Adaptive equity investing strategies via attention-based deep reinforcement learning

This paper develops a deep reinforcement learning framework that combines temporal self-attention (ATT) with a long short-term memory (LSTM) encoder for adaptive stock investment. Temporal self-attention captures salient market signals, including anomalies in volatility indices and search-heat shocks, while filtering noise. The LSTM extracts long-range dependencies from the attention-weighted features. The Proximal Policy Optimization (PPO) algorithm then produces risk-adjusted trading decisions. We test the framework on five regime-heterogeneous markets that span both developed and emerging economies: the CSI 300, the Hang Seng Index (HSI), the S&P 500, the Euro Stoxx 50 (SX5E), and the Nikkei 225 (N225). The proposed framework achieves the highest Sharpe ratio in every market. Holding the encoder fixed, PPO substantially outperforms Advantage Actor-Critic (A2C), Deep Q-Network (DQN), and Soft Actor-Critic (SAC). Its probability-ratio clipping is particularly useful in the very low signal-to-noise environment of equity markets. For interpretability, we build a multi-method attribution framework that fuses attention saliency, gradient sensitivity, integrated gradients, and occlusion analysis with adaptive weights based on Spearman consistency. A cross-market meta-analysis shows that search heat and the panic index are dominant universal signals, while the effectiveness of short-horizon technical indicators depends on market microstructure. The study provides a unified methodological framework for alpha discovery in global equity markets and offers empirical insights for dynamic cross-market hedging.