Short-Term Reversal in Chinese A-Shares: Statistical Predictability, Economic Tradability, and Multiple Testing

We study short-term reversal in Chinese A-shares on a fully traceable, point-in-time, delisting-inclusive database, using a monthly signal based on the past three months of returns that skips the most recent trading week. Over 2014:05–2026:02 (142 months), an equal-weighted loser-minus-winner long–short portfolio earns 1.72% per month (Newey–West t=3.64) and rises nearly monotonically across signal deciles (8 of 9 adjacent steps increasing). The effect is not absorbed by standard controls: a CH-3-style factor model leaves an alpha of 1.46% per month (t=3.33), and Fama–MacBeth regressions controlling for size, earnings-to-price, Amihud illiquidity, and industry retain a significant reversal coefficient (t=3.51). These risk-adjustment results survive multiple testing: under a Romano–Wolf step-down across the factor-model and Fama–MacBeth specification families, the featured specifications retain familywise-adjusted p-values of 0.0024 and 0.007, every one of the sixteen Fama–MacBeth specifications survives at adjusted p≤0.014, and all five factor-model specifications at p≤0.0032; the orthogonality also holds against the official Liu–Stambaugh–Yuan factors (alpha 1.88% per month, t=2.81, over the 2014–2021 overlap). Reversal is also not merely the volatility/lottery/turnover complex in disguise: when idiosyncratic volatility, the MAX lottery measure, turnover, beta, and momentum are measured over a window that does not overlap the signal-formation window, they leave the reversal coefficient intact (Fama–MacBeth t=4.3; three-factor alpha 1.32% per month, t=3.1). Such controls appear to absorb reversal only when measured over the same window as the signal—a mechanical artifact that vanishes monotonically as the control window is separated from it (§6.6). The tradable claim is weaker on every margin we examine. The long-only loser portfolio's excess over an equal-weighted universe is 0.25% per month (net-versus-net), loses single-test significance beyond 10–30 bps of proportional cost, and—measured against an investable cap-weighted index, the relevant opportunity cost—is economically large (about 1% per month) but statistically insignificant (t≈1.5). After a max-t correction across 72 strategy-selection candidates the zero-cost excess is not significant (adjusted p=0.288), whether the long side is the loser half or the extreme loser decile. Returns concentrate in less-liquid small- and mid-cap stocks, and the long–short is materially lower in later subsamples (no formally identified trend). We read A-share short-term reversal as a genuine cross-sectional regularity—robust to standard factors and to a stock's pre-existing volatility/lottery/turnover character—that is nonetheless economically constrained as a trade: predictability and tradability diverge once multiplicity and frictions are treated explicitly.