Shrinkage toward zero gave no clear mean return forecast gain for three US factors in 2005–2025 under assumed archive dates

We tested whether a fixed James–Stein rule that shrinks three past return means toward zero improves monthly forecasts. The targets were US market excess, size and value factor returns. Twenty annual origins gave 240 outcomes in October 2005–September 2025. Each fit used 120 months through July from its own annual source file. Forecasts stayed fixed for twelve later monthly targets. The test assumes that the annual files preserve their stated versions and were available by September 30. The primary covariance-scaled squared loss was 0.079% lower than for the past mean, with a paired 95% two-year-block interval of −0.159% to +0.302%. Loss was 0.168% lower than for zero, with interval −1.631% to +2.135%. The planned requirement for a clear gain over both benchmarks was not met. A fixed half mean also gave no clear gain. Training-window, covariance, source-revision and crisis checks did not establish improvement over both main benchmarks. The shrinkage rule kept every forecast sign unchanged. Its market MSE point estimate was 0.333% higher than the past mean. This is a dated-input extension on a reused market sample, not an independent replication or a trading test. Code, provenance, all forecasts, scores, uncertainty draws, figures, passed checks and a byte-exact replay are preserved. Individual source returns remain in a separate retained record under provider rights. The package also has unscored forecasts for November 2026–October 2027. The result does not establish reliable return prediction or profit after costs.

return prediction · mean shrinkage · James-Stein · dated data · forecast evaluation · negative result · Fama-French factors