How Accurate Are Football Predictions: A Practical Audit of Any Match Forecast
A forecast's accuracy can be audited with a graded public record, calibration, a Brier score against a baseline and enough matches; here is how, with Kickwise's record as the worked example.
Before you rely on a forecast in match preparation, in a board paper or in a fan product, you need a way to answer the question of how accurate football predictions really are. A single headline hit rate will not do it. The answer comes from an audit, and the audit has five parts: a public record that cannot be edited after the fact, calibration, confidence bands, a score measured against a plain baseline, and a sample large enough to mean something.
Start with the record itself. A forecast is only checkable if the call was fixed before kickoff and graded afterwards on a stated rule. Ask which call is kept when a forecast changes (the first one, the best one, or the last one before kickoff) and what result it is graded against. Ninety minutes, or extra time as well? Penalties as a draw or as a win? A provider that will not tell you these things is not offering a record, only a story. Also check that misses are visible alongside hits, and that the record covers every match the provider claims to cover, not a selection.
Calibration is the second test, and the one executives most often skip. A calibrated probability means that when a forecast says 70%, the outcome happens about seven times in ten over many matches, and a miss three times in ten is part of the deal, not a failure. A single 70% call that loses tells you almost nothing. What you want to see is whether the 70% calls, taken together, land near 70%, and whether the 40% calls land near 40%. A forecaster who is always confident and often right in a small run may simply be lucky, while one whose stated numbers match reality is giving you something you can plan around.
That leads to confidence bands. Overall accuracy blends easy matches with near coin flips, so read the record by band. High-confidence calls are the ones you might act on, in rotation or in how you frame a fixture for a board, so judge them separately and compare them with the obvious shortcut. If the home side wins 60% of the time in the same matches, then always backing the home team is the real benchmark, and a forecast has to beat it to justify its existence.
The Brier score is the fourth tool. It measures the gap between the probabilities you published and what happened, penalising confident misses heavily, and lower is better. On its own the number is hard to read, so always set it against a baseline, such as giving every match the sample's own home, draw and away rates. A model that cannot beat that baseline is adding noise. Any serious answer to whether AI can accurately predict football outcomes should be framed this way: not as a yes or no, but as a score against a dull alternative.
Kickwise's own record shows how this works, including the awkward parts. Its public record keeps the last call made before kickoff and grades it on the result after 90 minutes plus any extra time, with a penalty shootout counted as a draw. On calls made at 65% confidence or higher over the last 90 days, it stands at 43 right out of 53 (81%), in matches where always backing the home team would have been right 60% of the time. Across every graded call at every confidence level in the last 30 days, it is 68 right out of 111 (61%), with a Brier score of 0.56 against 0.654 for the baseline; the home side won 40% of those matches. The last 7 days read 23 right out of 37 (62%), with 0.575 against 0.643.
Now the limits, which matter as much as the figures. The draw is the hardest outcome to call. Friendlies and matches with thin data are close to coin flips, and a forecast should say so rather than dress them up. And a 70% call still misses about three times in ten. Then there is sample size: 37 matches in a week and 53 high-confidence calls in a quarter are enough to see a direction, not to settle a verdict. A 7-day window can swing on a handful of results, which is why the longer windows deserve more weight and why none of these should be read as a promise.
Put together, the audit is simple enough to hand to a head of data or a technology lead evaluating a vendor or building in-house. Demand the call timing and grading rule, read calibration by band, compare the Brier score with a stated baseline, and refuse conclusions from small samples. Used that way, a forecast is a useful input to pre match analysis and to opposition analysis reports, a calibrated view of the odds alongside scouting and tactics. It does not replace them, and it does not tell you what will happen on Saturday.