Two seasonal means gave no clear US market return forecast gain in 240 months in 2005–2025 under assumed archive dates

We tested separate November–April and May–October means for monthly US market excess returns. Twenty annual fits used official French Data Library archives, with inputs stopped in July. Historical use assumes each archive preserved its stated version and was available by September 30. In 240 months from October 2005 through September 2025, seasonal means had 0.012% more squared error than a pooled past mean. The paired 95% two-year circular block interval for error reduction was −1.76% to +2.16%. Both forecasts were always positive and matched an always-positive direction accuracy of 65.0%. At 10 basis points per dollar traded, a seasonal stock-and-cash paper rule earned 8.65% a year against 10.51% for the pooled rule, with lower exposure and risk. Its annual mean-variance utility difference was −0.745 percentage points, with interval −4.044 to +2.862. Fixed history, shrinkage, label and block checks gave no clear full-test seasonal gain. Separate scalar checks passed. A fresh run matched 76 analysis and account files. The package preserves code, provenance, forecasts, all fixed results, uncertainty draws, figures and repeat instructions. Direct source-derived outcome ledgers remain separately retained under provider terms. The test uses one known seasonal split, slow updates, assumed release dates and reused US history. It does not establish reliable prediction or profit. Twelve future paper forecasts remain unscored.

stock returns · seasonality · out-of-sample forecasts · historical data archives · trading costs · negative result