Correct Score and AI: Why It Is the Most Uncertain Forecast and How a Model Estimates It

Tablette affichant une analyse tactique et des probabilités de score

Can AI predict the correct score of a football match? It can put a probability on every possible score, not guess the right one. The correct score is the most detailed forecast there is, and therefore the most uncertain: in the worked example below, the single most likely score has only about a 12% chance of happening, which leaves close to 9 chances in 10 that it will be wrong. Here is why, and how a model built on expected goals (xG) and the Poisson distribution produces a full spread of scores.

Why the correct score is the least likely outcome

A correct score is one specific version of a result. A 2-1 is a home win, but only one of many: 1-0, 2-0, 3-1 and plenty of others are home wins too. Its probability is therefore always lower than that of the win it implies, which is itself lower than “win or draw”.

Football makes this worse: goals are rare, and a single goal at any moment is enough to change the final score. Where a 1X2 forecast splits the probability three ways, a score forecast spreads it across a long list of possible scorelines.

How a model estimates a spread of scores

1. Estimate each team's expected goals

The starting point is the average number of goals each team is expected to score in this particular match, written λ (lambda). Classic models build it from a team's attacking strength, the opponent's defensive strength and home advantage, the approach set out by M. J. Maher as early as 1982. More recent models often add expected goals (xG), which measure the quality of the chances created and conceded rather than goals alone. For more detail, read our guide to expected goals (xG), over/under 2.5 and both teams to score.

2. Turn an average into probabilities with the Poisson distribution

The Poisson distribution describes how often rare events occur over a fixed period, here a team's goals over 90 minutes. If λ is the expected average, the probability that the team scores exactly k goals is:

P(k) = e^(−λ) × λ^k / k!

where e ≈ 2.718 and k! is the factorial of k (3! = 3 × 2 × 1 = 6; by convention, 0! = 1).

3. Combine the two teams

If we assume each team scores independently of the other, the probability of a score such as 2-1 is the product of two numbers: the chance that team A scores 2 and the chance that team B scores 1. Doing this for every combination gives a grid of all scores. Adding up its cells gives back the win, draw and loss probabilities, or the probabilities for the total number of goals.

A worked example, step by step (invented figures)

Take an imaginary match: team A, at home, has 1.5 expected goals; team B has 1.1.

Team A (λ = 1.5), with e^(−1.5) ≈ 0.2231:

0 goals: 0.2231
1 goal: 0.2231 × 1.5 ≈ 0.3347
2 goals: 0.2231 × 1.5² / 2 ≈ 0.2510

Team B (λ = 1.1), with e^(−1.1) ≈ 0.3329:

0 goals: 0.3329
1 goal: 0.3329 × 1.1 ≈ 0.3662
2 goals: 0.3329 × 1.1² / 2 ≈ 0.2014

Probability of 1-1: 0.3347 × 0.3662 ≈ 0.1226, or about 12.3%. Repeating the sum for every cell gives the most likely scores:

Score (A-B)	Calculation	Probability
1-1	0.3347 × 0.3662	12.3%
1-0	0.3347 × 0.3329	11.1%
2-1	0.2510 × 0.3662	9.2%
2-0	0.2510 × 0.3329	8.4%
0-1	0.2231 × 0.3662	8.2%
0-0	0.2231 × 0.3329	7.4%

Three lessons stand out:

The most likely score is still unlikely. 1-1 comes top at about 12%: it would not happen in roughly 88% of cases.
The favourite does not necessarily own the most likely score. Adding up the whole grid, A wins about 46% of the time, the draw sits at 26% and B at 28%. Yet the single most likely score is a draw: A's wins are spread over many scorelines (1-0, 2-1, 2-0, 3-1…), whereas the draw is concentrated on a few. Our guide on reading 1X2 probabilities explains this in more depth.
Six scores barely cover more than one chance in two. The six scores in the table add up to roughly 57%.

The model is also sensitive to its inputs. With 1.8 expected goals for A instead of 1.5, A's chance of winning rises to about 54%; 1-1 stays on top but drops to 10.9%, and 2-1 (9.8%) almost catches 1-0 (9.9%). A gap of 0.3 goals, hard to settle before kick-off, is enough to reshuffle the order of scores.

The limits of the calculation

Independence is a simplification. Maher (1982) already found that allowing for a slight correlation in the two teams' goals describes observed scores more closely. Dixon and Coles (1997) later proposed an adjustment for low scores (0-0, 1-0, 0-1, 1-1), which the basic model estimates poorly.
Expected goals are estimates too. Form, absences and motivation all move them, and an average fixed before kick-off cannot see a red card or an early goal.
One scoreline neither proves nor disproves a model. Only a comparison over a large number of matches shows whether probabilities are well calibrated.

How Prediction AI presents the score

Prediction AI is an AI-powered predictive sports analysis platform. For a football match, the tool works in two stages. First it gathers verifiable facts through web search: each team's recent results, league position, absences, head-to-head meetings, context, key players and, once appointed, the referee. Form and the average goals scored and conceded are calculated from those results. An AI model then estimates the probabilities of a home win, a draw and an away win, over or under 2.5 goals and both teams scoring, and shows a most likely score as a simple reference point.

The three forecasts it offers (safe, main and bold) each carry a probability, and none of them is a correct score, precisely because that is the most uncertain outcome. The analysis also shows key factors, risks, the sources consulted and a data-quality indicator. The tool does not publish a full score grid like the one in the example. Take a look at the Prediction AI plans.

Frequently asked questions

Can AI predict the correct score of a match?

It can attach a probability to each score, not name the right one with certainty. In the example above and its variants, the most likely score carries only about 11 to 13%.

Why do 1-0 and 1-1 so often come out on top in this kind of model?

With expected goals of roughly 1 to 2 per team, the Poisson distribution gives most weight to 0, 1 or 2 goals. Low scores therefore carry the highest individual probabilities, even though all the other scores together still make up the majority.

Does the most likely score tell you the likely winner?

Not necessarily. In the example, the most likely score is a draw (1-1) while team A is favourite with about a 46% chance of winning. To see who is favourite, look at the 1X2 probabilities, not the score.

Probabilities describe uncertainty: they never promise a result.