A mathematical football prediction does not tell you who will win. It tells you how often a team should win a match like this one if it were played many times. That difference is the whole point. A prediction of 65% for Arsenal is not a promise that Arsenal win; it says that in 35 of every 100 versions of that match they don't.
This guide explains where those numbers come from, how to read them next to the bookmaker's price, and when to trust them less. If you want to see them first, every upcoming match is on our mathematical predictions board.
Updated 29 October 2026 with a worked example from 24 October.
What makes a prediction "mathematical"
A tipster's prediction is an opinion. A mathematical prediction is the output of a model: a fixed method that takes the same inputs for every match and turns them into probabilities. Give it the same data twice and it gives the same answer twice.
That has two consequences. It can be checked, because you can compare thousands of past predictions with what happened. And it has no feelings about the teams, which is a strength (no bias towards big clubs) and a weakness (it cannot see a manager's resignation or a striker's injury unless that shows up in the data).
The core idea: goals behave like a Poisson process
Most football models start from one observation. Goals are rare, roughly independent events. A team gets a handful of chances in 90 minutes and converts a few of them. Counts of rare, independent events follow a known shape, the Poisson distribution, which needs only one number: the average.
If a team is expected to score 1.5 goals in a match, the Poisson distribution says how often it scores exactly 0, 1, 2, 3 or more:
| Goals scored | Chance (expected goals 1.5) |
|---|---|
| 0 | 22% |
| 1 | 33% |
| 2 | 25% |
| 3 | 13% |
| 4 or more | 7% |
So the whole problem becomes: what is each team's expected number of goals in this match?
Step 1: rate each team's attack and defence
The model looks at each side's recent matches and asks two questions. How many goals does this team score compared with an average team? How many does it concede compared with an average team? A side that scores 30% more than average has an attack rating of 1.3; one that concedes 20% fewer has a defence rating of 0.8.
Three adjustments make those ratings sensible:
- Recent matches count more. Last month tells you more about a team than last season, so older games are weighted down.
- Small samples are pulled towards the average. A team that scored four in its only two games is not really a four-goals-a-game side. Ratings start near average and move away only as the evidence builds up.
- Home and away are treated differently. Home teams score more in almost every league, so the model allows for venue.
Shots, chance quality and expected goals (xG) help here too, because they are a steadier signal of how good a team is than goals alone.
Step 2: turn the ratings into expected goals for this match
Expected goals for the home side is roughly: the home team's attack, times the away team's defence, times the league's average home goals. The away side is the mirror image. Two strong attacks against two weak defences give high numbers on both sides; a strong defence pulls the other team's number down.
Step 3: build the scoreline grid
With an expected-goals figure for each side, the model works out the chance of every scoreline, 0–0, 1–0, 0–1, 1–1 and so on, by combining the two Poisson distributions. That grid of scoreline chances is the engine behind everything else.
Good models also correct a known weakness here. Plain Poisson slightly underrates draws like 0–0 and 1–1, because low-scoring games are not quite independent: a team leading 1–0 often settles for it. A small adjustment for those scores (the best-known version is the Dixon-Coles correction) fixes most of it.
Step 4: read every market off the same grid
Once the grid exists, every common market is just a sum of cells:
- Home win: every scoreline where the home side scores more.
- Over 2.5 goals: every scoreline with three or more goals.
- Both teams to score: every scoreline where neither side has zero.
- Correct score: one cell on its own.
Because all of them come from the same grid, they agree with each other. A model that says 70% home win and 20% over 2.5 is telling you it expects the home side to win something like 1–0 or 2–0, and you can see exactly that in the grid.
A real example: Arsenal 2–0 Everton, 24 October 2026
Before kick-off the model's chances were:
| Outcome | Model | Bookmaker (bet365, margin removed) |
|---|---|---|
| Arsenal win | 64.7% | about 69% (odds 1.36) |
| Draw | 21.4% | about 19% (odds 5.00) |
| Everton win | 14.0% | about 12% (odds 8.00) |
| Over 2.5 goals | 49.5% | |
| Both teams to score | 45.9% |
The most likely single scores were 1–0 (13.8%), 2–0 (12.7%) and 2–1 (10.2%). Arsenal won 2–0.
Two things are worth noticing. First, model and market were close, within five points on every outcome, which is what you want to see before trusting either. Second, even in a match with a clear favourite, the most likely exact score had less than a one-in-seven chance. That is normal, and it is why correct-score betting is so hard.
How to read a prediction next to the odds
A probability on its own does not tell you whether a bet is worth making. You need the price.
- Convert the odds to a probability. Decimal odds of 2.00 mean 50%; 1.36 means about 73.5%. Our odds converter does this for you.
- Take out the margin. The bookmaker's three 1X2 probabilities add up to more than 100%; the extra is the bookmaker's margin. Scale the three down so they total 100%.
- Compare. If the model's chance is clearly above the margin-free price, that is a possible value bet. If it is below, the price is too short for the model.
On the FootyPulse board, a pick the model rates at 55% or more is labelled Strong, anything below is a Lean, and a value angle is shown only where the model's chance is at least three points above the margin-free price. More on that idea in our guide to value betting.
Why a big gap usually means the model is wrong
It is tempting to look for the matches where the model and the bookmaker disagree most. Those are usually the matches where the model is missing something.
A bookmaker's price on a popular match is shaped by a lot of money and by people whose job is to correct mistakes. A goals model sees only historical numbers. It cannot know that a key player was ruled out this morning, that a team is resting players before a cup final, or that a newly promoted side spent heavily in the summer. When model and market are 15 or 20 points apart, the market is usually the one that knows.
That is why the FootyPulse predictions board, by default, only shows matches where there is a bookmaker price to check against and where the model is within 15 percentage points of the market on every outcome. The rest are still there behind "Show all matches", marked with a caution.
Where mathematical predictions work best, and where they don't
Strongest:
- Leagues with plenty of history and stable squads, where ratings settle quickly.
- Markets driven mainly by goal rates, such as over/under and both teams to score.
- Matches without unusual context: no derby, no dead rubber, no rotation.
Weakest:
- The first weeks of a season, before the new campaign's data has built up.
- Cup ties against lower-league sides and early-season continental matches, where the model has little head-to-head evidence.
- Anything decided by team news the data cannot see.
Two further checks help: our home and away statistics and the form guide show whether a team's recent numbers support what the model says.
Common mistakes with mathematical predictions
- Treating a probability as a tip. 60% loses four times in ten. Over a season that is a lot of losing bets on "good" predictions.
- Backing the model's pick at any price. A 70% favourite at odds of 1.30 is a bad bet even if the model is right about the 70%.
- Chasing the biggest edges. As above, the biggest gaps are usually model blind spots, not bargains.
- Judging a model on one weekend. A probability can only be judged over hundreds of matches. A run of five losses tells you very little.
Responsible gambling
Predictions describe chances, and chances include losing. Set a budget before you bet, never chase losses, and treat any stake as money you can afford to lose. If betting stops being fun, help is available from BeGambleAware.