How accurate are football prediction models, really?

It is the first thing anyone sensible asks, and the thing almost nobody answers straight: how accurate is this, really? The honest reply is not a number you will love. A good football model is a bit better than a coin and nowhere near a crystal ball, and the people who tell you otherwise are usually the ones with a tip to sell.

Accuracy is not the same as being right

Start with what “accurate” even means, because the intuitive version is a trap. You might think a good model is one that picks the winner most of the time. But favourites win a lot, so a lazy rule, “always back the stronger side”, already looks impressive on a scoreboard while telling you nothing you did not know.

A forecast deals in probabilities, so it has to be judged as a probability. The right question is calibration: when the model says 60%, does that thing actually happen about 60 times in 100? A model can be “wrong” on a Saturday, its 55% pick loses, and still be perfectly accurate, if over a season the 55% calls come in around 55% of the time. One result tells you almost nothing. The pattern across hundreds tells you everything.

The ceiling is lower than you think

Football is a low-scoring, high-variance game. A single deflection, a soft penalty, a red card, and the “right” team loses. That variance is not noise the model failed to remove; it is the sport. It puts a hard ceiling on how accurate any model can be, no matter how clever or how much data it eats.

On three-way results (home, draw, away), the best public models in the world land the outcome roughly half the time. That sounds mediocre until you remember that about a quarter of matches are draws, the hardest thing in football to call, and that the gap between the top models and a decent amateur guess is small. The edge is real but thin, and it only shows up in the aggregate.

How to spot an honest model

Because the real accuracy is modest, the tell of a trustworthy model is not a big claim, it is honesty about the limits. Three things to look for, and they are the three this site is built around:

It keeps its misses. A model that quietly deletes the calls it got wrong is not a model, it is marketing. Accuracy can only be checked against a record that includes the failures.

It shows its working. If you cannot see where the number came from, the inputs, the method, the sources, you cannot judge whether it is sound or just a confident-sounding guess.

It talks in probabilities, never certainties.“A 62% chance” is a claim you can grade. “Spain will win” is a claim that dodges accountability, right when it matters.

Our own record, in the open

So how does the model here do? Honestly, it is far too early to say, and that is the whole point. Every pre-match call is timestamped and kept next to the result, the hits and the misses, and the sample is tiny. Here is the record so far, exactly as it stands:

The uncomfortable truth is that if a football model were as accurate as its loudest promoters imply, the people building it would be quietly rich, not selling you a newsletter. Modest, honest, and calibrated is not a weak answer. It is the only truthful one, and it is worth more than certainty that cannot survive contact with a season.