Variance updates gave no clear forecast gain for at least two Nasdaq daily falls of 2% or more in 20 sessions in a 2015–2025 archive test

We tested whether variance updates within simulated paths improve forecasts of repeated large stock-market losses. The event was at least two Nasdaq Composite daily price falls of 2% or more in the next 20 exchange sessions. Each input came from a dated ALFRED record, with an observation cutoff two sessions before entry. We used 134 separate targets in 2015–2025. Four planned targets failed input freshness and remain in the record. The fixed and recursive models used the same starting variance and past shock pool. The recursive model reduced mean Brier loss by 0.001591, or 0.91%. The paired 95% six-origin block interval was −0.004517 to +0.007462. Thus the main test gave no clear gain. A planned secondary score for the full loss count improved by 0.018701, with interval 0.005136 to 0.032254. This gain was uncertain in 2020–2025 alone. Labels used the first complete common source record after each target. One label arrived 116 days late. Later data versions changed no count labels. All 134 states, 536 forecast rows and 134 receipt dates passed separate checks. A fresh run matched 12 core files byte for byte. The package preserves code, source hashes, all forecasts, both simulation seeds, exclusions, uncertainty draws, negative findings, figures and repeat instructions. This is a retrospective test on one index and a reused market sample. It does not establish reliable mean-return prediction or trading profit.

Nasdaq Composite · loss count · point-in-time data · filtered historical simulation · chronological evaluation · negative result