How Our AI Predictions Work
A transparent look inside our prediction engine: from raw data to high-confidence picks, powered by six statistical models working in harmony.
データ収集
Our system aggregates data from multiple professional football data providers, ensuring comprehensive coverage and cross-validated accuracy. Data is refreshed multiple times per day.
試合統計
Goals, shots, possession, corners, cards across 30+ leagues
選手データ
Ratings, injuries, suspensions, and squad availability
Historical Records
Head-to-head, home/away form, and seasonal trends
Odds Movements
Real-time odds from leading bookmakers for value detection
統計モデル
Six specialized models analyze the data from different angles. Each model brings unique strengths to the prediction process.
ポアソン分布
Goal probability modeling
Uses team attack and defense ratings to calculate the probability distribution of goals scored by each team. Generates probabilities for exact scores, over/under, and both teams to score markets.
ELOレーティング
Team strength comparison
Adapted from chess, our ELO system rates every team based on their match results. Ratings are updated after each game, accounting for opponent strength and home advantage. Higher-rated teams are predicted to win against lower-rated opponents.
Form Analysis
Recent performance trends
Analyzes recent match results to capture momentum and current form. Considers the last 5-10 matches with recency weighting, evaluating goals scored/conceded, clean sheets, and winning/losing streaks.
Value Bet Detection
Odds vs probability mismatch
Compares our calculated probabilities against bookmaker odds to identify value opportunities. When our model assigns a higher probability than the odds suggest, it signals a potential value bet.
怪我の影響
Squad availability adjustments
Evaluates the impact of injured and suspended players on team strength. Key player absences can significantly alter predictions, especially when star goalkeepers, central defenders, or top scorers are unavailable.
Context Service
Derby, referee & motivation factors
Post-ensemble adjustments for factors that pure statistics miss: derby match detection across 19 city groups, referee tendencies for cards and goals, title race and relegation motivation, and league-specific draw calibration.
Ensemble Prediction
Individual model outputs are combined using a weighted voting system. This ensemble approach consistently outperforms any single model because different models excel in different scenarios.
Weighted Model Combination
Injury Impact and Context Service apply post-ensemble adjustments to fine-tune the final prediction based on squad availability and match context.
Confidence & Banko Selection
The final prediction includes a confidence score. Our highest-confidence picks are labeled as "Banko" and must pass three strict criteria.
Confidence Score
The ensemble model must output at least 75% confidence in the prediction.
Model Agreement
At least two-thirds of prediction models must agree on the outcome.
Odds Range
Odds must fall within a safe range — not too short, not too risky.
Live Accuracy Tracking
Every prediction is settled and tracked. Our accuracy metrics are continuously updated based on real match outcomes.
試合結果
Win/draw/loss predictions tracked per model
オーバー/アンダー
Goal total predictions with Brier score analysis
両チーム得点
BTTS yes/no accuracy with rolling 30-day average
Brier Score
Probability calibration metric — lower is better