STORM: A Spatio-Temporal Factor Model Based on Dual Vector Quantized Variational Autoencoders for Financial Trading

Key findings from this study

• The study found that representing factors as multi-dimensional embeddings rather than single values enables capture of complex relationships and dependencies in asset pricing.
• The researchers demonstrate that dual vector quantized variational autoencoders extract complementary temporal and spatial features to improve individual stock pattern recognition.
• The authors report that discrete codebook-enforced orthogonality and diversity enhance factor robustness across different trading periods and downstream applications.

Overview

STORM proposes a spatio-temporal factor model for financial trading based on dual vector quantized variational autoencoders. The model extracts and fuses stock features from temporal and spatial perspectives, representing factors as multi-dimensional embeddings rather than single values. Discrete codebooks enforce orthogonality and diversity among learned factors to improve quality and robustness.

Methods and approach

The architecture employs dual vector quantized variational autoencoders to capture temporal patterns of individual stocks alongside overall market conditions. Features extracted from both temporal and spatial dimensions undergo fine-grained and semantic-level fusion and alignment. Discrete codebooks cluster factor embeddings to ensure orthogonality and enable systematic factor selection for trading applications.

Results

Extensive experiments on portfolio management tasks across two stock datasets and individual trading on six specific stocks demonstrate STORM's superior performance relative to baseline models. The model exhibits flexibility in adapting to diverse downstream trading tasks while maintaining competitive or improved predictive accuracy. The multi-dimensional factor embeddings with enforced diversity outperform representations derived from conventional latent factor approaches.

Implications

Enhanced factor quality and diversity through vector quantization and multi-dimensional embeddings may improve asset pricing accuracy and excess return capture in systematic trading strategies. The spatio-temporal architecture addresses limitations of prior variational autoencoder approaches that modeled aggregate market conditions without adequately capturing individual stock dynamics. Orthogonal factor representations could facilitate more interpretable factor selection and portfolio construction methodologies.

Scope and limitations

This summary is based on the study abstract and available metadata. It does not include a full analysis of the complete paper, supplementary materials, or underlying datasets unless explicitly stated. Findings should be interpreted in the context of the original publication.