AI Match Result Prediction System

AI system for predicting match results

Predicting sports results is a classic ML problem with a rich history of academic research. Practical applications include bookmakers, fantasy sports, and journalism. A key limitation is that sports contain significant randomness that cannot be eliminated by model accuracy.

Task setting

Target Variant:

• Win/Draw/Loss (3-class classification)
• Win/Loss (no draw, for overtime systems)
• Score prediction (regression) → the outcome is derived from the score
• xG prediction → result through simulation

Important EMH limitation for sports: Bookmakers' prices contain aggregated information. Beating Pinnacle's closing line is harder than it seems—the sharp money is already factored in.

Data for the football model

Team strength features:

team_features = {
    # Recent form
    'points_last_5': sum(results_last_5_games),
    'goals_scored_pg_last_10': avg_goals_last_10,
    'goals_conceded_pg_last_10': avg_conceded_last_10,
    'xg_scored_pg_last_10': avg_xg_for,  # OPTA/StatsBomb данные
    'xg_conceded_pg_last_10': avg_xg_against,

    # Shots quality
    'shots_on_target_pct': shots_on_target / total_shots,
    'conversion_rate': goals / shots_on_target,

    # Fatigue
    'days_since_last_match': rest_days,
    'travel_distance_km': travel_to_venue,
    'matches_in_last_14d': fixture_congestion
}

Player availability: Injuries and disqualifications of key players are one of the most significant predictors:

# Injury impact score: взвешенный по рейтингу отсутствующих игроков
injury_impact = sum(player_ratings[player] for player in injured_players) / squad_rating

Head-to-head history: Psychological factors and tactical patterns between specific teams. Limitation: when coaching staff changes, the history is less relevant.

Poisson Goal Model

Dixon-Coles (1997): A classic in football prediction.

from scipy.stats import poisson

def dixon_coles_probabilities(home_attack, away_attack, home_defence, away_defence, home_advantage=1.1):
    """
    lambda_home = exp(home_attack + away_defence + home_advantage)
    lambda_away = exp(away_attack + home_defence)
    P(score h:a) = Poisson(h, lambda_home) × Poisson(a, lambda_away) × correction_factor
    """
    lambda_home = np.exp(home_attack - away_defence + home_advantage)
    lambda_away = np.exp(away_attack - home_defence)

    max_goals = 10
    score_matrix = np.zeros((max_goals, max_goals))
    for h in range(max_goals):
        for a in range(max_goals):
            # Dixon-Coles low-score correction для 0-0, 1-0, 0-1, 1-1
            correction = dc_correction(h, a, lambda_home, lambda_away)
            score_matrix[h, a] = poisson.pmf(h, lambda_home) * poisson.pmf(a, lambda_away) * correction

    p_home = score_matrix[score_matrix > 0].sum(where=range(max_goals)>range(max_goals))
    return score_matrix, p_home_win, p_draw, p_away_win

ML Ensemble

Models in the ensemble:

1. Dixon-Coles Poisson: Statistical Basic Model
2. LightGBM on features: nonlinear interactions of features
3. Elo/Pi-rating system: rating model (Chess-style for football)
4. Market-implied probability (от Pinnacle): cleaning через margin removal

Stacking:

meta_model = LogisticRegression()
meta_model.fit(
    X=np.column_stack([poisson_preds, lgbm_preds, elo_preds, market_preds]),
    y=actual_results
)

Model quality assessment

Log Loss: Penalizes the uncertainty of incorrect predictions.

log_loss_score = log_loss(actual_results, predicted_probabilities)
# Baseline: uniform predictions (log_loss ≈ 1.099 для 3-class)
# Рыночный baseline: log_loss ≈ 0.95
# Хорошая модель: < 0.93

RPS (Ranked Probability Score): for ranked outcomes (loss < draw < win).

Calibration: A predicted probability of 70% should correspond to winning in 70% of cases:

from sklearn.calibration import calibration_curve
fraction_pos, mean_predicted_value = calibration_curve(y_true, y_prob, n_bins=10)

Limitations and Honesty

Structural unpredictability: The best models achieve 55-60% accuracy for three-digit outcomes. This is significantly higher than the 33% expected by chance, but far from 100%.

xG-based models: use more sophisticated statistics (xG, pressure, PPDA), but historically haven't significantly outperformed simple Elo models. The reason: random variance in xG conversion is high.

Information horizon: Match day events (latest squad news, motivation) are often more important than historical statistics - available only to betting syndicates.

Timeframe: Dixon-Coles baseline + LightGBM for a single sport – 3-4 weeks. Ensemble with market calibration, injury impact, and multi-sport coverage – 8-10 weeks.