Ask a prediction site how accurate it is and you'll usually get a number that ends in "95%". Ask a researcher and you'll get a much less exciting answer. This page gives you the less exciting answer, because it's the one that will actually save you money.
The baseline: guessing gets you 33%
A football match has three outcomes: home win, draw, away win. Pick one at random and you'll be right about a third of the time. Any accuracy figure only means something when you compare it against that baseline, and against the bookmaker's odds, which already encode a market full of informed opinions.
What experts and models actually achieve
Academic work on pundits puts expert accuracy at roughly 50% on Premier League match outcomes. Statistical and machine-learning models typically land between 50% and 60%, and the strongest published claims reach the 60s in favourable conditions. Nobody credible claims 95%.
The reason is the sport itself. Football is low-scoring, which means one deflection, one red card, or one goalkeeping mistake routinely flips a match that was analysed correctly. A model can call the probabilities right and still lose the result. That's not a flaw in the model. That's variance, and it never goes away.
Accuracy is the wrong question anyway
Two problems with a headline accuracy number.
First, it's easy to inflate. Predict "over 0.5 goals" in every match and your hit rate will be superb, and useless. A site quoting one big number without telling you which markets it counted is telling you nothing.
Second, hit rate isn't profit. Winning 60% of picks at odds of 1.50 loses money. Winning 40% at odds of 3.00 makes money. What matters is whether the predicted probability was better than the price, which is why serious services talk about calibration and ROI rather than raw accuracy.
What calibration means, in one paragraph
A calibrated predictor is one whose confidence matches reality. If you take every pick it labelled "70% confident" and check the results, about 70% of them should have won. Not 95%, not 50%. When a service shows you a calibration curve, you can see at a glance whether its confidence scores mean anything. When it doesn't show you one, you're taking its word for it.
We publish ours. Every WonlyAI pick carries a confidence score, every result is graded automatically, and the live performance page plots predicted probability against the actual hit rate for the last 30 days. Some weeks the curve looks good. Some weeks it doesn't. Both kinds are on the page, because an accuracy record you can't audit is just an advert.
How to sanity-check any accuracy claim
- Ask to see the losses. A graded archive with wins and losses is the minimum.
- Ask which markets the number covers. "Accuracy" across mixed markets is meaningless without a breakdown.
- Compare the claim against the research range. 50-60% on match outcomes is real skill. 90%+ across real odds is fantasy.
- Check whether confidence is stated per pick before kick-off, not claimed after full-time.
If a site fails all four, its number was invented for marketing. Plenty are.
The honest summary
Football predictions can be genuinely useful: a well-built model beats guessing by a wide margin and beats most casual analysis too. They cannot be certain, and any site that says otherwise is selling certainty it doesn't have. Treat predictions as probabilities, stake accordingly, and never bet money you can't afford to lose.
You can see today's calibrated picks, with the confidence and reasoning attached, on the predictions page. Read next: how AI football predictions actually work and how to judge a prediction site.