Prediction accuracy & track record
Verified pre-match predictions, graded against the result — including the ones we got wrong.
Covering 13 October 2026 to 16 October 2026. Every figure below carries the number of predictions behind it, because a percentage without a sample size is not something you can check.
Recent performance
Short windows move quickly, so they sit beside the all-time record rather than replacing it. Each period is anchored to the latest match graded in the record.
| Period | Graded | Result | AH hit rate | AH ROI | AH CLV |
|---|---|---|---|---|---|
| Last 30 days | 61 | 47.5% | 70.0% | 65.6% | +1.6% |
| Last 90 days | 61 | 47.5% | 70.0% | 65.6% | +1.6% |
| Last 365 days | 61 | 47.5% | 70.0% | 65.6% | +1.6% |
| All time | 61 | 47.5% | 70.0% | 65.6% | +1.6% |
By market
| Market | Correct | Graded | Hit rate |
|---|---|---|---|
| Match result | 29 | 61 | 47.5% |
| Exact score | 7 | 61 | 11.5% |
| Goals line | 36 | 59 | 61.0% |
| Both teams to score | 36 | 59 | 61.0% |
| Asian handicap | 3.5 | 5 | 70.0% |
The exact scoreline is listed alongside the rest even though it lands around one time in ten — it is genuinely hard, and dropping it would quietly lift every other number on this page. Asian handicap is included only where a generated line was stored. Markets that neither won nor lost, such as a goals line the match landed exactly on, are excluded from the graded count rather than counted as wins.
Are our confidence scores honest?
When we say a call is 70% likely, it ought to come in about seven times in ten. Here is what actually happened, grouped by the confidence we stated at the time:
| We said | Predictions | Actually right | Gap |
|---|---|---|---|
| 40–44% | 12 | 33.3% | -10.7 |
| 45–49% | 22 | 36.4% | -10.4 |
| 50–54% | 24 | 58.3% | +7.3 |
| 55–59% | 1 | 100.0% | +45.0 |
| 65–69% | 1 | 100.0% | +31.0 |
| 75–79% | 1 | 100.0% | +23.0 |
Are our Asian handicap probabilities calibrated?
Handicap cover probability is measured separately from match-result confidence. These are conservative lower bounds derived from the stored scoreline grid, grouped by the probability published before kick-off.
Closing-line value
We compare the exact Asian handicap side and line published with its 1xBet price shortly before kick-off. Positive CLV means the published price was better than the later market price. A different handicap line is never substituted.
| Model cover | Predictions | Win-equivalent | Gap |
|---|---|---|---|
| 50–54% | 4 | 62.5% | +10.3 |
| 70–74% | 1 | 100.0% | +26.8 |
By competition
Competitions with at least 50 graded predictions. Anything thinner is listed as a total rather than a rate — a hit rate from a dozen matches tells you nothing.
| Competition | Graded | Result | Score | Goals | BTTS | AH |
|---|---|---|---|---|---|---|
| 10 smaller competitions | 61 | too few to report a rate | ||||
Asian handicap records begin with predictions that contain a generated handicap. Quarter-line half wins count as 0.5 win-equivalents; half losses count as losses, and pushes are excluded. The confidence-calibration table above remains tied to match-result confidence, because that number does not describe the handicap bet.
This record is what improves the model
The point of grading every prediction is not the scoreboard — it is the correction. This record is the measurement the model gets corrected against, so it is judged on what happened rather than on its own assumptions.
The overconfidence in the table above is the first correction we have made. It was visible only because we grade ourselves; nothing else would have surfaced it. Predictions are now published with their confidence mapped onto what confidences like them have actually achieved, which is why the figures on the site are lower than they used to be.
The reliability table still shows the gap, because it covers every prediction we have ever graded — including the thousands made before the correction existed. As predictions made under it accumulate, this is where you will see whether it worked. If the gap does not close, that will be visible here too.
More results make each correction sharper and more specific. Competitions where we have hundreds of graded predictions can be adjusted on their own evidence; thinner ones will lean on the overall picture until they have accumulated enough to stand alone, rather than being tuned on a handful of matches and mistaking noise for a pattern.
Every prediction records which version of the model produced it. That is deliberate, and it cuts against us: it means a change either shows up in these numbers or it did not work, and you can check which. A track record is only worth anything if it can prove us wrong as readily as right — so expect these figures to move, and to move on the evidence.
How these are counted
- Markets settle on ninety minutes. A tie level at full time and won in extra time is graded as the draw it was. Extra time and penalties never turn a wrong call into a right one.
- Only verified pre-match publications count. Database-stamped forecast versions saved before kickoff are eligible regardless of the result. Historical records, backtests and imports without publication evidence are excluded.
- The forecast is frozen at publication. We grade the latest verified version saved before the fixture's kickoff. Later edits cannot replace a graded forecast.
- Saved previews may differ from the graded version. Accuracy uses the publication ledger, rather than the current article or its record timestamps.
Football is high variance, and a percentage over a few dozen matches is mostly noise. Judge any prediction service — including this one — over hundreds of results, which is why the sample size sits beside every figure above.