Prediction models
How AI football predictions work
An AI football prediction is an outcome probability that a statistical model computes from past match data: goals, results and team strength. A Poisson model turns each team's expected goals into scoreline probabilities, and from those come 1X2, over/under and both teams to score. An Elo rating tracks team strength across every match played, and calibration checks that the percentages match reality.
Contents
Key terms
- AI prediction
- The probability of a match outcome computed by a model, for example 49% for a home win. It is not a claim that the outcome will happen.
- Expected goals (λ)
- The average number of goals a model expects from a team in that match, for example 1.6. It usually depends on the team's attack, the opponent's defence and home advantage.
- Poisson distribution
- The distribution of the number of rare, independent events in a fixed time. It describes one team's goals in a match well.
- Elo rating
- A number that measures team strength. It rises after a better result than expected and falls after a worse one.
From statistics to probabilities
- Data: results, goals, home and away team, league and date for thousands of matches.
- Model: turns the data into expected goals or a rating for every team.
- Probabilities: the model gives every outcome of the match, from 1X2 to the exact score.
- Calibration: the percentages are compared with what actually happened and corrected.
- Measurement: after the match each prediction is scored (Brier score, log loss) and compared with the odds.
The Poisson goals model
If a team scores λ goals on average, the probability that it scores exactly k goals is:
P(k) = e^(−λ) × λ^k / k!The home team has 1.6 expected goals and the away team 1.1. The two teams' goals are treated as independent.
| Goals | Home (λ = 1.6) | Away (λ = 1.1) |
|---|---|---|
| 0 | 20.19% | 33.29% |
| 1 | 32.30% | 36.62% |
| 2 | 25.84% | 20.14% |
| 3 | 13.78% | 7.38% |
| 4 | 5.51% | 2.03% |
A scoreline's probability is the product of the two columns. A 1–1 draw has 32.30% × 36.62% = 11.83% and is the most likely single score. Adding up all scorelines:
| Outcome | Probability | Fair odds |
|---|---|---|
| 1 (home win) | 48.96% | 2.04 |
| X (draw) | 24.89% | 4.02 |
| 2 (away win) | 26.15% | 3.82 |
| Over 2.5 goals | 50.64% | 1.97 |
| BTTS (both teams to score) | 53.24% | 1.88 |
For total goals one Poisson distribution with 1.6 + 1.1 = 2.7 is enough, so over 2.5 goals is 1 − e^(−2.7) × (1 + 2.7 + 2.7²/2) = 50.64%. Both teams to score is (1 − e^(−1.6)) × (1 − e^(−1.1)) = 53.24%. Fair odds are 1 / probability, as in the implied probability guide.
Elo rating: team strength in one number
E = 1 / (1 + 10^(−(Rh + 60 − Ra) / 400))E is the home team's expected result (win 1, draw 0.5, loss 0), Rh and Ra are the home and away ratings, and 60 points is home advantage.
The home team is rated 1600 and the away team 1550. With home advantage the gap is 110 points, so E = 65.32%.
After the match the rating moves by K × (actual − expected result), with K = 20 and a bonus for a convincing win (×1.5 for a two-goal margin):
| Result | Home change | New home rating | New away rating |
|---|---|---|---|
| 1–0 | +6.94 | 1606.94 | 1543.06 |
| 2–0 | +10.40 | 1610.40 | 1539.60 |
| 1–1 | −3.06 | 1596.94 | 1553.06 |
| 0–1 | −13.06 | 1586.94 | 1563.06 |
Elo on its own does not separate a draw from half a win, so for 1X2 a draw probability is added that is highest when the teams are level.
What a model does not see
- Pre-match news: injuries, suspensions and a line-up announced an hour before kick-off, unless they are in the data.
- Motivation: cup ties, derbies, the end of the season and squad rotation.
- Changes in the team, such as a new coach or a sold top scorer, until they show in the results.
That is why an AI prediction is always compared with the odds, which often already price that information in (see AI vs bookmaker).
What StatsAI uses:
- An input model: 1X2, over/under 2.5, BTTS and half-time/full-time probabilities from an external statistical model that StatsAI imports every day.
- Its own club Elo: start 1500, K = 20, home +60 points and a larger change for a more convincing win.
- A Poisson model for goals per half, independent for the home and away team, and derived pick combinations.
- Calibration per kind of pick (see Probability calibration) and a blend with margin-free odds in the ticket generator.
- Daily measurement of every prediction on the AI model performance page.
Limitations
- The Poisson model assumes the two teams' goals are independent. In reality 0–0 and 1–1 happen slightly more often than it says.
- Elo depends on how many matches are in the data, so a team with few matches has an unreliable rating.
- The examples use made-up values (1.6 and 1.1 goals, ratings 1600 and 1550). Real models estimate them from data.
- An AI prediction is a probability, not a promise. Even an 80% forecast misses one time in five.
Related guides
Sources
- Maher, M. J. (1982). Modelling association football scores. Statistica Neerlandica, 36(3), 109–118.
- Elo, A. E. (1978). The Rating of Chessplayers, Past and Present. New York: Arco.