A four-session turn-of-month rule gave no clear daily Nasdaq return forecast gain in a 2015–2025 archive test

We tested a known four-session turn-of-month rule for daily Nasdaq Composite price returns. Each annual fit used an earlier ALFRED data version. The test had 2,766 exchange sessions in 2015–2025. The calendar model had 0.109% more squared error than a past-mean forecast. Its out-of-sample R squared was -0.109%, with a paired 95% six-month block interval of -0.326% to +0.043%. The turn-window mean was 1.28 basis points below the mean on other sessions, with interval -10.54 to +8.21. A fixed ten-year fit also gave no clear gain. All four fitted models had the same direction accuracy as an always-positive rule. At an assumed five-basis-point one-way cost, a turn-window paper allocation gave a compound annual price return of 1.08%, compared with 3.15% for constant 4/21 exposure. The latter had much lower volatility. These are price-only simulations with zero cash interest. The source-calendar check removed one holiday price row. The source also had a long update delay. The point-in-time claim concerns model inputs; the calendar is a current reconstruction. Code, provenance, all forecasts, fit states, aggregate scores, 32,000 resampling draws, figures, and checks are preserved. The result applies to one index, rule and period. It does not establish future prediction skill or trading profit.

return forecasts · turn of the month · Nasdaq Composite · point-in-time data · calendar effects · negative result · trading costs