Football Corners Forecast

A machine learning project that predicts total corners in football matches using historical data from 8 European leagues.

Predicts total corners in matches with MAE < 2.0 (average error less than 2 corners) to support betting analysis.

• API: https://huggingface.co/spaces/daniel-saed/futbol-corners-forecast-api

• Web: https://daniel-saed-corner-forecast.hf.space/

---

Web Page

Technologies

• Data: pandas, numpy, soccerdata
• ML: XGBoost, scikit-learn
• Tracking: MLflow
• Statistics: scipy (Poisson distribution)
• Visualization: matplotlib, plotly

---

📊 Dataset

Data Source

• Platform: FBref.com (via soccerdata library)
• Leagues: Premier League, La Liga, Bundesliga, Ligue 1, Serie A, Eredivisie, Primeira Liga, Pro League
• Seasons: 2018-2025 (years where advanced data is available)
• Total Matches: ~21,000

Features Collected

• Shooting: xG, shots, shots on target, shot distance, shot creation actions
• Passing: corners, crosses, passes (total attemps, progressive, last 1/3, long passes), assists
• Defense: tackles (Total, last 1/3), blocks, interceptions, clearances
• Possession: touches, carries (progressive, last 1/3, penalty area), possession %
• Goalkeeping: save %

---

Data Processing

1. Data Collection

FBref.com → Download stats → Merge leagues → Clean data → CSV

Output: dataset_cleaned.csv

2. Feature Engineering

Basic features used (9):

Processed with averages of their own leagues, example Average corners = Average corners team - Average corners league

- Average corners
- Varianze corners
- Average Xg
- Average sca
- Average crosses
- Average possession
- Average attemps in 1/3
- Average GF
- Average GA

Advanced Key engineered features used (15):

SHOTS

- shot accuracy
- xg shot
- possession_shot

PASSES

- progressive_pass_ratio
- final_third_involvement
- assist_sca
- creative_efficiency

DEFENSE

- interception_tackle
- clearance_ratio
- high press intensity

POSSESSION

- progressive_carry_ratio
- carry_pass_balance
- transition_index

ATTACK

- offensive index
- attacking presence

Other features (11):

POINTS PER GAME

- average points per game local team
- average points per game visit team
- difference poinst per game


LEAGUES ONE HOT ENCODING

- premier league
- ligue 1
- bundesliga
- la liga
- eredivise
- serie a
- primeira liga, 
- pro league

Created 269 features per match:

Category	Features	Examples
Local Team Averages	96	Form, General (Home/away) - Basic + Advance features
Visit Team Averages	96	Form, General (Home/away) - Basic + Advance features
Head-to-Head Averages	48	Last 3 matches (Home/away) - Basic + Advance features
Points Per Game Features	3	Poinst Local, Visit and Difference
League Encoding	8	One-hot encoded leagues
Team against Averages	18	Basic features against teams

Output: dataset_processed.csv

---

Model

Algorithm: XGBoost Regressor

Why XGBoost?

• Handles non-linear relationships
• Works well with 80+ features
• Resistant to overfitting
• Fast training/prediction

Training Process

Total: 21,000 matches
├── Train (70%):      14,700 matches
├── Validation (15%):  3,150 matches
└── Test (15%):        3,150 matches

Hyperparameters (found via GridSearchCV):

MLFlow image

{
    'n_estimators': 200,
    'max_depth': 4,
    'learning_rate': 0.03,
    'reg_alpha': 3.0,
    'reg_lambda': 5.0,
    'subsample': 0.7,
    'colsample_bytree': 0.7,
    'colsample_bylevel': 0.6,
    'best_gamma':1.0
}

---

📈 Results

Model Performance

MLFlow image

Set	MAE	R²	RMSE
Train	1.78	0.49	2.23
Validation	1.95	0.38	2.45
Test	1.93	0.39	2.42

✅ Test MAE = 1.93: Predictions are off by 1.93 corners on average

Currently my Model has overfit, I am still improving data and model configuration

Usual Error Distribution

Errors < 1 corner:    46%
Errors < 1.5 corners: 55%
Errors < 2 corners:   68%
Errors < 3 corners:   82%

Top 10 Most Important Features

Feature	Importance	Description
lst_team1_home_avg_ck	0.0842	Home team avg corners at home
lst_team2_away_avg_ck	0.0795	Away team avg corners away
lst_team1_home_xg	0.0623	Home team expected goals
lst_h2h_avg_ck	0.0581	Head-to-head avg corners
lst_team1_home_sh	0.0534	Home team shots
lst_team2_away_xg	0.0489	Away team expected goals

---

Prediction System

Input

predict_corners(
    local="Barcelona",
    visitante="Real Madrid",
    jornada=15,
    temporada="2526",
    league_code="ESP"
)

Output Example

🏟️  Barcelona vs Real Madrid
📅 Season 2526 | Round 15

🎯 PREDICTION: 10.3 corners
📊 Most probable: 10 corners (12.5%)
📊 80% confidence: 7-13 corners

🎯 OVER/UNDER PROBABILITIES:
Over 8.5:  72.3% @1.38 - HIGH ✅
Over 9.5:  58.1% @1.72 - MEDIUM ⚠️
Over 10.5: 43.2% @2.31 - LOW ❌

⚠️ RELIABILITY: VERY HIGH ⭐⭐⭐ (Score: 71/100)

Reliability Score

Measures team consistency:

Score = (100 - CV) × 0.4 + 
        consistency × 0.3 + 
        trend_stability × 0.3

- Score > 65: VERY HIGH ⭐⭐⭐
- Score > 50: HIGH ⭐⭐
- Score > 35: MEDIUM ⭐
- Score < 35: LOW ⚠️

---

Project Structure

futbol_corners_forecast/
│
├── config/
│   └── model_config.json          # Best hyperparameters
│
├── dataset/
│   ├── cleaned/
│   │   └── dataset_cleaned.csv    # Raw processed data
│   └── processed/
│       └── dataset_processed.csv  # ML-ready features
│
├── models/
│   ├── xgboost_corners_*.pkl      # Trained model
│   ├── scaler_corners_*.pkl       # Feature scaler
│   └── feature_importance_*.csv   # Feature rankings
│
├── mlruns/                        # MLflow experiments
│
├── src/
│   ├── models/
│   │   ├── train_model.py         # Training pipeline
│   │   └── test_model.py          # Prediction system
│   │
│   └── process_data/
│       ├── generate_dataset.py    # Data collection
│       └── process_dataset.py     # Feature engineering
│
├── EDA.ipynb                      # Exploratory analysis
└── README.md

Key Findings

What Works Well ✅

• Consistent teams → Better predictions (MAE ~1.9)
• Top leagues → More data = Better accuracy
• Mid-season matches → More historical data
• Matches where teams had low variance and low anomalies

Challenges ⚠️

• Inconsistent teams → Higher error (MAE ~2.3)
• Early season → Limited historical data
• uncertainty

---

License

Educational purposes only. Not financial advice.