Residual Mean Reversion in U.S. Large Caps, 2017–2026: A Reproducible Re-examination of Avellaneda–Lee Statistical Arbitrage

We re-examine the principal component analysis (PCA) residual, Ornstein–Uhlenbeck (OU) statistical arbitrage strategy of Avellaneda and Lee on a universe of 78 liquid U.S. large-capitalisation stocks over January 2017 to September 2026, using a strictly causal, fully reproducible implementation. With the original paper’s default parameters the strategy earns a gross annualised Sharpe ratio of 0.60 (annual return 2.8% on unit gross exposure, t = 1.86), with a one-way breakeven transaction cost of only 4.0 basis points (bps); at a realistic 2–3 bps the Sharpe ratio falls to 0.15–0.30. Its returns load significantly on the Fama–French short-term reversal factor (t = 5.3), and its alpha relative to a seven-factor model is 2.2% per year (t = 1.42). The edge does not survive standard corrections for multiple testing: across an 80-configuration parameter grid the median gross Sharpe ratio is 0.22, the deflated Sharpe ratio of the default configuration is 0.38, and a 20-day block bootstrap places zero inside the 95% confidence interval. Two methodological findings are of independent interest. First, the OU “speed of mean reversion” estimated by the standard procedure is largely mechanical: pure noise yields a median estimated reversion time of 8.4 days (real data: 7.6) and 98% of noise series pass the usual mean-reversion filter. Second, for non-daily rebalancing the realised Sharpe ratio depends dramatically on the arbitrary choice of rebalance day—weekly variants of the same strategy range from 0.09 to 0.86—and an exploratory single-phase test led us to an incorrect “crowding” inference that phase-averaged (staggered) portfolios overturn. A placebo experiment on simulated return panels with the real covariance structure but no time-series predictability shows the strategy’s realised Sharpe ratio exceeds all 40 placebo draws, so a small genuine effect is likely present, although a stricter sign-flip placebo that preserves volatility clustering puts it only at the 95th percentile. Extending the sample to 2003–2026 makes the gross edge statistically significant (Sharpe ratio 0.48, t = 2.35) and shows it is time-varying rather than monotonically decaying: negative in 2003–2006 and strongest in 2010–2015. With a tradeable sector exchange-traded fund (ETF) hedge whose trading is costed, breakeven falls to 2.9 bps. The effect is real but too small to be economically exploitable at retail-to-institutional trading costs, and the binding constraint is cost, not capacity. Declaration of generative AI use: All research questions, hypotheses, study designs and conclusions are the author’s own, including the decision to re-audit the author’s earlier exploratory results, which identified data errors in them. Under the author’s direction, generative AI tools (Claude Code, Anthropic) wrote most of the analysis code and ran it, and assisted with drafting, editing and formatting the text, and with typesetting. The author reviewed all code, results and text and takes full responsibility for the content of this paper.