AI Football Statistics

How Machine Learning Analyses Match Data to Find Value Bets

What AI football statistics analysis means

When most sites say "AI football statistics", they mean rule-based data scraping with a machine-learning label attached. Football Intelligence takes a different approach: the statistical model is a genuine Poisson distribution implementation, and the AI layer (Claude Sonnet 4.6) adds contextual analysis on top of it.

Two separate data sources feed the system. First, historical match results from the current season are used to calibrate Poisson attack and defensive ratings for every team across 10+ European leagues. Second, live odds from 40+ bookmakers are fetched via the-odds-api to detect where bookmaker prices diverge from the model fair odds.

The combination of calibrated statistical probabilities and live bookmaker price comparison is what makes the value bet detection possible. Pure statistics without live odds comparison tells you what the true probability is. Live odds without a statistical model tells you what the market thinks. The gap between the two is where value bets live.

The statistical pipeline

The complete pipeline from raw data to published AI pick:

1. Data ingestion: Historical match results are fetched from apifootball.com (primary) and football-data.org (fallback and Poisson calibration seed) for all covered leagues.
2. Attack and defensive rating calculation: For each team, attack strength = goals scored per game divided by league average goals scored. Defensive record = goals conceded per game divided by league average goals conceded. Both are normalised to 1.0 representing league average.
3. Bayesian regularisation: Early-season small samples are pulled toward the league average prior (1.0) with a regularisation weight that decreases as more matches are played. This prevents extreme ratings from 2-3 match samples.
4. Lambda derivation: Home lambda = home team attack × away team defensive record × league average goals per game × home advantage multiplier. Away lambda follows the same formula without the home advantage multiplier.
5. Scoreline matrix: The Poisson formula P(k) = (λ^k × e^−λ) / k! generates probabilities for 0–7 goals for each side, producing an 8×8 scoreline probability matrix with 64 cells that sum to 1.0.
6. Market probability extraction: Win/draw/loss, BTTS, Over/Under 2.5 and 3.5, and Asian handicap probabilities are all derived by summing the relevant cells of the scoreline matrix.
7. Odds comparison: Market probabilities are converted to fair odds (1 / probability). These are compared to live bookmaker odds. Where bookmaker odds exceed fair odds, a positive value edge exists.
8. Kelly Criterion sizing: For each value edge, the Kelly formula calculates the optimal stake fraction given the edge size and the model probability estimate.
9. Risk gate filtering: Picks are filtered by the data-driven risk gate built from 380+ settled historical bets. Only market-and-odds-band combinations with historically positive ROI are published.
10. Claude AI narrative: Claude Sonnet 4.6 reads the model output and generates a 150–200 word match analysis contextualising the probabilities with form, market context, and betting recommendations.

Attack and defensive ratings explained

Attack and defensive ratings are the core inputs to the Poisson model. Each is normalised to the league average, where 1.0 means exactly average performance.

• An attack rating of 1.3 means the team scores 30% more goals than the league average per game
• A defensive record of 0.85 means the team concedes 15% fewer goals than the league average per game
• A team with attack 1.3 playing against a team with defensive record 0.85: their expected goals against this opponent = 1.3 × 0.85 × league average × venue multiplier

This normalisation method allows meaningful cross-team comparison even when teams have played different numbers of games. It also handles league-specific goal frequencies correctly — Bundesliga averages around 3.0 goals per game, Ligue 1 around 2.7, so the ratings are always relative to that league's baseline.

Expected goals (xG) vs lambda

Football fans are familiar with xG (expected goals), which is calculated at the shot level: each shot is assigned an xG value based on its location, type, and assist quality, then summed to produce a match xG.

Football Intelligence uses lambda instead. Lambda is a match-level expected goals estimate derived from historical team performance rather than shot-level data. The key differences:

• xG requires shot-level data (location, type) for every shot — expensive to source for 10+ leagues in real time
• Lambda is calculated from readily available historical goals data and updates daily
• xG is more accurate for individual match analysis by tactical analysts; lambda is more practical for automated probability generation across hundreds of matches per day
• Over large samples, team attack strength derived from historical goals is a reliable predictor of future lambda

Lambda is a pragmatic proxy for xG that scales well and produces well-calibrated probabilities at the Poisson model level.

What the AI adds on top of statistics

The Poisson model produces numbers. Claude Sonnet 4.6 (Anthropic) produces context. The two layers are deliberately separate — Claude reads the model output but does not override it.

For each match, Claude receives the lambda values, win/draw/loss probabilities, value edges, and Kelly stakes, then generates an analysis covering:

• Which team the model favours and by what margin
• The most statistically likely scoreline range from the 8×8 matrix
• Which markets have the strongest value edge
• Any contextual factors the model cannot capture (cup context, known rotation, injury reports from public sources)

The SportGPT conversational interface (free: 2 queries/day, PRO: 15/day, Elite: 40/day) extends this to arbitrary football analysis questions in plain English.

Limitations of AI football statistics

• Poisson independence: The model assumes home and away goals are independent. In reality, tactical changes after a goal create some correlation. The model handles most matches well but may underestimate defensive outcomes in closely contested games.
• Season start volatility: Ratings are most accurate mid-season with 10+ matches of data. Early-season Bayesian priors reduce this volatility but do not eliminate it.
• No injury or team news data: The model does not know about injuries, suspensions, or likely team selections. These factors can significantly shift true probabilities away from the model estimate.
• High-variance competition types: Cup competitions (single-elimination), European away legs, and promotion/relegation playoff matches have higher variance and different tactical profiles than regular league matches. The model applies general adjustments but these are simplifications.