15 Jul 2026
How to Measure Football Prediction Accuracy: Sample Size, Hit Rate, Yield and Calibration
A football prediction record needs more than a list of correct calls. Use a fixed sample, clear probability forecasts, closing odds and complete records to judge whether results are meaningful.
Start with a defined prediction and a large enough sample
Measure one prediction type at a time: match winner, both teams to score, over 2.5 goals, correct score or another fixed market. Record every prediction before the match starts, using the same selection rules. Small samples mislead. A run of 15 correct picks can happen by chance, especially with favourites. Results become more useful after hundreds of comparable forecasts, though the required sample depends on the event type and the claimed advantage. Separate leagues, seasons and market types when their conditions differ.
Use hit rate correctly
Hit rate is the percentage of predictions that were correct: correct predictions divided by total predictions. If 58 of 100 match-winner calls win, the hit rate is 58%. That figure has no meaning without context. Picking the home favourite in every match can produce a high hit rate, while offering little information beyond public expectations. Compare the hit rate with a baseline, such as the historical rate for the same type of selection or the implied probability from the available odds. Correct-score predictions naturally have far lower hit rates than over/under calls.
Calculate yield only from complete stake and price records
Yield measures net return against total amount staked: net profit or loss divided by total stake, expressed as a percentage. For example, a record with 1,000 units staked and a 40-unit net profit has a 4% yield. Yield can describe the financial result of a forecasting method, but it is sensitive to the odds used, stake sizing, settlement rules and omitted losses. A claimed yield is not credible if it excludes losing selections, uses prices that were unavailable, changes stakes after results, or mixes single predictions with accumulators. Flat-stake results are easier to audit because each prediction carries the same weight.
Check calibration when forecasts include probabilities
Calibration asks whether stated probabilities match real-world frequencies. A model that labels 100 home wins as 60% likely should see about 60 of those home teams win over a sufficiently large sample. Group forecasts into probability bands, such as 40-49%, 50-59% and 60-69%, then compare the average stated probability with the actual win rate in each group. Calibration differs from hit rate. A model can correctly identify many favourites yet be poorly calibrated if it calls 80% chances that win only 65% of the time. Brier score and log loss add useful detail because they penalise overconfident wrong forecasts.
Use closing odds as a market benchmark, not as proof
Closing odds are the final widely available prices before kick-off. Their implied probabilities provide a strong benchmark because they incorporate team news, injuries, line-ups and broad market opinion. Compare a prediction's estimated probability with the closing implied probability after accounting for bookmaker margin. If forecasts repeatedly rate outcomes higher than the closing market and those estimates remain calibrated, that is stronger evidence than a short winning run. Be careful with the phrase closing-line value: getting a better price than the close does not guarantee a correct prediction or future profit, and a few price movements prove little.
Keep an auditable record and reject misleading claims
A useful log includes date, competition, fixture, market, prediction, stated probability, odds available when published, closing odds, stake method if yield is reported, result and settlement. Preserve timestamps and do not delete losing picks. Treat claims with caution when they show only wins, quote profit without total stakes, use vague phrases such as “high confidence,” cite a short streak, change the rules mid-season, or compare their hit rate with no baseline. The clearest report gives the full sample size, hit rate, yield where relevant, calibration results, closing-odds comparison and the exact period covered.
Analysis: pksport · our methodology
Analysis based on public data and market signals. For analysis only — not betting advice.