Deep Learning Enhanced Multivariate GARCH

Haoyuan Wang, Chen Liu, Minh-Ngoc Tran, Chao Wang · 2025

Read Deep Learning Enhanced Multivariate GARCH by Haoyuan Wang, Chen Liu, Minh-Ngoc Tran, Chao Wang (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2506-02796.txt.

Introduction

Haoyuan Wang and three colleagues present a 2025 finance paper grafting a neural network onto a classic volatility model. Markets swing together, and the BEKK family of models tracks those joint swings with sound statistical structure but stiff behavior. The authors thread a long short-term memory network through it, letting the model bend with nonlinear, shifting market moods while keeping the older model's readable form.

The hybrid, LSTM-BEKK, aims squarely at known weaknesses: sticky volatility clusters and lopsided co-movement when markets fall. Across several equity markets, it forecast portfolio risk out of sample better than the classic versions. The authors pitch the blend as a template: econometric bones for interpretability, deep learning muscle for the patterns the bones alone miss.

Catalog record

CREATORHaoyuan Wang, Chen Liu, Minh-Ngoc Tran, Chao Wang
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE99 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2506-02796

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Deep Learning Enhanced Multivariate GARCH Haoyuan Wang, Chen Liu, Minh-Ngoc Tran, Chao Wang Abstract This paper introduces a novel multivariate volatility modeling framework, named Long Short-Term Memory enhanced BEKK (LSTM-BEKK), that integrates deep learning into multivariate GARCH processes. By combining the flexibility of recurrent neural networks with the econometric structure of BEKK models, our approach is designed to better capture nonlinear, dynamic, and high-dimensional dependence structures in financial return data. The proposed model addresses key limitations of traditional multivariate GARCH-based methods, particularly in capturing persistent volatility clustering and asymmetric co-movement across assets. Leveraging the data-driven nature of LSTMs, the framework adapts effectively to time-varying market conditions, offering improved robustness and forecasting performance. Empirical results across multiple equity markets confirm that the LSTM-BEKK model achieves superior performance in terms of out-of-sample portfolio risk forecast, while maintaining the interpretability from the BEKK models. These findings highlight the potential of hybrid econometric-deep learning models in advancing financial risk management and multivariate volatility forecasting. Keywords : multivariate volatility modeling, Long Short-Term Memory, portfolio optimization, high-dimensional finance. 1 Introduction Modeling financial market volatility has long been a central topic in econometrics due to its critical role in risk management, asset pricing, and portfolio optimization. Engle, 1982 pioneered this line of research with the introduction of the Autoregressive Conditional Heteroskedasticity model, which characterizes time-varying volatility as a function of past shocks. This foundational framework was later generalized by Bollerslev, 1986 through the GARCH model, which incorporated both lagged innovations and past variances, enabling improved modeling of persistent volatility behavior. In parallel to the GARCH family, stochastic volatility (SV) models emerged as an important alternative, modeling volatility as an unobserved latent process governed by its own stochastic dynamics. This latent-state formulation allows SV models to capture the stylized facts often observed in financial time series ( Taylor, 1994 ; Asai et al., 2006 ) . While univariate volatility models are effective in capturing the dynamics of individual asset volatility, financial markets are inherently multivariate, with assets exhibiting strong comovements and spillovers. Accurate modeling of such joint dynamics is essential for systemic risk monitoring, portfolio allocation, and derivative pricing. In this context, the conditional covariance matrix plays a central role by describing the time-varying co-movement among asset returns. As highlighted in Bollerslev et al., 1988 , Engle, 2002 and Bauwens et al., 2006 , multivariate volatility modeling enables the quantification of interdependencies across assets and enhances the effectiveness of financial decision-making under uncertainty. To extend volatility modeling to multivariate settings, multivariate GARCH (MGARCH) models have been proposed as natural generalizations of the univariate framework. Among these, the BEKK model introduced by Engle and Kroner, 1995 stands out due to its flexible parameterization and being a direct extension of univariate GARCH. The BEKK specification guarantees the positive definiteness of the conditional covariance matrix by construction and can capture dynamic spillovers between asset returns. This structural advantage makes it particularly appealing for applications requiring robust covariance estimation in financial risk modeling and forecasting. Empirical studies ( Silvennoinen and Teräsvirta, 2009 ; Fang et al., 2015 , see, e.g.,) have