How we test these predictions
Each method is replayed on every past draw using only the draws before it, and scored on whether the drawn mark was in its top five. Chance is 5 in 36. Every draw is independent.
Prediction audit
| Method | Backtest hit rate | Chance | 95% interval | p | Status |
|---|---|---|---|---|---|
| Random top 5 | 11.2% | 13.9% | 9.3% to 13.4% (expected 11.7% to 16.1%) | ||
| History says: overdue | 15.3% | 13.9% | 13.1% to 17.7% | p = 0.12 | |
| Follower | 11.4% | 13.9% | 9.5% to 13.6% | p = 0.99 · perm 0.96 | |
| Slot | 13.3% | 13.9% | 11.3% to 15.7% | p = 0.70 · perm 0.48 | |
| Weekday | 12.9% | 13.9% | 10.9% to 15.2% | p = 0.82 · perm 0.64 | |
| Hot | 13.8% | 13.9% | 11.7% to 16.1% | p = 0.56 · perm 0.28 | |
| Weighted signals | 14.5% | 13.9% | 12.4% to 16.9% | p = 0.30 | |
| Model (softmax ensemble, default weights) | 13.5% | 13.9% | 11.5% to 15.9% | p = 0.63 |
No measurable evidence of a slot or weekday bias in this sample (p = 0.44 and 0.76)
Pre-registered method: how the audit works. Re-run after every draw with 1000 shuffles. A method enters a ranked list only when its interval clears 13.9% and holds on unseen draws.