Prior-day large-stock returns stopped predicting next-day US small-stock index returns after 2000 and gave no tradable ETF forecast gain in 2006–2026

Question: does the daily lead from large stocks to small stocks (Lo and MacKinlay 1990) still give out-of-sample forecasts, and does skill measured on CRSP portfolio indexes carry over to ETFs that can be traded? Method: one fixed OLS model predicts the next-day return from the prior-day large-stock return and the prior-day own return. The model uses a 1,260-session window and a monthly refit. The benchmark is the past mean. We wrote the plan before scoring. Targets: Kenneth French value-weighted size indexes (Lo 30, Lo 20, Lo 10, Med 40; Hi 10 as the lead) and the IWM, IJR, IWC and MDY ETFs (SPY as the lead), at close-to-close, overnight, next open-to-close (tradable) and one-day-delay timings. The main test has 5,197 sessions, 2006-01 to 2026-08. Results: for the Lo 30 index, out-of-sample R2 was positive in all fourteen five-year periods from 1931 to 2000 (0.9% to 21.6%), and -2.1% to +0.4% in the periods after 2000. The fitted lead coefficient changed from positive to negative (mean -0.16 in 2006-2026). In 2006-2026 the Lo 30 R2 was -0.46% (95% block interval -1.34% to +0.40%), so the primary hypothesis failed. IWM R2 was -0.17% close-to-close and -0.006% (-0.86% to +0.89%) open-to-close. No ETF case had a positive point estimate. The index-minus-IWM R2 difference was -0.29% (-0.72% to +0.13%). A long/short IWM open-to-close strategy earned -9.2% a year at 2 bp per side (-17.6% to -0.7%). No strategy beat buy-and-hold. A post hoc 1996-2000 check with a 252-session window gives weak evidence that MDY did not share the index predictability before it went away. Limitations: single current data vintages; one model form; ETF data cannot cover the pre-2001 period of strong index predictability; fills at the official open are assumed. Code, plan, all forecasts, fits, scores, bootstrap draws, hashes and a byte-identical replay are included. Realised returns are in a retained SCI dataset.

lead-lag · cross-autocorrelation · stale prices · small-cap stocks · ETF · out-of-sample R2 · daily return forecasting · paper trading · transaction costs · Kenneth French data library · IWM · negative result