Every forecast, graded and published
We record how each prediction resolved and publish the result, whether it flatters us or not. Sample sizes are shown on every figure. Nothing here is selected after the fact.
Data as of October 29, 2026View the raw data (JSON)
How accurate are ScorePredicts football predictions?
Best Picks
The single highest-conviction selection per fixture, across all markets.
Value Analysis
Selections where the model's probability diverges most from consensus market pricing.
Hit rate by period
| Prediction type | 7 days | 30 days | All time |
|---|---|---|---|
| Best Picks | 73.2%n = 1,531 | 73.2%n = 5,086 | 71.4%n = 10,470 |
| Value Analysis | 69.1%n = 333 | 66.0%n = 788 | 60.0%n = 6,294 |
Monthly reports
Each month's figures on a page of their own, final from the 2nd of the following month.
Compared to guessing
An accuracy figure means nothing on its own. The only useful question is what a naive rule would have scored on exactly the same matches.
Measured across every graded fixture, not a selected subset. Home advantage is why the naive rule scores as well as it does; the gap above it is what the model contributes. n = 8,178.
Hit rate over time
Each line is a seven-day rolling average of settled selections, with a node marking the days behind it. Days with nothing settled break the line rather than being bridged across.
| Date | Best Picks | Value Analysis |
|---|---|---|
| 29.09.2026 | 74.7% | 59.3% |
| 30.09.2026 | 75.0% | 60.0% |
| 01.10.2026 | 74.9% | 58.5% |
| 02.10.2026 | 73.9% | 56.8% |
| 03.10.2026 | 72.7% | 58.9% |
| 04.10.2026 | 71.3% | 54.2% |
| 05.10.2026 | 71.3% | 51.9% |
| 06.10.2026 | 70.3% | 53.3% |
| 07.10.2026 | 70.5% | 53.3% |
| 08.10.2026 | 71.0% | 54.2% |
| 09.10.2026 | 72.5% | 59.2% |
| 10.10.2026 | 73.3% | 54.8% |
| 11.10.2026 | 74.1% | 54.9% |
| 12.10.2026 | 73.5% | 57.7% |
| 13.10.2026 | 73.6% | 57.0% |
| 14.10.2026 | 73.7% | 56.8% |
| 15.10.2026 | 73.8% | 57.8% |
| 16.10.2026 | 73.5% | 60.8% |
| 17.10.2026 | 72.5% | 67.7% |
| 18.10.2026 | 73.1% | 68.9% |
| 19.10.2026 | 73.3% | 68.4% |
| 20.10.2026 | 74.1% | 68.7% |
| 21.10.2026 | 73.7% | 69.8% |
| 22.10.2026 | 73.7% | 68.0% |
| 23.10.2026 | 73.8% | 67.7% |
| 24.10.2026 | 74.4% | 67.6% |
| 25.10.2026 | 73.2% | 67.7% |
| 26.10.2026 | 73.4% | 68.9% |
| 27.10.2026 | 73.0% | 68.9% |
| 28.10.2026 | 73.2% | 69.1% |
Are the probabilities honest?
This is the test that matters. We group every match-result forecast by the probability we published, then plot what actually happened. A calibrated model sits on the diagonal.
What this chart coversMatch result across every graded fixture, not just selected picks - a selected subset could not test calibration honestly. Buckets holding too few fixtures to plot are merged into the end points, and point size scales with sample. Vertical bars are 95% intervals: a thin bucket is an uncertain reading, not a precise miss.
Brier score: 0.1955Mean squared error between published probability and outcome. Lower is better; 0.25 is what you would score by predicting 50% on everything.
Calibration by probability bin
| Probability band | Probability we published | How often it actually happened | 95% interval | Fixtures |
|---|---|---|---|---|
| 0.05-0.20 | 14.8% | 12.6% | 11.5% – 13.8% | 3294 |
| 0.20-0.25 | 22.9% | 21.8% | 20.6% – 23.0% | 4367 |
| 0.25-0.30 | 27.3% | 26.3% | 25.1% – 27.5% | 5443 |
| 0.30-0.35 | 32.3% | 32.5% | 30.7% – 34.3% | 2649 |
| 0.35-0.40 | 37.5% | 37.5% | 35.5% – 39.6% | 2136 |
| 0.40-0.45 | 42.4% | 44.5% | 42.2% – 46.9% | 1771 |
| 0.45-0.50 | 47.4% | 48.8% | 46.2% – 51.4% | 1398 |
| 0.50-0.55 | 52.4% | 55.1% | 52.1% – 58.1% | 1034 |
| 0.55-0.60 | 57.3% | 61.9% | 58.5% – 65.2% | 819 |
| 0.60-0.65 | 62.3% | 64.3% | 60.4% – 68.2% | 575 |
| 0.65-0.70 | 67.4% | 72.3% | 68.0% – 76.3% | 452 |
| 0.70-0.75 | 72.4% | 75.3% | 69.8% – 80.0% | 271 |
| 0.75-0.80 | 77.3% | 80.9% | 74.4% – 86.1% | 173 |
| 0.80-0.95 | 85.1% | 84.9% | 78.3% – 89.7% | 152 |
Interval is the 95% range the observed figure could take at that sample size. Narrower means more matches in the bin.