demonstrated the adaptability of MGARCH frameworks in capturing time-varying correlations and volatilities under extreme events, reinforcing their relevance in financial risk management. Traditional MGARCH formulations such as BEKK encounter severe computational bottlenecks in high-dimensional settings due to the rapidly expanding parameter space ( Ledoit and Wolf, 2012 ; Ledoit and Wolf, 2015 ) . To address this scalability issue, Engle, 2002 proposed the Dynamic Conditional Correlation (DCC) model, which simplifies estimation by decoupling univariate volatility and correlation dynamics. This reduction in complexity allows DCC to accommodate a larger number of assets while still capturing time-varying dependencies. Extensions such as the Student- t DCC ( Ku, 2008 ) and Asymmetric DCC ( Lai and Sheu, 2011 ) further enhance the modeling robustness under heavy tails and asymmetric shocks, while the Dynamic Equicorrelation model ( Engle and Kelly, 2012 ) further improves computational tractability for the DCC framework. These methodological contributions collectively underpin the development of modern multivariate volatility modeling, which is central to understanding the dynamic behavior of financial markets and the evolving interdependencies among assets. While models such as BEKK, DCC and their extensions have significantly advanced the multivariate volatility modeling, their reliance on the simple summation of lagged covariance matrices and outer products of return vectors limits their adaptability to complicated patterns and structural shifts often observed during periods of financial stress. These limitations motivate the integration of more expressive modeling techniques, such as deep learning models, into multivariate volatility modeling, offering potential new directions for advancing volatility modeling in high-dimensional financial applications. Recent advances in deep learning provide compelling techniques for modeling complicated and high dimensional sequential data. Recurrent Neural Networks (RNNs), and Long Short-Term Memory (LSTM) networks in particular, offer a powerful mechanism for learning temporal and comovement dependencies in sequential multivariate data. Their ability to capture long-range interactions and nonlinear patterns has led to various successful applications across a range of large-scale industry level tasks ( Goodfellow et al., 2016 ) . This paper proposes a novel hybrid model, the LSTM-BEKK, which combines the econometric rigor of the BEKK model with the adaptive capabilities of LSTMs. The LSTM-BEKK model leverages the structural strengths of BEKK while enhancing it with LSTM’s ability to capture nonlinear and long-term dynamics. By allowing LSTM-generated components to directly influence the time-varying covariance matrix, this framework provides a highly flexible tool for analyzing the evolving relationships among financial assets, particularly in high-dimensional contexts. The design of the LSTM-BEKK model is inspired by recent advancements in integrating deep learning with univariate volatility models, which have demonstrated superior predictive accuracy and the ability to capture nonlinearities in volatility. For example, the work by Nguyen et al., 2022 explored the effectiveness of deep learning in enhancing GARCH-based volatility modeling. These studies highlight the advantages of incorporating neural networks into traditional econometric frameworks, improving the adaptability and forecasting performance of volatility models. Extending this approach to a multivariate setting introduces unique challenges, such as ensuring positive definiteness of the covariance matrix and managing the curse of dimensionality ( Ledoit and Wolf, 2012 ) . The primary innovation of the LSTM-BEKK model lies in its ability to utilize economic information instruments and adapt to changing market conditions. Unlike traditional BEKK models constrained by their inflexible parametric structure, the LSTM-BEKK model dynamically adjusts itself to capture evolving market relationships. Furthermore, the LSTM’s capacity to learn complex patterns enhances the model’s responsiveness to turbulent market periods, improving the accuracy of volatility forecasts. Substantial empirical results demonstrate that the LSTM-BEKK model outperforms traditional BEKK and DCC models in terms of predictive accuracy, as measured by standard evaluation metrics such as out-of-sample negative log-likelihood and annualized volatility of global minimum variance portfolio. These findings are robust across datasets covering portfolios constructed from the top companies by market capitalization in Japan, the U.S., and the U.K., reflecting the model’s strong generalization capability across different market environments. Moreover, in low-dimensional settings, the interpretability of the LSTM-BEKK framework is enhanced by visualizing individual variance and covariance trajectories, which reveal the model’s ability to capture both abrupt volatility spikes and directional shifts in inter-asset correlations—especially during periods of market stress. These insights affirm not only its

Read the complete text