Experimental pregame estimates
Model Track Record
Recorded before tipoff. Compared with confirmed outcomes. Only aggregated results are public.
Individual picks
Exact-line More and Less forecasts, counted separately.
Model joint-availability-five-v5
Observed hit rate
50.0%
Decisive results
54
Pushes / Voids
0 / 0
Pending / Recorded
102 / 156
Recorded Nov 6, 2026 – Nov 7, 2026 · Results checked Nov 7, 2026
| Forecast range | Mean forecast | Observed | Hits / n | Push / Void |
|---|---|---|---|---|
| 0–10% | — | — | 0 / 0 | 0 / 0 |
| 10–20% | — | — | 0 / 0 | 0 / 0 |
| 20–30% | 25.4% | 0.0% | 0 / 2Small sample | 0 / 0 |
| 30–40% | 34.2% | 72.7% | 8 / 11Small sample | 0 / 0 |
| 40–50% | 46.3% | 66.7% | 10 / 15Small sample | 0 / 0 |
| 50–60% | 54.1% | 28.6% | 4 / 14Small sample | 0 / 0 |
| 60–70% | 65.4% | 30.0% | 3 / 10Small sample | 0 / 0 |
| 70–80% | 74.6% | 100.0% | 2 / 2Small sample | 0 / 0 |
| 80–90% | — | — | 0 / 0 | 0 / 0 |
| 90–100% | — | — | 0 / 0 | 0 / 0 |
Whole slips
A hit requires every selected leg to hit; not an average of individual chances.
Model joint-availability-five-v5
Observed hit rate
28.3%
Decisive results
46
Pushes / Voids
0 / 0
Pending / Recorded
36 / 82
Recorded Nov 6, 2026 – Nov 7, 2026 · Results checked Nov 7, 2026
| Forecast range | Mean forecast | Observed | Hits / n | Push / Void |
|---|---|---|---|---|
| 0–10% | 5.0% | 0.0% | 0 / 3Small sample | 0 / 0 |
| 10–20% | 15.0% | 41.2% | 7 / 17Small sample | 0 / 0 |
| 20–30% | 25.6% | 30.8% | 4 / 13Small sample | 0 / 0 |
| 30–40% | 33.9% | 12.5% | 1 / 8Small sample | 0 / 0 |
| 40–50% | 44.0% | 20.0% | 1 / 5Small sample | 0 / 0 |
| 50–60% | — | — | 0 / 0 | 0 / 0 |
| 60–70% | — | — | 0 / 0 | 0 / 0 |
| 70–80% | — | — | 0 / 0 | 0 / 0 |
| 80–90% | — | — | 0 / 0 | 0 / 0 |
| 90–100% | — | — | 0 / 0 | 0 / 0 |
Methodology & privacy
- Selection · first-calculated-v1
- Starting November 6, 2026, retain the first successfully calculated, schedule- and roster-verified forecast for each unique game/player/market/period/scoring-rule/line/side/model combination. Individual analysis records both sides. Whole slips use the first unique sorted set of legs per model version. Repeated visits, leg ordering and availability revisions do not add samples. Custom Lines are included. This is a visitor-requested cohort, not every board offer or a random sample; it can have selection bias.
- Before tipoff only
- The server rechecks the scheduled game before recording. The database supplies the recording time and rejects timestamps at or after the earliest tipoff. Forecast time, exact lines/sides, model version and eventual results stay server-side. Old archives are excluded because they were collected under a different protocol; no predictions are reconstructed after outcomes.
- Denominators & forecast ranges
- Observed hit rate = hits ÷ (hits + misses). Pending, pushes and voids are excluded and reported separately. A whole slip with any void is classified void; otherwise any push is classified push, even if another leg missed. A hit requires all legs to hit. Mean forecast uses the same decisive sample; forecast chances originally include separate push/void probability, so observed rates are conditional on decisive results. Ranges include their lower bound and exclude their upper bound, except 90–100% includes 100%.
- Confirmed results & corrections
- Full-game box scores include overtime. Missing required stats, unconfirmed absences and postponements remain pending. Only explicit DNP or cancellation is void. The scheduled five-minute result check rotates through completed candidates, including settled results, so provider corrections replace the current result without changing the original forecast or adding a sample. Large backlogs can take multiple checks.
- Small samples & privacy
- Fewer than 30 decisive results is a small sample; even larger samples do not establish reliability. Overlapping picks, opposite sides and shared-game slips are correlated, not independent evidence. Model versions are never pooled. No account identifiers, player names, saved lines or individual slip contents are published.