How our predictions work
Every probability on our prediction pages is calculated from results in our own database. Nothing is guessed, nothing is copied from bookmakers, and we publish how often we are right.
1. Team strength
Each team has a LeagueStats Elo rating, updated after every match in our data: clubs mostly since 2010-11, national teams since 1872. The rating gap between the two teams, plus home advantage (65 points for clubs, 100 for national teams, none at a neutral venue), is the only input to the probabilities.
2. From strength to goals
A larger rating gap means more expected goals for the stronger side. The relationship was fitted once on 52,880 matches played before 2022: expected home goals = e0.246 + 0.720·d and expected away goals = e0.262 − 0.718·d, where d is the rating gap divided by 400.
3. From goals to results
Goals are modelled as Poisson counts, with the Dixon-Coles correction for 0–0, 1–0, 0–1 and 1–1 (low scores are a little more and less common than simple Poisson suggests). That gives the probability of every score from 0–0 to 10–10. Adding up those scores gives the chance of a home win, a draw and an away win, the most likely score and the goals outlook.
4. When we do not predict
We would rather show no prediction than a confident wrong one. A match is not predicted when:
- our results for that competition are not up to date (some fixtures in the past still have no result in our data);
- either team has fewer than 30 rated matches, or has not played recently;
- our sources disagree about a result that the ratings depend on.
We know nothing about injuries, suspensions, line-ups, transfers or team news, and the predictions do not pretend to. They describe what the results so far imply.
5. Frozen at kick-off, checked afterwards
Predictions are updated twice a day while new results come in. Each version is stored for good and can never be edited; at kick-off the last version is frozen, and after the match it is checked against the result after 90 minutes. Every fixture has a public record of all its versions.
6. How accurate is it?
Live record: 0 predictions graded so far. We publish a live accuracy figure from 200 graded matches, because smaller samples swing too much to mean anything.
Historical backtest. We fitted the model on matches before 2022 and tested it on the 19,288 club matches since, which it had never seen:
| Approach | Outcomes called | Brier score |
|---|---|---|
| LeagueStats model | 50.5% | 0.600 |
| Always pick the home side, use average frequencies | 44.0% | 0.648 |
Football is hard to predict: draws are common and upsets frequent, so no model calls much more than half of all outcomes. What matters is that the probabilities are honest. When our model says 70%, does the favourite win about 70% of the time?
| Favourite's stated chance | Matches | Stated (avg) | Actually won |
|---|---|---|---|
| 30–40% | 3,821 | 37.9% | 37.8% |
| 40–50% | 7,260 | 44.7% | 45.1% |
| 50–60% | 4,318 | 54.5% | 53.7% |
| 60–70% | 2,282 | 64.5% | 64.9% |
| 70–80% | 1,122 | 74.4% | 74.7% |
| 80–90% | 452 | 83.9% | 82.5% |
| 90–100% | 33 | 91.8% | 69.7% |
By confidence: high-confidence calls (favourite at least 60%) were 20% of matches and 70% correct; medium 51%; low 40%. The most likely exact score came up in 12% of matches.
We show the expected number of goals but make no "over 2.5 goals" or "both teams to score" calls: in the backtest those were barely better than always guessing the average, so presenting them as picks would be misleading.
Formula 1
Every driver and every constructor has a rating, updated after each Grand Prix since 1950 from who finished ahead of whom. A driver's strength in a race is their own rating plus their car's, so a fast car lifts both of its drivers. Car ratings are pulled strongly back towards the average between seasons, because cars change; driver ratings only a little. This season's points add to that strength.
From those strengths we work out each driver's chance to win, and we simulate the race 20,000 times to get the chances of a podium and of a points finish. We predict from the previous race's line-up before the weekend, so the model does not know qualifying, weather, penalties, upgrades or reliability problems. The first race of a season is not predicted. Each prediction is frozen at the start of the race and checked against the winner.
7. Not betting advice
These are statistical estimates for fans. We do not publish odds, do not link to bookmakers and do not recommend bets. A 70% favourite still fails to win three times in ten.
About the name
"LeagueStats AI Prediction" names the automated system behind these pages: a statistical model running on our data, not a chatbot.
AI-written text
Some prediction cards carry a short AI insight: two or three sentences that explain the numbers in plain English. A language model writes them from a fact sheet that contains only the numbers already on the card and the facts listed under them. It does not choose or change any probability.
Before an insight is published, our software checks it: every number must appear in the fact sheet, the only names allowed are the two teams and the competition, and team news, betting language and links are rejected. A text that fails twice is dropped, and the card shows the numbers alone. When the prediction changes, the old text is no longer shown.
For a few of the most interesting matches each week we also publish an AI Preview: a short article on form, head-to-head, the table and the model's view. It is written with AI from a fact sheet of our own data, passes the same automatic checks plus a second AI pass that looks for claims the data does not support, and is read and approved by our editor before it appears. If the prediction changes after approval, the preview is taken down and rewritten. Previews know nothing about team news, and we never describe players or managers.
Method version 1.0.0 (elo-poisson-dc), released 21 Nov 2026. Ratings and predictions use the results shown on this site, including display-only sources; they are not part of our public API.