Beyond Black Boxes: An Energy-Based Unified Framework for Interpretable Stock Selection

Abstract

Quantitative stock selection from large-scale market data is critical for achieving excess returns in financial markets. A persistent challenge is tail risk, which triggers non-stationary regime shifts that invalidate learned patterns and distort portfolio logic. Despite advances in deep learning, existing approaches cannot jointly model how tail events reshape market dynamics or offer interpretable tools for navigating such disruptions. To address these limitations, we propose PRIME (Potential Robust Integrated Macro Energy), a framework that grounds return generation and tail-risk control in energy-based modeling and game theory from a momentum perspective. A semantic encoder decomposes market forces into bullish momentum, bearish resistance, and frictional dissipation, casting stock valuation as an energy state within a long-short game. A macro-aware modulation mechanism reshapes the energy landscape across bull-bear transitions, adapting scoring geometry to tail-driven non-stationary dynamics. The aggregation process is formulated through potential game theory, whose Nash equilibrium properties guarantee stable ranking under distributional shifts. A risk-guardian module closes the loop by detecting anomalous energy spikes as crash signals and filtering them from portfolio construction. Experiments on S&P 500 and CSI 500 benchmarks, including ablation studies, demonstrate that PRIME achieves approximately 10% improvement over state-of-the-art baselines with theoretical guarantees on convergence and ranking consistency.