A one-year weight half-life gave no clear joint forecast gain for three US equity factors in 2005–2025 archive tests

We tested whether recent returns improve joint monthly forecasts of US market, size, and value returns. Each of 20 annual decisions used its own French Data Library archive. Fits used 120 months through July. Forecasts stayed fixed from October through the next September. The test has 240 monthly vectors in 2005–2025. All models shared the same mean forecast and a 50% cut in estimated correlations. The primary covariance rule halved input weights for each extra year of age. Against equal weights, its mean Gaussian log score gain was -0.0487 nats per month. A paired 95% annual-block interval was -0.3690 to +0.2318. The test showed no clear gain and did not prove harm. A fixed half-and-half covariance blend gained 0.0716 nats, with interval -0.0288 to +0.1802. Data-version checks also gave no clear primary gain. Nominal 90% joint regions covered 82.9% to 87.1% of outcomes across models. The main limits are only 20 annual decisions, old inputs, a reused US factor sample, working normal distributions, and assumed archive availability by September 30. Labels are the earliest annual records in the collection, not first monthly releases. We evaluated no trading strategy. The package preserves the fixed plan, code, provenance, all forecasts and scores, uncertainty draws, figures, independent checks, and unscored future paper forecasts. A clean run matched all 14 analysis files.

covariance forecasts · historical data releases · exponential weights · equity factors · chronological tests · negative result