How Accuratswe Are AI Football Predictions?

Most football prediction sites advertise an accuracy figure, often 90% or more. The honest number is a lot lower, and it depends on what is being predicted. In the 2017 Soccer Prediction Challenge, run alongside a special issue of the journal Machine Learning, entrants trained on more than 200,000 past results and the best of them got about 52% of the 206 test matches right. Later work on the same problem found the bookmakers' own odds forecasting about as well as the best models.So a site that has genuinely built a good model is doing well to get the result right a little over half the time. Anything advertised far above that is describing a different market, a different sample, or nothing at all.

The market decides the difficulty

Whether a match produces more than two and a half goals is close to a weighted coin flip, driven by how freely the two teams tend to score. The match result is harder, because a draw is always on the table and no model gets to ignore it. The exact final score is a different problem again. More than a dozen scorelines have a realistic chance in an ordinary fixture, and the single likeliest of them is wrong about six times out of seven.

A site quoting one accuracy figure across all of those has averaged three unrelated problems into a number that describes none of them. The first question about any accuracy claim is therefore which market it covers. An 80% record on over and under markets is plausible. An 80% record on exact scores is not, and a site claiming it should be able to show the rows.

What the models produce

The serious work in this field is probabilistic. A model estimates how many goals each side is likely to score and turns that into a distribution across every possible result. The best known approach is the double Poisson model, which treats each team's goals as a Poisson process driven by attacking and defensive strength, and it holds up in competition. In 2022, Penn and Donnelly published in PLOS ONE an analysis of a double Poisson model whose predictions won the Royal Statistical Society's Euro 2020 prediction competition. A relatively simple statistical model, carefully fitted, competes with anything in the field.

That work also shows what an honest claim looks like. A model that puts a home win at 60% expects to be wrong four times in ten, so four losses in one week is ordinary variance and says nothing about whether the model is broken. Any site presenting its output as a verdict rather than a probability has already told you something about how it will report its results.

Five checks you can run yourself

Most accuracy claims fall apart under five questions, and none of them takes long to answer.

  1. Which market does the number cover? An accuracy figure with no named market is marketing. Look for the market, the size of the sample, and the date range in the same place.
  2. Is there a dated public archive? The claim should sit on a list of individual past predictions with dates attached, and not only on a badge in the page header. If the record exists as a percentage and nothing else, there is nothing to audit.
  3. Does each row show the prediction as it stood before kickoff? A forecast stored as editable text can be adjusted after the match, and a results page assembled afterward proves nothing about what was published beforehand. What matters is whether the archive shows the forecast frozen at publication, with the inputs behind it.
  4. Does the site say what counts and how it settles? Two quiet rules move every percentage: whether the record covers every fixture on the site or only a shortlist, and whether a pick settles on the 90-minute score or includes extra time. A site that states neither has left itself room to move.
  5. Does the site say where its model is weak? Forecasting how many goals a match produces is a different skill from forecasting who wins it, and most models are far better at one than the other. A site that names which is which is telling you where its number can be trusted. A site claiming the same accuracy everywhere has not looked.

Some of the bigger sites do parts of this well. Forebet, one of the biggest, colors every archived prediction green or red against the result. Fewer sites get through the last three checks. Superior Tips publishes its graded archive with the frozen pre-match snapshot sitting behind every row, states that a forecast is fixed when it is published and never recalculated from the result, and says plainly that what enters the record is its shortlisted selections rather than every fixture on the site, settled on the 90-minute score.

It is also unusually direct on the fifth check. The site says its model is strongest on goal totals and tight, low-scoring games and weakest at picking the winner, and its daily selections are drawn only from the markets and the price range where its own graded record holds up. The probability printed beside each pick has been fitted to that record rather than taken raw from the model, so it is built to promise slightly less than lands. When the model and the market disagree sharply, the site treats the disagreement as its own error and refuses the pick. Whether you agree with the picks is a separate question. The point is that all five answers are on the page, which is what lets a reader form a view.

What a full record still cannot do

Even an auditable archive has limits worth understanding before anyone leans on it. Football is low scoring and high variance, so small samples swing a long way, and a month of results in either direction is mostly noise. Reading a hot two weeks as a trend is how people talk themselves into bad decisions.

Hit rate on its own is incomplete too. A service picking heavy favorites and a service picking outsiders can post the same percentage with completely different consequences for anyone following along, so the price attached to each pick carries as much information as whether it landed.

Every price also carries the bookmaker's margin, and a pick that lands exactly as often as its price implies still loses that margin over time. No model removes it. A site that prints its own probability beside the price at least lets you see how much of the price is margin, which is more than any banner will tell you.

The same discipline matters more once a match is under way. A goalless first half makes the over look cheap, and the live price moves to say so. Superior Tips checked that situation across its own match history and found the bet loses more often than it wins, which is why it reports a live price as a price and never as value. A number that has moved is not the same thing as a number that is wrong.

The arithmetic underneath most of these models is sound. What goes wrong is the banner someone hangs above it.

Test it in five minutes

Pick any prediction site and open its archive. Choose a week that has already finished, count the predictions, check the dates, and work out whether the page shows what was published before kickoff or only what happened afterward. Then ask whether the accuracy figure on the homepage could have been calculated from the rows you counted, and whether the site anywhere admits what its model gets wrong.

Superior Tips is one site where that exercise can be done end to end. A surprising number of others do not allow it, and finding out which is which is the quickest way to tell a model from a marketing department.

Enjoyed this article? Share it!

Zalven Draxen
Written By

Zalven Draxen

2070 Articles

Zalven is a U.S.-based SEO strategist and digital marketing expert known for helping businesses grow through search optimization, online visibility, and smart content strategies. With deep experience in technical SEO and local search, he simplifies complex marketing concepts into clear, actionable insights for brands of all sizes.

1 thought on “How Accuratswe Are AI Football Predictions?”

Comments are closed.