Scientific Analytics Alliance · Evidence
Validation recordRealized backtestMethodologySample resultsMonitor status
Model validation record
Scientific Analytics Alliance keeps a prospective record of its research models: each forecast is recorded before its outcome is known, and each model is judged only on what it predicted after it was registered. This page is the public summary of that record.
As of 17 Nov 2026. 7 research models under evaluation; 0 have graduated; 1 model rejected (2 cohorts). Next pre-registered evaluation: 1 Dec 2026.
Models and their status
| Model | What it tries to do | Strict forward record | Status |
|---|---|---|---|
| Directional momentum | Predict the direction of the 10-day return | since 6 Oct 2026, 1843 matured observations | Observing. Ranking advantage not established (AUC 0.499, interval 0.452-0.541 includes 0.500). |
| Sentiment probability | Predict the 10-day direction from a sentiment posture | since 6 Oct 2026, 1587 matured observations | Observing. Not established. AUC 0.575 (interval 0.525-0.620), but other pre-registered checks fail: probability loss not better than the base rate; hit rate not above the majority baseline. |
| Sentiment divergence | Flag divergence between sentiment and price | since 6 Oct 2026, 1876 matured observations | Observing. Not established. AUC 0.573 (interval 0.523-0.615), but other pre-registered checks fail: probability loss not better than the base rate; hit rate not above the majority baseline. |
| Early-warning drawdown | Warn before a 10-day drawdown | since 6 Oct 2026, 1819 matured observations | Observing. No verdict yet. Warning ranking interval 0.378-0.443. |
| VaR-limit warning | Warn before a day as large as the -8% house limit | since 6 Oct 2026, 1819 matured observations | Observing. The test passes on the pooled sample, but the forward record spans 31 days; at least 90 are required before a verdict. |
| GARCH 10-day volatility | Forecast volatility better than a naive benchmark | since 6 Oct 2026, 1757 matured observations | Observing. Superiority not established (t = -2.09; at least 1.96 in size is required). |
| 1-day VaR95 coverage | Keep the share of days beyond VaR at 5% | since 6 Oct 2026, 2340 matured observations | Rejected. On the research universe, 10.0% of asset-days fell beyond VaR against 5% nominal: risk understated. The earlier diagnostic cohort was rejected in the opposite direction (2.6%, conservative). |
Two different checks. This page tests single assets, one day ahead: 10.0% of days went past VaR (expected 5%). The Realized backtest counts portfolio records whose time windows overlap; its low count (4 against 33.1 expected) is not evidence that the model is cautious: it is too early to test.
Integrity of the record
87 evaluation runs since 26 Sep 2026; every run is hashed and chained to its predecessor; verified on 17 Nov 2026: confirmed with a recorded correction. Independent time-stamping by an outside authority is not yet in place.
As of 17 Nov 2026.
Research information only. Not investment advice, not a recommendation, not a solicitation to buy or sell any asset. Figures are computed from the stated sources as of the stated date and carry the limits shown with them. Where the evidence is insufficient, the figure is withheld. Market data: CoinGecko. Scientific Analytics Alliance Inc., Delaware.
MethodologyPricingMarket data: CoinGecko
Math first · Agents second · Governor always · Audit forever