Calendar inputs gave no clear daily Nasdaq return forecast gain with monthly fits in a 2018–2025 archive test

We tested daily Nasdaq Composite log-return forecasts from weekday and month boundary inputs. Each model used ten years of earlier returns and fit once each month. Training prices came from the ALFRED version dated seven calendar days before the first target session of the month. The final test had 2,011 sessions in 2018–2025. The combined calendar model gave 0.085% more mean squared error than the rolling mean. Relative forecast skill was -0.0846%, with a paired 95% three-month block interval of -0.3897% to +0.1394%. Turn-of-month inputs alone also gave no clear gain. Combined sign accuracy was 54.45%, against 55.79% for an always-positive rule. Planned period, block, and outcome-version checks gave no clear primary gain. A post-plan check removed both months with unscheduled full-day closures and gave the same result. The test assumes the target calendar state was known before entry. The package preserves code, source hashes and requests, all 2,766 forecast rows, 132 monthly fit records, uncertainty draws, figures, and reproduction instructions. Independent checks passed. This is a negative forecast result for one index and model family. It does not establish equal performance or trading profit. Raw source prices retain provider rights and remain outside the public package.

return prediction · calendar effects · Nasdaq Composite · dated data · chronological testing · negative result