Model

Why is football so hard to predict?

Published 2026-09-01

Football forecasting has a reputation for being bad at its job, and the reputation is partly deserved. But the usual explanations — that the sport is too emotional, too tactical, too human to model — get the cause wrong. Basketball is emotional and tactical too, and it is far more predictable. The difference is arithmetic.

Football is hard to predict because it is low-scoring, and every consequence follows from that one property.

One goal decides almost everything

In a sport where a typical match produces two or three goals, a single goal is an enormous fraction of the total. Reverse one goal in a football match and you have very often reversed the result. Reverse one basket in a basketball game and you have changed the score from 104-98 to 104-96.

That is the whole story in one sentence. The smaller the number of scoring events, the larger the share of the outcome that any individual event controls — and individual events in football are extremely noisy. A deflection, an offside call measured in centimetres, a goalkeeper’s fingertip, a post. None of these are things a model can know in advance, and each of them routinely decides matches.

A better team is genuinely more likely to win. It is just that “more likely” tops out much lower here than in higher-scoring sports, because the sport does not give superiority enough events to express itself.

The draw is the fingerprint of this

Watch what happens across our own data. These are the nine leagues we model, ordered by how many goals they produce:

The four lowest-scoring competitions all draw between 26.1% and 27.6% of the time. The three highest-scoring sit between 23.9% and 24.9%. The pattern is real, and it is not perfect — the Bundesliga scores most and still draws more often than the Premier League — but the direction is clear and the mechanism is obvious once stated. Fewer goals means more matches where neither side manages to get ahead and stay there.

That matters for prediction because the draw is the outcome no model can ever be confident about. It is rarely the most likely of the three results, so it is rarely anyone’s pick, and yet it happens in a quarter of all matches. Every drawn match is a match where the favourite did not win, and a league that draws 27% of the time has a hard ceiling on how often anyone’s favourite can come in.

The consequence, in the market’s own numbers

The clearest way to see the ceiling is to look not at models but at the betting market, which has more money and better information than any public forecaster.

Across 16,628 matches carrying closing odds in our database, the market’s own favourite won 53.1% of the time. That is the aggregated judgement of professional traders and everyone willing to bet against them, and it is right about half the time — because the sport does not permit better.

In the Championship, the same market’s favourite comes in 47.4% of the time. A competition of 24 closely-matched teams playing 46 rounds produces so little separation that the best available estimate of who will win is wrong more often than it is right.

What this does and doesn’t mean

It does not mean predictions are worthless. A model that reliably distinguishes a 70% match from a 45% match is doing real work even though it will be “wrong” often in both categories. The value is in the probability being honest, not in the pick being right.

It does mean accuracy claims should be read against the ceiling. When a site advertises 85% or 90%, the question is not whether they are unusually good. It is which subset of fixtures produced the figure, because no honest process gets there across a full slate of matches. We wrote out where the ceiling actually sits and how to test any claim against it.

It means the right metric is not accuracy at all. If being right about half the time is the natural state of a good football forecast, then hit rate cannot distinguish good from bad. What can is whether the stated probabilities match reality over many matches — which is what calibration measures and what the Brier score scores.

And it means large samples are not optional. In a sport this noisy, dozens of predictions tell you essentially nothing about whether a forecaster is skilled. Hundreds begin to.

What we do about it

Nothing that makes football easier. Our model produces probabilities rather than confident picks, publishes them before kick-off with the model version attached, and grades every one of them — including the many that miss — on a public track record. The methodology page shows the model’s own backtest against the closing-odds market, gap included.

A forecaster who admits the ceiling is not being modest; they are describing the sport accurately. Treat football probabilities as probabilities, and stake only what you can afford to lose.