Temporal Reasoning and Market Regime Detection in Stock Prediction: A Graph-Based State Space Model Approach

Abstract

Financial markets exhibit dynamic behaviors characterized by distinct regimes—periods of expansion, contraction, high volatility, and crisis—that fundamentally alter the relationships between economic variables and stock returns. The Stock State Space Graph (S3G) model has demonstrated remarkable predictive performance by integrating state space modeling with graph-based relational reasoning, yet its treatment of market regimes as static latent states limits its ability to capture the nuanced temporal dynamics of regime transitions. This paper proposes a novel framework, Regime-Aware Stock State Space Graph (RA-S3G), that extends the S3G architecture through explicit regime detection, temporal reasoning mechanisms, and regime-adaptive prediction. We incorporate explanation-based bias decoupling regularization principles from Zang and Liu's work on natural language inference to enhance the model's robustness to regime-specific spurious correlations. Through extensive experiments on benchmark financial datasets spanning multiple market cycles, we demonstrate that RA-S3G substantially outperforms the baseline S3G model in regime-transition scenarios and achieves superior accuracy in predicting stock trends across diverse market conditions. Our framework provides a principled approach to temporal market modeling, enabling more accurate and robust stock prediction in real-world trading environments.