How Accurate Are Football Predictions, Really?

Search for football predictions and you will find a lot of confident numbers. Ninety per cent accuracy. Guaranteed winners. Verified tipsters with records that only ever seem to go one way.

Almost none of them show you the losing weeks.

So here is a different kind of article: what prediction accuracy actually looks like when you measure it honestly, what a good number is, and why anyone promising you much more than that is either measuring something very easy or not measuring at all.

Start with what guessing gets you

Every accuracy claim needs a baseline, otherwise the number means nothing.

A football match has three possible outcomes: home win, draw, away win. Pick one at random and you will be right about 33% of the time. That is the floor. Any model, tipster or pundit who cannot beat 33% is actively worse than a coin with three sides.

Exact scorelines are much harder. Most matches finish with a scoreline from a pool of maybe thirty realistic options, though a handful dominate. Guess at random and you land around 3%.

So those are the two numbers to hold in your head. 33% for the result. 3% for the exact score. Everything else should be judged against them.

What a good model actually achieves

Here are our numbers, taken from the same results page anyone can open, covering every prediction we have published:

• Match outcome: around 50%. Half of our predictions call the right result.
• Exact scoreline: around 12%. Roughly one in eight.

Put next to the baselines, that is a meaningful edge. Fifty per cent against a 33% floor means we are right about half again as often as guessing. Twelve per cent on exact scores is four times the random rate.

Put next to the marketing you will see elsewhere, it sounds modest. That is the point. A 50% result accuracy is roughly what a genuinely good football model achieves, and anyone quoting you 80% or 90% across all matches is not describing the same thing you are imagining.

Why half of them are wrong

This is the part worth sitting with, because it is not a flaw in the model. It is the sport.

Football is low-scoring, and low-scoring games are noisy. A single deflection, a marginal offside call, a goalkeeper having the afternoon of his life — any one of those flips a result that the underlying play did not deserve to flip. Over ninety minutes, the better team does not win nearly as often as people assume.

That noise is exactly why we build predictions from expected goals rather than from recent results. A team that has lost three on the bounce while creating the better chances in all three is not a team in crisis; it is a team due a correction. Results tell you what happened. xG tells you what was likely to happen, and that is a far better guide to the next match.

But even a perfect read on chance quality cannot tell you whether the shot goes in. No model can. Anyone who claims otherwise is selling something.

Where the accuracy is higher

Accuracy is not evenly spread, and this is where a model earns its keep.

Some matches are genuinely more predictable than others. When our model gives a home side a 65–75% chance of winning, those predictions have come in 74.6% of the time across 873 of them and eleven backtested seasons. That is the band we flag as Top Picks.

Two details about that number matter more than the number itself.

It only applies to home wins. At the same confidence level, our away predictions come in 59.5% of the time against 73.6% for home ones. Home advantage is partly made of things expected goals cannot see — the crowd, the travel, the way marginal decisions tend to fall. When a model makes an away side a heavy favourite on shot quality alone, it is usually missing something.

It is capped at 75%, not open-ended. This is the part most sites would quietly leave out. Above about 75% confidence our model becomes over-confident: those predictions land 67.8% of the time, not the 78% the model claims. So our most certain-looking calls deliberately do not get our top badge. The badge reflects how often that kind of prediction has actually landed, not how sure the model says it is.

A model that knows where it is unreliable is more useful than one that insists it is right everywhere.

How to judge anyone's accuracy claim

If you are weighing up any prediction service, including this one, four questions sort the serious from the rest.

Over how many predictions? A 90% hit rate over twenty matches is noise. You need hundreds before a rate means anything, and thousands before it means much.

Measured against what baseline? "70% accurate" is meaningless until you know what was being predicted. Seventy per cent on "will there be a goal in this match" is worse than guessing. Seventy per cent on exact scorelines would be extraordinary and is almost certainly false.

Are the losses published? This is the quickest test. A record with no bad weeks in it has been curated. Ours has two seasons in red on the backtest table because they fell short, and they stay there.

Can you check the working? Every figure on our results page comes out of a script in our own codebase, run against data anyone can obtain from Understat and football-data.co.uk. If a number cannot be reproduced, it cannot be verified.

What to do with a prediction

A prediction is a piece of information, not an instruction. Ours tells you what the chance quality in a fixture suggests, nothing more. We do not know about the injury that happened in training on Thursday, the manager resting players for a cup tie, or the fact that one side has nothing left to play for. The model reads shot quality and nothing else, and we would rather say that plainly than imply a completeness we do not have.

Used well, that information tells you where the data disagrees with the obvious narrative — which is most of the value in it. Used as a guarantee, it will let you down about half the time, exactly as the numbers above say it will.

You can see this week's predictions across all five leagues for free, scorelines and all. The percentages behind them, and the matches we are most confident about, are what a subscription covers.

Either way, the record stays public.