MarkovFormer: Market Regime-Aware Hierarchical State Space Architecture for Stock Movement Prediction

Abstract

Stock market prediction is uniquely challenging because financial time series are fundamentally non-stationary: the statistical properties of returns, correlations, and volatility change over time as markets transition between distinct regimes such as bull markets, bear markets, high-volatility crises, and low-volatility consolidations. Most existing deep learning approaches—including recurrent networks, transformers, and graph neural networks—treat market dynamics as stationary processes and learn a single set of model parameters that applies uniformly across all market conditions. This design choice is inherently limiting, as the optimal predictive model for a quiet bull market is fundamentally different from that for a volatile bear market. In this paper, we propose MarkovFormer (Markov Switching Transformer), a novel hierarchical architecture that explicitly models market regime transitions using a Markov Switching mechanism and adapts its predictive parameters accordingly. MarkovFormer consists of three key components: (i) a Regime Detection Module (RDM) that uses a latent Markov switching process to classify the current market state from historical price data; (ii) a Regime-Conditioned State Space Encoder (RSE) that maintains regime-specific stock state representations learned through regime-conditioned transition matrices; and (iii) a Cross-Regime Attention Mechanism (CRAM) that enables information flow across regimes, allowing the model to leverage knowledge from similar historical regimes to improve predictions in the current one. By conditioning predictions on the detected market regime, MarkovFormer dynamically adjusts the weighting of short-term momentum, mean-reversion, and volatility signals based on which regime is currently active. Extensive experiments on three major markets (S&P 500, CSI 300, and KOSPI 200) demonstrate that MarkovFormer consistently outperforms ten competitive baselines, including S3G (Lu, Hu, and Zhang, 2026), the state-of-the-art stock state space graph model, achieving an average improvement of 5.4% in directional accuracy and 9.7% in Matthews Correlation Coefficient (MCC) across all datasets. Most notably, during high-volatility crisis periods—including the 2020 COVID-19 market crash and the 2022 global rate-hike cycle—MarkovFormer substantially outperforms S3G, which lacks explicit regime awareness, demonstrating the critical importance of modeling market structure non-stationarity.