Scalable inference for a full multivariate stochastic volatility model

Πέτρος Δελλαπόρτας, Michalis K. Titsias, Kateřina Petrová, Anastasios PlataniotisJournal of Econometrics 2015 · arXiv

Abstract

We introduce a multivariate stochastic volatility (MSV) model for asset returns that imposes no restrictions on the structure of the volatility matrix and treats all its elements as functions of latent stochastic processes. When the number of assets is prohibitively large, we propose a factor multivariate stochastic volatility model in which the variances and correlations of the factors evolve stochastically over time. Inference is achieved via a carefully designed feasible and scalable Markov chain Monte Carlo (MCMC) algorithm that combines two computationally important ingredients: it utilizes invariant-to-the-prior Metropolis proposal densities for simultaneously updating all latent paths and achieves quadratic, rather than cubic, computational complexity when evaluating the multivariate normal densities required. We apply our modelling and computational methodology to 571 stock daily returns of the Euro STOXX index over a 10-year period.

Introduction

In financial econometrics, accurately predicting the time-varying covariance matrix ($\Sigma_t$) of asset returns is essential for portfolio selection, asset pricing, and risk management. However, modeling high-dimensional covariance matrices presents two significant challenges. First, the matrix must remain positive-definite at every time step to be mathematically valid. Second, the "curse of dimensionality" causes the number of parameters to grow quadratically with the number of assets ($N$). For a universe of 500 stocks, a researcher must track over 125,000 values per time step.

To manage this complexity, traditional models often rely on restrictive assumptions, such as constant correlations or independent underlying factors. This paper proposes a "full" MSV model that allows every element—both variances and correlations—to evolve stochastically, while maintaining computational feasibility for hundreds of assets.

The Full Multivariate Stochastic Volatility Model

The core of the proposed model is a spectral decomposition of the covariance matrix: $\Sigma_t = P_t \Lambda_t P_t^T$. This approach decomposes the matrix into two components:

  1. Eigenvalues ($\Lambda_t$): These represent the variances of the principal components. The authors model the log of these eigenvalues as latent Gaussian autoregressive processes.
  2. Eigenvectors ($P_t$): To ensure $P_t$ remains an orthogonal rotation matrix, it is decomposed into a product of $N(N-1)/2$ Givens rotations. Each rotation is controlled by a single angle, which is also modeled as a latent stochastic process.

Scaling via Factor MSV

For very large datasets, the authors introduce a Full Factor MSV model. Unlike standard factor models that assume factors are independent, this model allows the factors themselves to have time-varying correlations. This captures "contagion" effects where volatility in one market sector spills over into others, a phenomenon frequently observed during financial crises.

Computational Innovation

The primary barrier to using full MSV models has historically been the computational cost. Standard Gaussian likelihood evaluations require matrix inversions and determinants, which typically scale at $O(K^3)$ complexity.

Quadratic Complexity

The authors developed a recursive algorithm that exploits the sparse structure of Givens rotations. This innovation reduces the computational burden to $O(K^2)$, making it possible to scale the model to hundreds of factors. As the number of factors ($K$) increases, the efficiency gains over standard methods grow linearly.

Auxiliary Langevin Sampler

To perform Bayesian inference, the authors utilize an Auxiliary Langevin Sampler for MCMC. In high-dimensional spaces, standard MCMC often struggles to converge (or "mix"). This sampler uses gradient information to propose updates for all latent paths simultaneously. Crucially, the proposal is designed to be "invariant to the prior," allowing the Markov chain to take larger, more effective steps through the parameter space without getting stuck.

Empirical Application: Euro STOXX 600

The methodology was tested on daily returns for 571 stocks from the Euro STOXX 600 index over a 10-year period (2007–2014). This dataset included over 36,000 missing values due to varying market holidays, which the model's factor structure handled seamlessly.

Key Findings

  • Predictive Power: Using Bayes factors, the "Full Factor" model (with correlated factors) demonstrated overwhelming evidence of superior performance compared to models assuming independent factors.
  • Optimal Dimensionality: For the 571-stock dataset, approximately 30 factors were found to provide the best out-of-sample forecasting ability.
  • Crisis Identification: The model successfully captured major economic events, showing visible spikes in simultaneous correlation and volatility during the 2009 global financial crisis and the 2012 Eurozone crisis.

Limitations and Future Directions

While the model represents a significant step forward in scalable volatility modeling, certain areas remain for future research:

  1. Distributional Assumptions: The current framework assumes Gaussian (normal) returns. Financial data often exhibits "fat tails," where extreme events occur more frequently than a normal distribution predicts.
  2. Leverage Effects: The model does not currently account for the "leverage effect," where volatility tends to increase more sharply following price drops than price gains.
  3. Complexity of Extensions: Incorporating non-Gaussian distributions or leverage effects while maintaining $O(K^2)$ scalability remains a significant technical challenge.

Conclusion

By combining spectral decomposition with a novel recursive algorithm and a prior-invariant MCMC sampler, this paper provides a robust framework for high-dimensional stochastic volatility modeling. The ability to model correlated factors at scale allows for a more realistic representation of financial markets, providing better tools for risk assessment in complex portfolios.

MATLAB software for this paper is available at: http://www.aueb.gr/users/mtitsias/code/msv.zip

Summary updated

Abstract

We introduce a multivariate stochastic volatility model for asset returns that imposes no restrictions to the structure of the volatility matrix and treats all its elements as functions of latent stochastic processes. When the number of assets is prohibitively large, we propose a factor multivariate stochastic volatility model in which the variances and correlations of the factors evolve stochastically over time. Inference is achieved via a carefully designed feasible and scalable Markov chain Monte Carlo algorithm that combines two computationally important ingredients: it utilizes invariant to the prior Metropolis proposal densities for simultaneously updating all latent paths and has quadratic, rather than cubic, computational complexity when evaluating the multivariate normal densities required. We apply our modelling and computational methodology to $571$ stock daily returns of Euro STOXX index for data over a period of $10$ years. MATLAB software for this paper is available at http://www.aueb.gr/users/mtitsias/code/msv.zip.
BibTeX
@article{2015scalable,
  author = {Πέτρος Δελλαπόρτας and Michalis K. Titsias and Kateřina Petrová and Anastasios Plataniotis},
  title = {{Scalable inference for a full multivariate stochastic volatility model}},
  journal = {Journal of Econometrics},
  year = {2015},
}
Figure