[Paper Review] Copula Variational LSTM for High-dimensional Cross-market Multivariate Dependence Modeling

Jia Xu, Longbing Cao|arXiv (Cornell University)|May 9, 2023

Stock Market Forecasting Methods4 citations

TL;DR

This paper proposes WPVC-VLSTM, a novel variational LSTM architecture integrated with regular vine copulas to model high-dimensional, non-normal cross-market dependencies across heterogeneous financial time series. By jointly learning temporal dynamics and asymmetric latent dependencies, it outperforms benchmarks in cross-market portfolio forecasting and risk estimation.

ABSTRACT

We address an important yet challenging problem - modeling high-dimensional dependencies across multivariates such as financial indicators in heterogeneous markets. In reality, a market couples and influences others over time, and the financial variables of a market are also coupled. We make the first attempt to integrate variational sequential neural learning with copula-based dependence modeling to characterize both temporal observable and latent variable-based dependence degrees and structures across non-normal multivariates. Our variational neural network WPVC-VLSTM models variational sequential dependence degrees and structures across multivariate time series by variational long short-term memory networks and regular vine copula. The regular vine copula models nonnormal and long-range distributional couplings across multiple dynamic variables. WPVC-VLSTM is verified in terms of both technical significance and portfolio forecasting performance. It outperforms benchmarks including linear models, stochastic volatility models, deep neural networks, and variational recurrent networks in cross-market portfolio forecasting.

Motivation & Objective

• To address the challenge of modeling complex, high-dimensional cross-market dependencies across heterogeneous financial markets.
• To capture both observable and latent dependencies in non-normal, non-stationary, and asymmetric multivariate financial time series.
• To integrate variational sequential learning with copula-based dependence modeling for improved representation of dynamic, hierarchical couplings.
• To enhance cross-market portfolio forecasting and risk management by modeling explicit and implicit inter-market influences.

Proposed method

• Proposes WPVC-VLSTM, a hybrid model combining variational autoencoders with long short-term memory (LSTM) networks for sequential dependence learning.
• Uses regular vine copulas to model non-normal, long-range, and asymmetric dependence structures among latent variables.
• Employs variational inference to learn deep, probabilistic representations of cross-market couplings and interactions.
• Introduces a weighted partial vine copula to capture hierarchical and heterogeneous dependence patterns in high-dimensional settings.
• Trains the model end-to-end using backpropagation through time and likelihood maximization under non-Gaussian assumptions.
• Validates the architecture on real-world cross-market financial data, focusing on portfolio risk and return prediction.

Experimental results

Research questions

• RQ1 How can high-dimensional, non-normal cross-market dependencies be effectively modeled in multivariate financial time series?
• RQ2 What is the impact of integrating variational sequential learning with copula-based dependence modeling on forecasting accuracy?
• RQ3 Can the proposed model capture both explicit and implicit couplings between heterogeneous financial markets more effectively than existing benchmarks?
• RQ4 How does the model perform in cross-market portfolio risk estimation under non-IID and non-stationary market conditions?
• RQ5 To what extent does the use of regular vine copulas improve the modeling of asymmetric and long-range dependencies in latent variable spaces?

Key findings

• WPVC-VLSTM significantly outperforms linear models, stochastic volatility models, deep neural networks, and variational recurrent networks in cross-market portfolio forecasting.
• At the 99% confidence level, WPVC-VLSTM achieves a Value at Risk (VaR) of 0.81%, compared to 4.07% for ARMA-GARCH and 10% for GP-Vol, indicating superior risk estimation.
• The LRUC test shows WPVC-VLSTM’s p-values of 0.857 (99%), 0.536 (95%), and 0.294 (90%) are well above the 0.05 threshold, indicating no significant rejection of the model’s accuracy.
• In the LRIT test, WPVC-VLSTM achieves p-values of 0.128 (99%), 0.665 (95%), and 0.208 (90%), further confirming its statistical robustness.
• The model reduces VaR estimation errors by over 80% compared to ARMA-GARCH and GP-Vol at the 99% level, demonstrating strong predictive performance.
• WPVC-VLSTM effectively captures hierarchical, asymmetric, and non-linear couplings across heterogeneous markets such as equities, FX, and commodities, outperforming models that assume Gaussian or symmetric dependencies.