Track record
How good are our forecasts?
The honest version
Every game page carries a forecast: the chance each side wins, and the most likely score. This page grades those forecasts against the results, using only what the model said before kick-off. Nothing here is re-predicted with hindsight.
Two things to know. Youth soccer is streaky, so no forecast will be right every week. And a good forecast is not one that is always right, it is one that is right as often as it says: when we say 70%, the favourite should win about 70% of the time. The calibration table below is the real test.
Calibration
when we said X, how often it happened
| We said the favourite had… | Games | Average we said | Favourite actually won | Fit |
|---|---|---|---|---|
| Toss-up (under 45%) | 2,779 | 42% | 44% | on the money (+2) |
| 45–55% | 4,162 | 50% | 54% | on the money (+4) |
| 55–65% | 3,718 | 60% | 63% | on the money (+4) |
| 65–75% | 3,158 | 70% | 73% | on the money (+4) |
| 75–85% | 2,477 | 80% | 85% | we were too cautious (+6) |
| 85% and up | 2,296 | 91% | 95% | on the money (+3) |
"Favourite" is whichever outcome the forecast rated highest: a home win, an away win or, rarely, a draw. Bands with no games are left out. Draws happen in 13% of games and are the hardest call for any model.
By league
| League | Games | Favourite won | Always picking the home team | Brier score |
|---|---|---|---|---|
| ECNL Regional League Boys | 5,060 | 66% | 47% | 0.466 |
| ECNL Regional League Girls | 4,504 | 69% | 46% | 0.438 |
| ECNL Boys | 2,396 | 62% | 45% | 0.499 |
| ECNL Girls | 1,804 | 67% | 46% | 0.459 |
| Pre-ECNL Boys | 1,740 | 70% | 46% | 0.428 |
| Pre-ECNL Girls | 1,201 | 69% | 47% | 0.441 |
| Elite Academy League | 1,036 | 67% | 46% | 0.440 |
| Elite Academy League 2 | 847 | 68% | 50% | 0.426 |
| Elite Academy League National | 2 | 50% | 50% | 0.560 |
Headline numbers
The Brier score adds up the squared gap between each forecast and what happened, so 0 is perfect and a coin-flip style guess scores about 0.667. The baseline is a forecast that only knows how often home wins, draws and away wins happen overall.
Where the forecasts come from
2 of the graded games use the forecast recorded on the site the night before kick-off. 18,588 older games, from before recording began, were graded by giving the same model only the results that existed before that date.
The model is an opponent-adjusted goals model: every team's attack and defence are rated from its results, weighted by who it played and how recently, with a small home edge. It knows nothing about injuries, weather, travel or referees, and says so on every page.