Jump-Residual State Learning for A-Share Stock Selection: A Residual Identification and Portfolio Validation Study

Abstract

This study examines whether discontinuous price-movement information can improve cross-sectional stock selection in the Chinese A-share market after conventional return-prediction signals are controlled. A jump-residual state learning framework is proposed. The framework first estimates a non-jump machine-learning baseline, then learns the residual return component using robust jump indicators, idiosyncratic jump intensity, post-jump memory, graph-based peer confirmation and auditable semantic-event variables. This residual design tests a stricter incremental-value question than directly appending raw jump factors to a predictive model. Using a filtered HS300-100 universe, a five-day forward excess-return target, five-day rebalancing and transaction-cost-adjusted top-10 portfolios, JRS_Residual achieves a test-period cumulative return of 44.08% and a Sharpe ratio of 2.419, compared with CSI300 at 9.03%, EqualWeight at 21.44% and ML_Base at 24.87%. However, Newey-West RankIC and bootstrap tests indicate that statistical significance is mixed, and semantic augmentation is constrained by incomplete text-event coverage. The evidence therefore supports a cautious conclusion: residualized jump-state information exhibits sample-period out-of-sample value, while semantic enhancement remains a reproducible but not yet statistically confirmed extension.