The FFBS Estimation of High Dimensional Panel Data Factor Stochastic Volatility Models

Guobin Fang, Huimin Ma, Michelle Xia, Bo ZhangarXiv 2019 · DOI · arXiv

Abstract

This paper proposes a novel Panel Data Factor Stochastic Volatility Model (PDFSVM) that integrates both observable and unobservable factors within a multivariate stochastic volatility framework. The model is structured into three components: a mean equation, a volatility equation, and a factor volatility evolution process. To address the computational challenges of high-dimensional parameter estimation, the authors employ a Bayesian Markov Chain Monte Carlo (MCMC) approach utilizing the Forward Filtering Backward Sampling (FFBS) algorithm and the Kalman Filter Recursive Algorithm (KFRA). Numerical simulations demonstrate the robustness and consistency of the estimation method. An empirical application to the Chinese stock market reveals that while observable factors influence Internet finance and traditional financial firms similarly, their responses to unobservable latent factors differ significantly.

Introduction

In modern financial econometrics, accurately modeling the time-varying volatility of asset returns is essential for risk management and portfolio optimization. While traditional models like GARCH or standard Stochastic Volatility (SV) models are effective for individual assets, they often falter when applied to high-dimensional panel data. The "curse of dimensionality"—where the number of parameters to be estimated grows exponentially with the number of assets—makes many multivariate models computationally intractable.

Furthermore, most existing frameworks focus exclusively on either observable market factors (such as indices or trading volume) or unobservable latent factors (hidden shocks). This paper bridges that gap by introducing the PDFSVM, a unified model that captures both types of influences in a high-dimensional setting, providing a more comprehensive view of market dynamics.

The PDFSVM Framework

The proposed model decomposes asset returns into three distinct, interconnected equations:

  1. The Mean Equation: This captures the linear relationship between asset returns and a combination of observable and unobservable factors.
  2. The Volatility Equation: This models the time-varying variance of the error terms, following a stochastic process rather than a deterministic one.
  3. The Factor Volatility Evolution: This describes how the volatility of the latent (unobservable) factors themselves changes over time, allowing for a deeper layer of risk modeling.

By structuring the model this way, the authors can account for common market shocks while still allowing for idiosyncratic volatility within specific sectors or assets.

Estimation via FFBS and MCMC

The primary technical contribution of the paper is the application of the Forward Filtering Backward Sampling (FFBS) algorithm within a Bayesian MCMC framework. Estimating high-dimensional SV models is notoriously difficult because the likelihood functions are often intractable.

To overcome this, the authors use a "blocking" strategy. They group parameters—such as regression coefficients, factor loadings, and volatility terms—into blocks and sample them iteratively. The FFBS algorithm, supported by the Kalman Filter, operates in two stages:

  • Forward Filtering: The algorithm moves through the time series to predict the state of the system one step ahead based on available information.
  • Backward Sampling: The algorithm then moves backward through the data to "smooth" those estimates, ensuring that the latent factor paths are consistent across the entire sample period.

This approach ensures that the estimation remains computationally efficient even as the number of assets ($N$) and time periods ($T$) increase.

Empirical Analysis: Internet Finance vs. Traditional Banking

The authors applied their model to a dataset of 20 companies listed on the Chinese stock market, consisting of 10 Internet financial firms and 10 traditional regional commercial banks. The data spanned from November 2017 to October 2018 and included variables such as daily stock returns, trading volume, and circulation market value.

Key Findings

  • Consistency in Observables: The study found a significant positive correlation between a company's circulation market value and its expected return. This effect was consistent across both Internet-based and traditional financial firms.
  • Divergence in Latent Factors: A striking difference emerged in how the two sectors responded to unobservable shocks. Internet financial companies exhibited highly consistent factor loadings (specifically, positive loadings on the first two latent factors and negative on the third). In contrast, traditional banks showed fragmented and inconsistent loadings. This suggests that while these sectors share the same broader market environment, they are driven by fundamentally different hidden risk drivers.
  • Algorithm Robustness: Prior to the empirical study, the authors conducted simulations across eight different model configurations (varying $N$, $T$, and the number of factors $p$). In all cases, the FFBS algorithm produced small Root-Mean-Square Errors (RMSE) and highly significant t-statistics, confirming its reliability for high-dimensional data.

Conclusion and Future Directions

The PDFSVM and its associated FFBS estimation method provide a powerful tool for analyzing complex financial panels. By successfully integrating observable and unobservable factors, the model offers a more nuanced understanding of market risk than traditional methods.

However, the authors note several avenues for future research. Current model diagnostics like AIC and BIC are difficult to apply to MCMC-based models, and future work could refine the use of the Deviance Information Criterion (DIC). Additionally, incorporating "jumps" (sudden price movements) and spatial correlations—where the volatility of one firm affects its economic neighbors—could further enhance the model's predictive power in volatile markets.

Summary updated

Abstract

In this paper, We propose a new style panel data factor stochastic volatility model with observable factors and unobservable factors based on the multivariate stochastic volatility model, which is mainly composed of three parts, such as the mean equation, volatility equation and factor volatility evolution. The stochastic volatility equation is a 1-step forward prediction process with high dimensional parameters to be estimated. Using the Markov Chain Monte Carlo Simulation (MCMC) method, the Forward Filtering Backward Sampling (FFBS) algorithm of the stochastic volatility equation is mainly used to estimate the new model by Kalman Filter Recursive Algorithm (KFRA). The results of numeric simulation and latent factor estimation show that the algorithm possesses robustness and consistency for parameter estimation. This paper makes a comparative analysis of the observable and unobservable factors of internet finance and traditional financial listed companies in the Chinese stock market using the new model and its estimation method. The results show that the influence of observable factors is similar to the two types of listed companies, but the influence of unobservable factors is obviously different.
BibTeX
@misc{fang2019ffbs,
  author = {Guobin Fang and Huimin Ma and Michelle Xia and Bo Zhang},
  title = {{The FFBS Estimation of High Dimensional Panel Data Factor Stochastic Volatility Models}},
  year = {2019},
  eprint = {1901.10516},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1901.10516},
}
Figure