Ten-year sign counts gave no clear US monthly sign forecast gain over a normal model in 2005–2025 with assumed archive dates

We tested whether past positive-month counts improve monthly US market excess-return sign probabilities over a normal model. Both primary models used the same 120 past months. Twenty annual French Data Library archives supplied 240 later outcomes from October 2005 to September 2025. Historical release by September 30 and preservation of the old archive files are assumptions. The count model had Brier loss 0.229218, versus 0.236032 for normal. Its gain was 0.006814, with a paired 95% two-year-block interval from -0.003094 to +0.016486. The primary test gave no clear gain. The count rule always predicted a positive direction and matched the always-positive accuracy of 65%. A descriptive loss split added after scoring showed that a better average rise probability supplied most of the point gain. Fixed release-delay, revision, crisis-exclusion and block-length checks retained an uncertain primary gain. An expanding-count model beat an expanding-normal model on a secondary score interval, but both always selected a positive direction. Independent checks and a complete numerical repeat passed. The package preserves code, provenance, 6720 forecasts, fit summaries, all planned results, 5.04 million uncertainty gain draws, figures, negative findings and seven unscored November 2026 paper probabilities. Restricted outcomes remain separately retained under provider terms. This is a retrospective benchmark on a reused US market path. It does not establish timing skill or trading profit.

stock returns · direction forecasts · Brier loss · data vintages · chronological evaluation · negative result