How Accurate Are Betting Odds? Calibration and the Brier Score

Betting odds have no single accuracy percentage. You can measure how often the favourite wins, whether implied probabilities match observed frequencies, or how well the full probability forecasts score. These answer different questions. Results also depend on the league, market, period and time at which the odds were recorded.

A team given a 60% chance can lose without invalidating that forecast. Across many comparable 60% forecasts, however, an observed win rate close to 60% would support calibration. This guide explains that distinction and shows how to calculate the Brier score.

See the football market study →

For 1X2 hit rate, the pick is often the highest-probability outcome among home, draw and away. The tie rule must be specified. Predicting 51% and 90% for the same winning pick earns the same hit, but a probability score distinguishes them.

For decimal odds d, the raw implied probability is 1/d. Across a bookmaker's complete market, these values usually sum to more than 100%. That excess is the overround; it does not tell you the true probability of each outcome.

A common adjustment divides each raw probability by their sum. For 1X2, include home, draw and away from the same bookmaker and snapshot. This produces a distribution summing to 100%, but the result still depends on the margin-removal assumption. See how to convert odds into probabilities. Normalisation is not evidence that the resulting forecasts are calibrated.

Suppose a group of home-win forecasts averages 60%. If home wins occur in about 60% of that group, the frequencies align. An observed 45% rate suggests the forecasts overestimated home-win probability; 75% suggests underestimation. Whether either gap is convincing depends on the number of matches and sampling uncertainty.

A forecast that always reports a stable event's base rate can be calibrated while making no distinction between fixtures. Useful evaluation therefore considers both calibration and the ability to separate different levels of risk.

For a binary event such as “home wins,” use probabilities between 0 and 1 and encode the result as 1 if the event happened, otherwise 0. Average the squared errors:

Brier = (1/N) × Σ(p − y)2

The binary score ranges from 0 to 1; lower is better. Zero means every forecast in the evaluated sample assigned probability 1 to what happened. A confident error receives a large penalty.

Brier score is not calibration alone. It reflects calibration, resolution and the event uncertainty in the sample. A lower score need not mean better calibration; the scikit-learn calibration documentation explains this distinction.

Rounded to three decimals, the score is 0.136. Using a 50% threshold for home win versus no home win, all four binary classifications are correct. The probabilities still receive non-zero errors because none predicted the realised event with certainty.

These four examples explain the arithmetic. They are too few to establish calibration or a repeatable advantage over another forecaster.

There is no universal cutoff. Always forecasting 0.50 gives a binary Brier score of 0.25, regardless of the results. But consider an event whose stable underlying frequency is 80%: always predicting 0.80 has an expected score of 0.8 × 0.2 = 0.16, despite using no fixture-specific information.

Compare forecasts on the same matches against a relevant baseline, such as a constant probability estimated from earlier data and a margin-adjusted market forecast. Keep the evaluation period separate from data used to fit or select the model.

For three-way 1X2, sum the squared errors for home, draw and away on each match, then average across matches. This unscaled convention ranges from 0 to 2. Some tools divide by two; check the convention before comparing scores. The binary 0.25 reference cannot be transferred directly to an unscaled 1X2 score.

Use the Football Probability Evaluation Calculator for full 1X2 vectors, Brier score, log loss and reliability tables.

A reliability diagram groups predictions into probability bins. Its horizontal axis shows the mean forecast probability in each bin; the vertical axis shows the observed event frequency. On these axes, points above the diagonal indicate underestimated event probabilities; points below indicate overestimation.

Show the number of forecasts per bin and uncertainty intervals. A small bin can move substantially after a few results. Check home, draw and away separately, and inspect relevant leagues or periods: pooled averages can hide opposing errors. The diagram diagnoses calibration; Brier provides a complementary overall score.

1. Define the event and timing. Use the same settlement period and matched odds snapshots. Keep opening and closing prices distinct.
2. Save forecasts before results. Record full probabilities, odds source and margin-removal method. Compare models with information available at the stated forecast time.
3. Evaluate the same held-out matches. Report sample size, baseline and probability scores alongside calibration diagnostics. Avoid selecting only successful leagues or seasons after seeing results.
4. Quantify uncertainty. A small score difference may be sampling noise. Check stability on later data before treating a historical improvement as repeatable.

A probability score improvement does not itself establish profitable betting. Returns also depend on the prices actually available and the decisions made.

The published market calibration study covers 19,381 matches across five leagues. It compares implied and observed frequencies in historical groups. Its displayed gaps describe that sample; selected subgroup differences do not establish a persistent market bias.

The World Cup 2026 tracker is a completed 104-match archive. It records results against pre-match score maps and includes observed versus expected summaries. One close aggregate comparison does not prove calibration across all probability levels. The score probability map guide explains the separate meanings of expected coverage and observed hit rate.

The modelling methodology explains how odds become a score distribution. Matching input market probabilities is a fitting check; testing predictions against later outcomes is an evaluation task.