Direct answer · Metrics before marketing
How accurate are AI football predictions?
There is no single defensible accuracy rate. The answer changes with the market, league, period, sample and baseline. A headline percentage becomes meaningful only when every forecast was fixed before kick-off, every eligible miss remains in the ledger and the complete probabilities are scored with uncertainty.
Published by Football Proof AI · Updated · Editorial technical note; not externally peer reviewed · Educational analysis, not betting adviceCanonical publication record
Abstract
A direct, evidence-first answer to how accurate AI football predictions are, separating markets, hit rate, probability quality, sample uncertainty, publication integrity and relevant baselines.
- Author and publisher
- Football Proof AI
- Technical report
ai-football-prediction-accuracy-guide/1.0.0- Published
- Last modified
- Release status
- Current release
- Review status
- Editorial technical note; not externally peer reviewed
- Version history
ai-football-prediction-accuracy-guide/1.0.0: Initial public release.
The useful answer
Accuracy is a measurement contract, not a universal number
“AI football accuracy” can mean the largest 1X2 probability was correct, a BTTS call landed, an over/under line settled, or an exact score matched. Those are different tasks with different base rates. They cannot be placed in one leaderboard without a common target, sample and scoring rule.
Home, draw or away
Top-pick hit rate is readable, but it ignores how probability mass was distributed across the two alternatives.
BTTS or over/under
A binary percentage is not directly comparable with a three-outcome 1X2 percentage.
Correct score
Many possible scorelines make raw hit rate structurally lower than broad market accuracy.
Same headline · Different evidence
“90% accurate” can describe three very different samples
All rows below have the same observed hit rate. Their Wilson 95% intervals differ because a small sample leaves much more uncertainty. Even the largest row still does not prove that the ledger is complete or that forecasts were public before kick-off.
| Correct / settled | Observed hit rate | Wilson 95% interval | What is still unknown |
|---|---|---|---|
| 9 / 10 | 90.0% | 59.6%–98.2% | Publication timing, denominator completeness, market definition and baseline |
| 90 / 100 | 90.0% | 82.6%–94.5% | Publication timing, denominator completeness, market definition and baseline |
| 900 / 1000 | 90.0% | 88.0%–91.7% | Publication timing, denominator completeness, market definition and baseline |
Enter any advertised top-pick hit-rate sample in the PROOF-5 AI football prediction claim checker.
Measure the probability, not only the pick
A credible evaluation needs more than hit rate
Hit rate asks whether the most likely outcome happened. Brier score and log loss use the complete probability distribution. Calibration asks whether events labelled 60% occur about 60% of the time.
- Hit rate
- Easy to read, but sensitive to outcome prevalence, decision rules and market selection
- Brier score
- Squared error across the declared probabilities; lower is better and the convention must be named
- Log loss
- Penalises confident mistakes sharply; lower is better and probabilities must stay away from impossible zeroes
- Calibration
- Compares stated probability with observed frequency in sufficiently large, pre-declared buckets
- Uncertainty
- Intervals and sample warnings prevent a short hot streak from masquerading as stable model skill
The accuracy-metrics research note defines the formulas and explains why summed and class-averaged multiclass Brier scores must not be compared without conversion.
The hurdle matters
Random guessing is rarely the only honest baseline
Equal random 1X2 choices imply a 33.3% expected hit rate, but real competitions do not produce home wins, draws and away wins equally. A useful model should also be compared with a frozen base-rate, common-outcome or market-implied reference on the same matches.
- 01
Freeze the rule
Choose the baseline before inspecting the final evaluation period.
- 02
Pair the rows
Score model and reference on exactly the same settled matches and exclusions.
- 03
Keep the timeline
Estimate league priors or market probabilities only from information available at forecast time.
- 04
Publish the loss
Show negative and inconclusive comparisons instead of reporting only favourable splits.
Fast credibility screen
Treat these accuracy claims as unverified
- “99% accurate” without correct and settled counts.A rate without its denominator has no visible sampling uncertainty.
- Mixed markets inside one percentage.Double chance, 1X2, BTTS and correct-score hits answer different questions.
- Screenshots instead of a complete pre-match ledger.Selected images cannot establish timing or preserve every miss.
- No model version or evaluation cutoff.Readers cannot separate a current system from a retrospectively tuned backtest.
- No calibration or relevant baseline.A top-pick percentage alone cannot show whether probability estimates add information.