Accuracy
Backtest results (2026-08-20)
Tested on 5 leagues × 2023-2025 seasons — 5,256 played matches. The 2025/26 evaluation season covered 1752 fixtures: 1709 evaluated and 43 skipped (insufficient match history for a new/promoted team).
| Test | Result |
|---|---|
| Beats a frequency baseline | 5/5 leagues, statistically significant |
| Calibration (within 0.05 tolerance) | 4/5 leagues |
Calibration error by league
Closing line value (CLV)
Did the market move towards our pick after we published it? For every settled match in this site's competitions we take the first pick we published with market prices beside it, and compare the market's de-vigged probability for that pick at that moment with the closing probability just before kick-off. Positive means the market moved towards our pick.
| Measure | All picks | 2+ pt edge |
|---|---|---|
| Picks | 2,053 | 407 |
| Average CLV | +0.2 pts | +0.9 pts |
| Moved our way | 47% | 56% |
| Moved against | 41% | 34% |
| Unchanged | 12% | 10% |
Why we don't claim "99% accurate"
You will see other sites advertise a "99% accurate", "sure win" or "zero-risk" football prediction. These claims are lies, for a simple mathematical reason: football is not that predictable. Even a well-calibrated model on the favourite side of a 1X2 market typically lands in the 45–65% probability range for its top pick — not 99%.
A site claiming near-certainty is either (a) cherry-picking a tiny, favourable sample after the fact, (b) quietly redefining "accurate" to mean something other than "the outcome I predicted happened", or (c) simply making the number up to sell a tip. We publish 1709 real evaluation-season predictions, including the ones we got wrong, because a track record you can't fail is not a track record.
Remember: predictions here are probabilistic estimates, not betting advice. If gambling stops being fun, see our responsible gambling page. 18+ only.