How Football Intelligence Predicts Matches

Last updated 2026‑05‑31·Version 14.9· hello@footballinteligence.com

Data sources & league coverage

Football Intelligence ingests historical and live match data from three primary providers:

Leagues covered. Premier League, La Liga, Bundesliga, Serie A, Ligue 1, UEFA Champions League, Europa League, Conference League, Eredivisie, Primeira Liga, Ekstraklasa, Scottish Premiership, Belgian Pro League.

The Poisson distribution model

The Poisson distribution is a discrete probability distribution that expresses the probability of a given number of events (goals) occurring in a fixed interval, given a known average rate λ (lambda). Maher (1982) and Dixon & Coles (1997) showed that goals in football closely follow independent Poisson processes for each team.

The core formula:

By computing P(home_goals = i) and P(away_goals = j) for all i, j from 0–7, we generate a full 8×8 scoreline probability matrix. Summing the appropriate cells gives home win probability (Σ cells where i > j), draw probability (Σ diagonal), and away win probability (Σ cells where j > i).

Attack strength & defensive record

For each team playing at home, we compute two relative indices from their season-to-date results:

These indices measure performance relative to the league average. A value of 1.0 means average; above 1.0 means better than average in that role.

Lambda calculation for a domestic match:

League averages are maintained in a calibration table updated each season. Example: Premier League home 1.53, away 1.19; Bundesliga home 1.72, away 1.28.

Bayesian regularisation

Raw attack and defense indices are unreliable when a team has played few matches (e.g. early in the season). We apply Bayesian regularisation — shrinkage toward the prior mean of 1.0:

This ensures teams with fewer than 6 matches do not generate extreme predictions. The sample window caps at 14 games to keep the index recency-weighted.

Home advantage weighting

Home advantage is encoded in the separate home and away league averages used as multipliers. We do not apply a fixed additive home-advantage constant; the multiplicative structure naturally captures the league-specific home edge through the league_avg_home_goals parameter, calibrated from 5+ seasons of historical data per competition.

European competition adjustments

For Champions League and Europa League matches, teams come from different domestic leagues with different strength levels. We normalise using the UEFA country coefficient ranking (5-year method):

Each team's domestic lambda is scaled by the ratio of their league's strength factor. European matches use neutral-ground averages (home 1.40, away 1.18) as the baseline, reflecting the reduced home advantage in two-legged European ties.

Scoreline probability matrix

Using λ_home and λ_away, we compute P(i, j) for all scorelines i ∈ [0,7] and j ∈ [0,7]. From the 8×8 matrix we derive:

• Home win probability: Σ P(i,j) where i > j
• Draw probability: Σ P(i,j) where i = j
• Away win probability: Σ P(i,j) where j > i
• BTTS (both teams to score): 1 − P(home=0) − P(away=0) + P(0,0)
• Over 2.5 goals: Σ P(i,j) where i + j ≥ 3
• Exact score probabilities: each P(i,j) cell directly

Fair decimal odds are computed as the inverse of the probability: fair_odds = 1 / probability. Unlike bookmaker odds, these contain no overround margin.

Value betting detection

A value bet exists when a bookmaker's offered decimal odds are higher than the fair odds derived by our model:

We fetch live odds from 40+ sportsbooks and display the highest available odds for each market. Only bets with a positive value edge are surfaced. An edge of 5% or more is generally considered meaningful for long-term profitability — smaller edges sit within model uncertainty.

Kelly Criterion stake calculator

The Kelly Criterion (Kelly, 1956) is a mathematical formula for the theoretically optimal fraction of a bankroll to stake on a bet with known edge and odds:

We display both Full Kelly and Half-Kelly (50% of the calculated fraction). Use Half-Kelly or lower, because:

1. No model has perfect probability estimation — model error pulls actual edge below the calculated value.
2. Full Kelly leads to high variance and bankroll drawdowns even with a genuine edge.
3. The Kelly formula assumes independent bets and a long-run perspective.

Risk tiers — data-driven pick filtering

After a value pick survives the Kelly and edge gates, it is classified into one of three risk tiers using historical-ROI bands derived from a 380-pick analysis. Picks that historically lost money in their (market × odds) band are rejected by the algorithm and never reach users — even if the raw value calculation is positive.

The tiers below reflect only the (market × odds) bands that survive our historical-ROI filter, not every theoretically positive-EV calculation. The percentages below are hit-rate (how often a pick wins), not profit — a high-hit-rate, low-odds tier can still be net-negative on ROI. See our published accuracy & losses for the honest, settled numbers.

Each tier is calibrated against historical settled picks. Bands are reviewed quarterly and adjusted when sample size and observed ROI justify it. Public dashboard cards "Low + Medium" show the recommended-tier ROI and exclude the speculative High tier from the headline number.

Claude Sonnet AI layer — SportGPT

SportGPT is a conversational AI assistant powered by Claude Sonnet 4.6 (Anthropic). It receives the Poisson model output for a specific match as context and answers natural-language questions about probability breakdowns, form analysis, staking recommendations, and market interpretation.

SportGPT does not independently calculate probabilities — it interprets and explains the outputs of the Poisson model. The AI layer adds natural-language reasoning, contextual knowledge (injuries, team news, head-to-head from training data), and personalised staking advice based on the user's bankroll and risk profile.

Model limitations & accuracy

We are direct about what the model cannot do:

• Football is high-variance. Even an accurate statistical model cannot reliably predict an individual match. A team with a 70% win probability still loses 30% of the time.
• In-match events are not included. The model does not account for red cards, tactical substitutions, or weather in real time.
• Injury data is partial. The model uses publicly available data. Undisclosed injuries or late team-news changes are not reflected until the index is rebuilt.
• Market efficiency is high. Modern bookmaker odds are very efficient. Genuine value edges of 5%+ are rare and require large samples to realise.
• No profit guarantee. Past model performance does not predict future results. Value betting requires discipline over hundreds of bets to see statistical advantage materialise.
• Small-sample leagues. For leagues with fewer historical seasons in our dataset, predictions carry higher uncertainty and lambda values are more regularised toward league averages.

Football Intelligence displays model accuracy metrics and a settled-pick history so users can evaluate actual performance over time. We do not cherry-pick results or hide losses.

Legal disclaimer

Questions about the methodology? Email hello@footballinteligence.com.

<iframe
  src="https://footballinteligence.com/widget/picks?limit=3&accent=00e5a0&compact=true"
  width="380"
  height="290"
  frameborder="0"
  scrolling="no"
  title="Statistical football picks — Football Intelligence"
  loading="lazy"></iframe>

<iframe
  src="https://footballinteligence.com/widget/picks?limit=5&accent=00e5a0"
  width="600"
  height="400"
  frameborder="0"
  scrolling="no"
  title="Statistical football picks — Football Intelligence"
  loading="lazy"></iframe>