Machine-Learning Ranking for Short-Horizon Return Targets in U.S. Mid-Cap Equities

This paper develops and evaluates a machine-learning ranking framework for U.S. mid-capequities under a strategy-consistent short-horizon event definition. A weekly observationis labelled positive only when a stock reaches a 5% upside barrier before its ATR-basedprotective stop within 21 U.S. trading sessions, with both barriers measured from the same5-basis-point slippage-adjusted next-session entry used by the portfolio simulator. Yearlyexpanding-window estimation is purged using each observation’s actual label-end tradingdate rather than a generic business-day approximation. The full weekly panel contains83,283 stock-date observations from July 2022 to August 2026, of which 40,962 (49.18%)satisfy the corrected target-first label. Across 53,387 out-of-sample observations in 2024–2026,Precision@2 is 69.12% against a 48.95% base rate, a lift of 1.412; a five-week moving-blockbootstrap gives a 95% interval of 1.266–1.559 for the lift. The raw gradient-boosting outputis useful for ranking but is not well calibrated as a probability: its Brier score is 0.2609versus 0.2499 for a constant base-rate forecast. The unfiltered Top-2 portfolio earns 86.31%cumulatively over the historical OOS experiment, with 26.65% CAGR when annualized from8 January 2024 through the last actual exit on 27 August 2026, a 1.58 profit factor, and an11.46% maximum drawdown. A pre-specified nested score-threshold procedure selects 0.80 for2024–2025 and 0.70 for 2026 using only prior OOS portfolio evidence; the resulting 42-tradecontinuous portfolio earns 25.72%, with a 1.88 profit factor and 4.87% maximum drawdown.The evidence supports economically meaningful ranking information while remaining subjectto survivorship bias, raw-price corporate-action risk, a short OOS history, and small portfoliosamples. A forward shadow protocol is therefore retained as the next source of genuinely newevidence.