramer (2001) to the multivariate case. The core of this method is a transformation to unit volatility. By transforming the diffusion process into one with an identity diffusion matrix, the parameters are decoupled from the latent paths. This ensures that the MCMC algorithm remains efficient even as the discretization becomes arbitrarily fine.

A critical insight of the paper is the "Reducibility Condition." While scalar diffusions can always be transformed to unit volatility, multivariate ones cannot. The authors provide specific mathematical conditions (Propositions 3.1 and 3.2) defining when a multivariate system is reducible. For systems that are not fully reducible, such as certain stochastic volatility models, the authors demonstrate how to apply these transformations to the volatility components specifically, making inference possible for partially observed systems.

Experimental Validation

Simulation Study: 3D CIR Model

The methodology was first validated using a 3-dimensional Cox-Ingersoll-Ross (CIR) model, a standard for interest rate modeling. Using 500 simulated equidistant observations, the MCMC scheme demonstrated high efficiency, with the independence sampler used for path updates achieving a 98.14% acceptance rate. The posterior means and medians were in close agreement with the true parameters, successfully recovering complex correlations (e.g., a posterior $\rho_{21}$ of 0.480 against a true value of 0.45).

Empirical Application: Foreign Exchange Rates

The framework was applied to daily EUR/USD and GBP/USD exchange rates from 2005 to 2006, paired with their monthly implied volatilities. This created a 4-dimensional time series used to fit a bivariate Heston stochastic volatility model. The model successfully captured high correlations between the currency pairs ($\rho \approx 0.82$) and their respective volatilities ($\rho \approx 0.57$), demonstrating its ability to extract the correlated market structure that independent models would overlook.

Conclusion and Future Work

This research provides a robust framework for performing likelihood-based inference on complex, correlated systems. By combining Cholesky factorization with generalized data augmentation, the authors overcome the traditional pitfalls of MCMC degeneracy and matrix constraints. While the unit volatility transformation is limited by the reducibility condition, the workarounds provided for stochastic volatility models expand the reach of these techniques. Future research may focus on comparing this approach to other likelihood approximations in settings with extremely sparse data.

Summary updated

Abstract

We address the problem of likelihood based inference for correlated diffusion processes using Markov chain Monte Carlo (MCMC) techniques. Such a task presents two interesting problems. First, the construction of the MCMC scheme should ensure that the correlation coefficients are updated subject to the positive definite constraints of the diffusion matrix. Second, a diffusion may only be observed at a finite set of points and the marginal likelihood for the parameters based on these observations is generally not available. We overcome the first issue by using the Cholesky factorisation on the diffusion matrix. To deal with the likelihood unavailability, we generalise the data augmentation framework of Roberts and Stramer (2001 Biometrika 88(3):603-621) to d-dimensional correlated diffusions including multivariate stochastic volatility models. Our methodology is illustrated through simulation based experiments and with daily EUR /USD, GBP/USD rates together with their implied volatilities.
BibTeX
@inproceedings{2011likelihood,
  author = {Κωνσταντίνος Καλογερόπουλος and Πέτρος Δελλαπόρτας and Gareth O. Roberts},
  title = {{Likelihood-based inference for correlated diffusions}},
  booktitle = {Canadian},
  year = {2011},
  doi = {10.1002/cjs.10096},
}
Figure