Does The Method Hold Up?
Every Past Match, Forecast Again
742 matches742 finished matches were each forecast again using only the matches played before them. The settings in use called 53.1% of results (basic blend 52.6%) and were closer to the truth on 430 of 742 matches; their chances were within 2.0 points of what happened.
Run Sun 18 Oct, 22:15 UK. Rerun with npm run model:backtest after a model change; this page reads the newest report.
Live Against Rebuilt
graded forecastsSaved before kickoff
- Matches
- 398
- Called
- 56.5%
- Exact
- 9.0%
- Honesty gap
- 3.1 pts
Rebuilt from history
- Matches
- 740
- Called
- 50.1%
- Exact
- 10.3%
- Honesty gap
- 2.3 pts
Forecasts saved before kickoff have called more results than rebuilt ones (+6.4 pts) so far.
Every Setting Compared
same matches for each| Setting | Results called | Exact scores | Over/under 2.5 | Both score | Error score | Honesty gap | Better / worse |
|---|---|---|---|---|---|---|---|
| Basic goals blend | 52.6% | 10.0% | 57.7% | 57.3% | 0.602 | 3.2 pts | – |
| Basic + goal calibration | 52.6% | 9.7% | 56.7% | 56.2% | 0.602 | 3.2 pts | same |
| Dixon-Coles goals | 53.1% | 12.3% | 56.6% | 56.6% | 0.584 | 2.0 pts | 430 / 312 |
| Dixon-Coles + goal calibrationin use | 53.1% | 11.2% | 55.5% | 55.9% | 0.584 | 2.0 pts | 430 / 312 |
Goal Markets
settings in use- Over 1.5about rightsaid 78.2%happened 80.1%
- Over 2.5model too lowsaid 55.6%happened 59.7%
- Over 3.5model too lowsaid 34.5%happened 39.4%
- Both teams scoremodel too lowsaid 55.0%happened 57.8%
Model's average chance how often it happened. Below the actual rate means the model expects fewer goals than are scored.
Where It Helped
results called| Competition | Matches | Basic | In use | Change |
|---|---|---|---|---|
| Turkish Super Lig | 99 | 50% | 52% | +2.0 pts |
| La Liga | 85 | 53% | 49% | −3.5 pts |
| Eredivisie | 79 | 51% | 54% | +3.8 pts |
| Primeira Liga | 71 | 54% | 55% | +1.4 pts |
| Premier League | 69 | 46% | 49% | +2.9 pts |
| Serie A | 68 | 62% | 56% | −5.9 pts |
| Ligue 1 | 63 | 54% | 54% | 0.0 pts |
| Scottish Premiership | 63 | 38% | 41% | +3.2 pts |
| Bundesliga | 54 | 54% | 50% | −3.7 pts |
| UEFA Champions League | 36 | 69% | 72% | +2.8 pts |
| UEFA Europa League | 36 | 53% | 58% | +5.5 pts |
| UEFA Conference Leaguesmall sample | 18 | 67% | 67% | 0.0 pts |
| Belgian Pro Leaguesmall sample | 1 | 100% | 100% | 0.0 pts |
Are The Chances Honest?
per 10% bandHover or tab to a band to compare. Dots above the line: things happened more often than the model said.
Show as table
| Model said | Dixon-Coles + goal calibration | Basic goals blend |
|---|---|---|
| 0–10% | 9% → 5% (37) | – |
| 10–20% | 16% → 14% (244) | 18% → 9% (80) |
| 20–30% | 25% → 24% (937) | 26% → 24% (1098) |
| 30–40% | 35% → 34% (399) | 34% → 33% (461) |
| 40–50% | 45% → 48% (287) | 45% → 48% (368) |
| 50–60% | 55% → 54% (184) | 54% → 65% (187) |
| 60–70% | 64% → 78% (87) | 63% → 81% (32) |
| 70–80% | 74% → 73% (45) | – |
| 80–90% | 82% → 100% (6) | – |
What Each Input Adds
older run- Restkept, less weightslightly less accurate, chances 0.6 pts more honest
- Travelkept, less weightslightly less accurate, chances 0.8 pts more honest
- Everything togetherkept, less weightslightly less accurate, chances 1.0 pts more honest
- Not testable yetno historyInjuries/Lineups, Weather, Club Context, Game Theory: there is no record of what was known before past kickoffs to test them on.
An older run: Premier League only, 380 test matches from 2018–25, written Sun 13 Sept. Each input added to the base model on its own.