A fixed evidence rule crossed 20 for scaled Nasdaq loss forecasts in 802 archive tests from 2018 to 2025

We tested a sequential evidence rule on an existing Nasdaq Composite loss-probability benchmark. The rule waits for each selected outcome before it selects another forecast. This gives 802 tests and 75 large-loss events in 2018 to 2025 under stated archive-date assumptions. Filtered historical simulation had 3.34% less mean Brier loss than unscaled historical simulation. Its equal mixture of four fixed evidence products ended at 25.49 and reached 31.86, above the fixed threshold of 20. Against a normal reference, final evidence was 20.58. Sign forecasts were unchanged. The source audit found one revised event label and a maximum label delay of 119 days. A direct proof and a two-event example show why a product of overlapping delayed factors need not be valid under an issue-time null. The package has first eligible labels, complete selection records, all nine historical paths, code, source hashes, 30000 simulation records, figures, scalar checks, and an exact repeat of 13 result files. All 24201 parent forecasts also matched a raw-source rebuild. This is a method check on reused history. The favorable parent result was known, and the wait rule was added after the initial audit. Nominal evidence values are not new confirmatory tests. The result does not establish reliable direction prediction or profit after trading costs.

stock forecast evaluation · sequential evidence · delayed outcomes · point-in-time data · Brier score · Nasdaq Composite · reproducibility · negative direction result