Likelihood-based inference for correlated diffusions

Κωνσταντίνος Καλογερόπουλος, Πέτρος Δελλαπόρτας, Gareth O. Roberts

TL;DR

This paper introduces a novel Bayesian MCMC framework for analyzing correlated diffusion processes, tackling the challenges of ensuring positive definite diffusion matrices and handling intractable likelihoods. By employing Cholesky decomposition and generalizing data augmentation techniques, the method allows for robust inference on complex financial models, demonstrated on simulated and real-world currency exchange rate data.

What Problem Does This Paper Address?

Modeling financial markets often involves continuous-time stochastic processes called diffusions. When multiple such processes are related (e.g., interest rates for different maturities, or currency exchange rates), their interactions are captured by "correlated diffusions." Inferring the parameters of these models using likelihood-based methods is challenging for two main reasons:

  1. Positive Definite Constraint: The diffusion matrix, which describes the instantaneous variances and covariances of the processes, must be positive definite. Standard parameterizations can make it difficult to ensure this constraint is met during the inference process.
  2. Intractable Likelihood: For most diffusion models, the exact likelihood function (which quantifies how well the model explains the observed data) cannot be computed analytically, especially when observations are only available at discrete time points. This makes direct parameter estimation difficult.

Methods

The authors propose a Bayesian approach using Markov Chain Monte Carlo (MCMC) techniques, combined with data augmentation. Key methodological contributions include:

  • Cholesky Factorization for Diffusion Matrix: To address the positive definite constraint, they reparameterize the diffusion matrix $A$ using its Cholesky decomposition. This allows them to work with a lower triangular matrix $\Sigma$, where $A = \Sigma\Sigma'$. By updating the elements of $\Sigma$ (specifically, through its Cholesky decomposition $C$), they can ensure that $A$ remains positive definite throughout the MCMC sampling. This facilitates component-wise updates, which is crucial for high-dimensional models.
  • Generalized Data Augmentation: They extend the data augmentation framework of Roberts and Stramer (2001) to handle $d$-dimensional correlated diffusions. This technique treats the unobserved continuous paths of the diffusion process as "missing data" and imputes them within the MCMC algorithm. This bypasses the need to directly compute the intractable likelihood.
  • Transformation to Unit Volatility: For certain classes of diffusions (those satisfying specific conditions), they introduce a transformation that effectively sets the diffusion matrix to the identity matrix. This simplifies the inference process and helps avoid degenerate MCMC algorithms, particularly when dealing with high levels of augmentation. This transformation is then used to derive a likelihood with a parameter-free dominating measure.

Datasets

  • Simulated Data: The methodology is first tested on simulated data from a 3-dimensional Cox-Ingersoll-Ross (CIR) model, allowing for controlled analysis of parameter estimation and the impact of augmentation levels.
  • EUR/USD and GBP/USD Exchange Rates: Real-world data consisting of daily exchange rates and their corresponding month implied volatilities from January 2005 to December 2006 are used to demonstrate the model's applicability to financial time series.

Key Insights and Ideas

  • Cholesky Decomposition as a Constraint Manager: The Cholesky decomposition is a powerful tool not just for matrix factorization but also for enforcing the positive-definiteness constraint in a computationally efficient way during MCMC sampling. It allows for component-wise updates, which is a significant advantage for complex, high-dimensional models.
  • Data Augmentation for Intractable Likelihoods: The paper highlights how data augmentation, when properly implemented, can effectively circumvent the problem of intractable likelihoods in continuous-time models. By imputing the latent path, the problem is transformed into a series of conditional sampling steps.
  • Reducibility for MCMC Efficiency: The concept of "reducibility" (transforming the diffusion to have unit volatility) is shown to be crucial for ensuring that the MCMC algorithm remains efficient and does not degenerate as the amount of data augmentation increases. This is a key insight for practical implementation.
  • Unified Framework for Volatility Models: The proposed framework is general enough to handle various correlated diffusion models, including multivariate stochastic volatility models, which are prevalent in finance.

Key Results

  • The MCMC scheme successfully estimates parameters for both simulated and real-world financial data.
  • The autocorrelation plots of the posterior draws show no sign of increased suspicion against the irreducibility of the chain, even with increasing levels of data augmentation.
  • Kernel density plots of posterior draws and log-likelihood values suggest that the data augmentation scheme provides good approximations, with densities becoming similar for higher levels of augmentation.
  • The application to EUR/USD and GBP/USD exchange rates demonstrates the model's ability to capture substantial correlations between exchange rates and their implied volatilities, which is important for accurate financial modeling.

Limitations and Future Work

  • Approximation Error: While the data augmentation scheme allows the discretization error to become arbitrarily small by increasing the level of augmentation, it still relies on an approximation of the likelihood.
  • Generalization to Partially Observed Diffusions: The paper notes that while the framework is presented for directly observed diffusions, adaptations to partial observation regimes (where some components might be unobserved or observed with error) are provided when necessary. However, generalizing some techniques to partially observed diffusions might present major difficulties.
  • Future Work: The authors mention that current work investigates the behavior of different approaches in various settings regarding dimensionality, correlation, and data sparseness. They also suggest that the posterior samples can be used for option pricing purposes, potentially in a Bayesian option pricing framework.

Summary updated

Abstract

The authors address the problem of likelihood-based inference for correlated diffusions. Such a task presents two issues; the positive definite constraints of the diffusion matrix and the likelihood intractability. The first issue is handled by using the Cholesky factorization on the diffusion matrix. To deal with the likelihood unavailability, a generalization of the data augmentation framework of Roberts and Stramer [Roberts and Stramer (2001) Biometrika 88(3), 603–621] to d-dimensional correlated diffusions, including multivariate stochastic volatility models, is given. The methodology is illustrated through simulated and real data sets. The Canadian Journal of Statistics 39: 52–72; 2011 © 2011 Statistical Society of Canada Les auteurs considèrent l'inférence sur les diffusions corrélées basée sur la fonction de vraisemblance. Une telle tâche fait face deux obstacles: les contraintes pour s'assurer que la matrice de diffusion est définie positive et l'impossibilité d'évaluer la fonction de vraisemblance. Le premier obstacle est résolu en utilisant la decomposition de Cholesky sur la matrice de diffusion. Pour contourner celui avec la fonction de vraisemblance, nous utilisons une généralisation de la structure de données augmentées proposée par Roberts et Stramer [2001 Biometrika 88(3), 603-621], pour des diffusions corrélées de dimension d, incluant les modèles de volatilité stochastique multidimensionnels. Cette méthode est illustrée avec des données simulées et réelles. La revue canadienne de statistique 39: 52–72; 2011 © 2011 Société statistique du Canada
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