Why picks land below their stated probability
Both products land below their own stated probabilities. Best Picks carry an average published probability of % and hit %. Value Analysis averages % and hits %. Neither gap is hidden here, because the explanation matters more than the number.
It is a selection effect, not a broken model. Both products rank the model's own estimates and surface the top of the list, and conditioning on the highest outputs preferentially selects the cases where the model happened to read too high. Any system that ranks its own output and publishes the best of it shows this pattern - it is the same reason a fund's strongest backtested trades underperform once they are traded live.
The calibration curve above is the check, because it covers every graded fixture rather than a selected subset. Across the range where most matches sit, our published probabilities track observed frequencies closely, and if anything run slightly low rather than high. That is the opposite of an overconfident model, and it is what tells you the gap above is selection rather than miscalibration.
The practical reading: treat a stated probability as a probability rather than a promise, and expect a ranked shortlist to run a few points under its headline figure.
Which markets the picks came from
A hit rate across a mixed basket is not the same claim as a hit rate on match result alone. These are the six most-used markets in each basket, with the remainder grouped, so you can judge both headline figures for yourself.
Best Picks
- Home Win or Draw46.3% of picksn = 2,357
- Draw or Away Win20.6% of picksn = 1,050
- Home Win17.0% of picksn = 867
- Away Win6.2% of picksn = 315
- Over 1.5 Goals5.9% of picksn = 300
- Over 2.5 Goals3.4% of picksn = 173
- Other markets (2)0.5% of picksn = 24
Value Analysis
- Home Win50.9% of picksn = 401
- Home Cards Under 2.510.2% of picksn = 80
- Away Cards Under 2.56.1% of picksn = 48
- Corners Over 8.55.8% of picksn = 46
- Corners Under 10.54.3% of picksn = 34
- Away Cards Over 1.53.3% of picksn = 26
- Other markets (20)19.4% of picksn = 153
How these numbers are produced
Graded automatically, once
Every prediction is settled against the final result after the match finishes. Figures are never revised afterwards to look better, and nothing is removed from the record.
Abandoned fixtures are excluded
Matches that were cancelled, abandoned, walked over or administratively awarded never receive a win or loss. fixtures fall into this category in the period shown, and they are excluded from every figure rather than counted as either outcome.
Every figure carries its sample size
A percentage drawn from a handful of results is unreliable, so it is never shown on its own: every hit rate on this page is printed beside the number of graded predictions behind it. Nothing is withheld for being small. Calibration buckets holding fewer than fixtures are merged into the end points or left off the chart, because a dozen fixtures plotted as a point would bend the curve.
Past performance is not future performance
Everything here describes what has already happened. It is not a projection, not a target, and not a rate you should expect to personally experience. Football remains genuinely uncertain and no model changes that.
These are the rules this page follows. For what the numbers actually mean, why 85% accuracy claims are not credible, and how to check any prediction site including this one, read Are AI football predictions accurate?.
Measures used on this page
The three statistics behind the figures above, and where they come from. None are ours; all are standard.
- Brier score
- The scoring rule used to grade probability forecasts. It squares the distance between the stated probability and what happened, so a confident wrong answer is penalised more than a hedged one. Lower is better; 0.25 is what you get by saying 50% to everything. Brier, G. W. (1950). Verification of forecasts expressed in terms of probability. Monthly Weather Review, 78(1), 1-3.
- Wilson score interval
- The 95% interval shown beside each observed figure in the calibration table. It is used instead of the textbook normal interval because it stays sensible with small samples and with rates near 0% or 100%, which is exactly where calibration bins sit. Wilson, E. B. (1927). Probable inference, the law of succession, and statistical inference. Journal of the American Statistical Association, 22(158), 209-212.
- Isotonic regression calibration
- The method that maps raw model output onto probabilities that mean what they say. It fits a monotonic curve against graded results, so ordering is preserved while the stated 70% is pulled towards whatever 70% forecasts actually did. Zadrozny, B., & Elkan, C. (2002). Transforming classifier scores into accurate multiclass probability estimates. KDD 2002, 694-699.
See today's forecasts
Up to four best picks a day are free, with every live score, fixture and table on the site.