How to Judge Football Predictions

Every tipster and prediction site claims to be accurate. Very few show the evidence in a form you can check. This guide covers the questions worth asking of any source of football predictions, ours included: how many predictions the record is based on, whether a high hit rate actually means anything, and whether the stated probabilities hold up.

1. Is the whole record visible?

The simplest test comes first. Predictions should be published before kick-off and stay up afterwards, hits and misses alike. A page of winning tips with the losers quietly deleted tells you nothing. Watch for:

• tips that appear only after the match, or have no visible date;
• results pages that list wins but not losses;
• claims of “guaranteed” or “fixed” matches, which are always false and often a scam;
• paid “VIP” tips sold on the strength of a short winning streak.

On GoalStatsHub every prediction stays on its match page after the final whistle. Tips and scores that landed turn green and misses are struck through. Estimates from an outside data provider are marked ext and are never counted as ours.

2. How many predictions is the record based on?

Short records are mostly luck. Suppose a source hits 60% of 50 tips. The statistical margin of error on that figure is about ±14 points, so its true rate could plausibly be anywhere from about 46% to 74%. Over 500 tips the same 60% is accurate to roughly ±4 points. As a rough guide:

• under 100 predictions: too few to separate skill from chance;
• a few hundred: enough to spot a source that is clearly bad;
• thousands, across many leagues and seasons: enough to judge a model.

3. A high hit rate isn’t the same as good predictions

Hit rate means little without the prices. Picking heavy favourites at odds of 1.20 and landing 80% of them sounds impressive, but 0.8 × 1.20 = 0.96: you lose 4% of everything staked. Landing 40% of picks at odds of 3.00 sounds poor, yet 0.4 × 3.00 = 1.20 would be a 20% profit. Any record that quotes a hit rate without the odds, or the probabilities, of the picks is incomplete.

Our guide to odds and probability shows how to convert between the two.

4. Calibration: do the percentages mean what they say?

For a source that publishes probabilities, the key test is calibration. Collect every outcome it rated at around 60%. If roughly 60% of them happened, the model is well calibrated at that level. Do the same for 20%, 40%, 80% and so on.

• If outcomes rated 70% happen only 55% of the time, the model is overconfident.
• If outcomes rated 30% happen 40% of the time, it is underrating them.

Calibration is what makes a probability useful. A well-calibrated 60% lets you compare the figure with a bookmaker’s price; a badly calibrated one doesn’t, however often its favourites happen to win.

5. Scoring rules: one number for the whole record

Statisticians score probability forecasts with rules that reward being confident and right and punish being confident and wrong. The most common for football is the Brier score: for each prediction, take the squared difference between the probability and what happened (1 if it happened, 0 if not), then average.

• Saying 70% for something that happens scores (1 − 0.7)² = 0.09.
• Saying 70% for something that doesn’t happen scores 0.7² = 0.49.
• Always saying 50% scores 0.25 every time.

Lower is better, and a useful model must beat simple baselines such as “always 50%” or the long-run frequency of each outcome. Log loss is a similar rule that punishes confident mistakes even more heavily.

6. Losing streaks are part of the deal

Even a perfect model loses often. Ten independent picks at 70% all land only about 3% of the time, and a run of four or five misses in a row will happen regularly over a season. A streak of losses is not by itself proof that a model is broken, just as a streak of wins is not proof that it works. Only a long, complete record answers that question.

How we apply this to ourselves

We keep every in-house prediction after the match so that the record can be judged in exactly these terms: hit rates by market and confidence level, and calibration across the probability range. A public accuracy report built on that record is in preparation. Until it is live, every match page shows what we predicted and how it turned out, and the methodology page explains how the numbers are produced.

18+. No prediction source can guarantee results. If you bet, set a limit you can afford to lose and read our responsible gambling page.