What is a good Brier score for football?

For three-way match results, around 0.57–0.62 is typical for strong forecasts (lower is better); guessing a third for every outcome scores about 0.67. Context matters: compare forecasters on the same matches.

A measure that penalises confident wrong predictions heavily: it is the average of −ln(probability given to what actually happened). Lower is better.

Where can I see your full record?

On our accuracy page, which grades every prediction after full time, with hit rates, calibration and Brier scores by market and a downloadable CSV.

How Accurate are football predictions?

By ProbaPredict Data Desk · Last reviewed Fri, 2 Oct 2026

Good football predictions are well calibrated, not always right: even the best models pick the winner only about half the time because draws and upsets are common. Accuracy is measured with the Brier score and log loss. In our back-test, margin-free closing odds are slightly more accurate than our model, and we publish that comparison.

Why can't predictions be right most of the time?

Football is low-scoring, so one deflection or penalty decides many matches. Draws happen about a quarter of the time. Even when a model is excellent, the favourite often has only a 45–55% chance. A forecaster who "picks the winner" in half of matches can be doing very well. That's why we judge probabilities, not hit rates alone.

What is the Brier score?

The Brier score is the average squared difference between the probability you gave and what happened (1 if it happened, 0 if not). For a three-way market, add the squared errors for home, draw and away. Example: you gave home 50%, draw 30%, away 20% and the home team won. Brier = (0.5 − 1)² + (0.3 − 0)² + (0.2 − 0)² = 0.25 + 0.09 + 0.04 = 0.38. Lower is better; 0 is perfect. Giving every outcome one third scores 0.667 every time.

What is log loss?

Log loss is the average of −ln(probability given to the actual outcome). In the example above, −ln(0.50) = 0.693. Had the away team won, it would be −ln(0.20) = 1.609. Log loss punishes confident mistakes much more than the Brier score, so a forecaster that sometimes says 2% for things that happen gets hammered.

Our live record and our back-test

Every prediction we publish is graded after full time and stays on the record. Separately, we ran a back-test: we predicted 18,807 past matches week by week, refitting the model each Monday on data available before that week, and compared it with margin-free closing odds.

Our live record (graded after every match)

Market	Hit rate	Brier score
1X2 match result	–	–
Over/Under 1.5 goals	–	–
Over/Under 2.5 goals	–	–
Over/Under 3.5 goals	–	–
All markets	–	–

Source: ProbaPredict graded predictions (0 graded picks), updated Fri, 2 Oct 2026. Each pick is our most likely outcome in that market; Brier score is per pick (0 is perfect).

Back-test: our model vs margin-free closing odds

Measure (lower is better)	Our model	Closing odds
1X2 Brier score	0.609	0.592
1X2 log loss	1.016	0.991
Over/Under 2.5 Brier score	0.246	0.239

Source: ProbaPredict analysis of 18,807 matches, updated Fri, 2 Oct 2026. Back-test, not live predictions: walk-forward out-of-sample predictions with the model refitted weekly on earlier data only. Closing odds made margin-free with the proportional method.

What does the back-test show?

Closing odds were slightly more accurate. On the match result, our Brier score was 0.609 against 0.592 for closing odds, and log loss 1.016 against 0.991. On Over/Under 2.5 it was 0.246 against 0.239. That's a gap of about 3% on the 1X2 Brier score. Closing odds aggregate a huge amount of money, team news and information our results-only model doesn't use, so this is the toughest benchmark there is. We think publishing the comparison is more useful than claiming to beat the market.

Worked example: comparing two forecasters

Two forecasters give a match: A says 60/25/15, B says 45/30/25. The away team wins. A's Brier = 0.36 + 0.0625 + 0.7225 = 1.145; B's = 0.2025 + 0.09 + 0.5625 = 0.855. B scores better on this match because it gave the actual result more weight. Over hundreds of matches, the forecaster whose probabilities track reality best ends up with the lower average.

How should you use accuracy figures?

Look for large samples, public grading of every pick (not just the winners), calibration tables, and a comparison with a strong benchmark. Be wary of anyone advertising only hit rates or "winning streaks". Our full record, including a CSV, is on the accuracy page.

Common mistakes when judging accuracy

Counting hit rates only. Picking 1.20 favourites gives a high hit rate and says little.
Small samples. A month of results is mostly noise.
Ignoring the benchmark. Compare with closing odds, not with guessing.
Mixing back-test and live results. We label which is which.

What is a realistic goal for a football model?

Beating random guessing is easy; beating closing odds is very hard. A realistic aim for a transparent, results-based model is to be well calibrated and to sit close to the market on Brier score and log loss, while explaining every number it publishes. Our back-test shows the gap to closing odds is a few percent on these measures and is smallest in some second-tier leagues.

We publish both our live record and the back-test because each answers a different question. The live record shows how the published predictions have actually done; the back-test shows how the method holds up over many more matches than we've published so far.

Frequently asked questions

What is a good Brier score for football?: For three-way match results, around 0.57–0.62 is typical for strong forecasts (lower is better); guessing a third for every outcome scores about 0.67. Context matters: compare forecasters on the same matches.
Are bookmakers more accurate than prediction models?: Closing odds are very hard to beat. In our back-test, margin-free closing odds are slightly more accurate than our model on both 1X2 and Over/Under 2.5.
What is log loss?: A measure that penalises confident wrong predictions heavily: it is the average of −ln(probability given to what actually happened). Lower is better.
Where can I see your full record?: On our accuracy page, which grades every prediction after full time, with hit rates, calibration and Brier scores by market and a downloadable CSV.

18+. Gambling can be addictive; only bet what you can afford to lose. Free help: BeGambleAware.org · GamStop.