AI Football Predictions — How It Works
LaPreBet publishes home/draw/away win probabilities for football fixtures, generated by a third-party statistical model and compared directly against bookmaker implied odds. This page explains the data pipeline, the comparison methodology, and how we measure whether the model is actually any good.
Data sources
We use the API-Football data provider (api-sports.io) for both fixture data and pre-match prediction probabilities. This is a commercial sports data API that aggregates statistical model outputs from independent quantitative analysts. The model considers recent form, head-to-head records, home/away performance, xG (expected goals), squad strength, and injury data.
We are explicit about this: we do not claim to run an in-house AI model. The predictions are from a third-party statistical model, labelled as such in the product. Under the EU AI Act's transparency requirements (in force from August 2026), automated predictions must be disclosed — we do that by default.
What we store
Every 30 minutes, a scheduled ingest job runs per configured league. For each fixture with status NS (not started), it writes a prediction_snapshot row containing:
| Field | Description |
|---|---|
home_win_pct | Home team win probability (0–100) |
draw_pct | Draw probability (0–100) |
away_win_pct | Away team win probability (0–100) |
captured_at | Timestamp of the snapshot |
Simultaneously it writes an odds_snapshot row from the bookmaker odds (best available price per market). This means the full history of how both the model and the market moved in the days before kickoff is preserved — which is what powers the probability history chart on each fixture detail page.
Bookmaker comparison — de-vigging
Raw bookmaker odds include an overround (also called "vig" or "juice") — the built-in margin that ensures the bookmaker profits regardless of outcome. To compare fairly with the model's probabilities, we strip the overround:
raw_draw = 1 / draw_decimal_odds
raw_away = 1 / away_decimal_odds
sum = raw_home + raw_draw + raw_away
implied_home = raw_home / sum (and so on for draw, away)
The resulting three numbers sum to exactly 1.0 and represent the market's best-estimate probability for each outcome, with the bookmaker's margin removed. This is the standard method used in academic sports forecasting research.
How we measure accuracy — Brier score
Hit rate ("got it right / total predictions") is a bad accuracy metric for probabilistic forecasters because it ignores confidence. A model that says "home 51%, draw 25%, away 24%" and a model that says "home 91%, draw 5%, away 4%" both "predict home" — but the second model is making a much stronger claim, and should be penalised much more if it's wrong.
We use the multi-class Brier score instead:
Where p is the stated probability (0–1) and o is the one-hot actual outcome (1 for the outcome that happened, 0 for the others). The score ranges from 0 (perfect prediction) to 2 (maximally wrong). A random three-way forecast scores approximately 0.67.
We compute this for both the AI model and the bookmaker market on every finished fixture, using the last pre-kickoff snapshot as the reference. You can see both values on any completed fixture detail page.
What the numbers show
Across our historical Premier League data:
| Predictor | Mean Brier score | vs random baseline |
|---|---|---|
| Random (33%/33%/33%) | ~0.67 | baseline |
| Always-home (100%/0%/0%) | ~1.0 | worse |
| AI statistical model | lower than random | better |
| Bookmaker market (de-vigged) | slightly lower than AI | better; more calibrated |
The bookmaker market typically outperforms standalone statistical models slightly — this is expected, because market prices incorporate live bet-flow from sharp bettors and is continuously updated. The more interesting question is where they diverge by more than a few percentage points, and which one turns out to be right.
Crowd predictions
In addition to AI and bookmaker probabilities, LaPreBet shows a real-time crowd prediction barometer on every fixture. Any visitor can submit a prediction (exact score, result, over/under 2.5 goals, or both teams to score) without creating an account. The barometer aggregates all submissions into a home/draw/away consensus. After the final whistle, each crowd predictor gets a Brier score on their result pick — so crowd calibration is also tracked, using the same metric as the AI and bookmaker comparison.
Try it on a live fixture
Pick any upcoming match on LaPreBet, record your own home/draw/away prediction, and come back after the final whistle to see your Brier score alongside the AI's and the bookmaker's. It takes 30 seconds and requires no account.
See today's fixtures →What we don't do
- We don't accept bets or hold stakes
- We don't pay out winnings
- We don't earn affiliate commissions from bookmakers
- We don't claim proprietary AI we don't have
- We don't sell "sure things" or "guaranteed picks"
See Predictions without betting for the full positioning.