Machine Learning Models

Overview

BetIntel uses an ensemble of ML models to generate accurate football predictions.

Available Models

PRODUCTION_V3_GBM (Recommended)

The most advanced model combining Production V3 with GBM Ensemble.

Metric	Value
Accuracy 1X2	62-65%
Over/Under 2.5	64%
BTTS	52%
RPS	0.160
Features	140+

Composition:

• 55% PRODUCTION_V3 (stability)
• 45% GBM Ensemble (precision)

# Combination weights
PRODUCTION_WEIGHT = 0.55
GBM_WEIGHT = 0.45

combined_probs = PRODUCTION_WEIGHT * prod_probs + GBM_WEIGHT * gbm_probs

GBM Ensemble

Ensemble of gradient boosting machines:

Model	RPS	Weight
LightGBM	0.1612	50%
CatBoost	0.1609	50%

Files:

• models/advanced/lightgbm_optimized.pkl
• models/advanced/catboost_optimized.pkl
• models/advanced/best_params.json (feature columns)

PRODUCTION_V3

Base model with good stability.

Metric	Value
Accuracy 1X2	60.1%
Over/Under 2.5	64.1%
BTTS	52.4%
RPS	~0.18

Files:

• models/ensemble_v3.pkl - 1X2 predictions
• models/over25_model.pkl - Over/Under
• models/btts_model.pkl - BTTS

Pi-Rating System

Dynamic rating system superior to ELO for football.

class PiRatingSystem:
    def __init__(self, initial_rating=0.0, lambda_decay=0.035):
        self.lambda_decay = lambda_decay
        self.home_ratings = {}
        self.away_ratings = {}

Based on: Constantinou & Fenton (2012) - "Determining the level of ability of football teams by dynamic ratings"

Prediction Pipeline

Match Data
    │
    ▼
┌─────────────────────┐
│  Feature Extractor  │  140+ features
└─────────────────────┘
    │
    ▼
┌─────────────────────┐
│   PRODUCTION_V3     │  Base predictions
└─────────────────────┘
    │
    ▼
┌─────────────────────┐
│   GBM Predictor     │  LightGBM + CatBoost
└─────────────────────┘
    │
    ▼
┌─────────────────────┐
│   Combine (55/45)   │  Weighted average
└─────────────────────┘
    │
    ▼
┌─────────────────────┐
│   Calibration       │  Temperature scaling
└─────────────────────┘
    │
    ▼
Final Prediction

Supported Markets

1X2 (Match Result)

• prob_home_win: Home win probability
• prob_draw: Draw probability
• prob_away_win: Away win probability

Over/Under 2.5

• prob_over_25: Probability > 2.5 goals
• prob_under_25: Probability ≤ 2.5 goals

BTTS (Both Teams To Score)

• prob_btts_yes: Both teams score
• prob_btts_no: At least one doesn't score

Double Chance

• prob_1x: Home or draw
• prob_x2: Draw or away
• prob_12: Home or away

Asian Handicap

• asian_handicap_line: Handicap line (e.g., -0.5, -1.0)
• asian_handicap_home_odds: Home odds
• asian_handicap_away_odds: Away odds

Evaluation Metrics

RPS (Ranked Probability Score)

Measures probability calibration.

def rps(probs, actual):
    """
    RPS = 0: perfect prediction
    RPS = 1: worst possible prediction
    """
    cum_probs = np.cumsum(probs)
    cum_actual = np.cumsum(actual)
    return np.mean((cum_probs - cum_actual) ** 2)

Benchmark	RPS
Random (33/33/33)	~0.22
SOTA literature	0.195
BetIntel	0.160

Accuracy

Percentage of correct predictions for 1X2 market.

Calibration

Verifies that predicted probabilities match actual frequencies.

Training

Dataset

• 5+ seasons of historical data
• 5 top European leagues (Serie A, Premier League, La Liga, Bundesliga, Ligue 1)
• ~10,000+ matches for training

Cross-Validation

from sklearn.model_selection import TimeSeriesSplit

tscv = TimeSeriesSplit(n_splits=5)
for train_idx, val_idx in tscv.split(X):
    # Train and evaluate

Hyperparameter Tuning

# LightGBM best params
params = {
    'n_estimators': 500,
    'max_depth': 6,
    'learning_rate': 0.05,
    'num_leaves': 31,
    'feature_fraction': 0.8,
    'bagging_fraction': 0.8,
}

Future Development

1. Error Analysis Notebook - Error pattern analysis
2. Probability Calibration - Fix for low confidence
3. League-Specific Models - Per-league models