Premier League Match Outcome: Multiclass Probabilistic Classification

This project implements a multiclass probabilistic classification model to predict the outcomes of English Premier League matches (Home Win, Draw, Away Win). Instead of simple labels, the model focuses on generating calibrated probabilities, which are crucial for assessing risk and value in sports analytics.

📌 Project Overview

The goal is to predict the probability of the three possible outcomes (H, D, A) using historical match data while strictly avoiding data leakage. The project employs a rolling season-by-season training strategy to simulate real-world conditions where the model is updated as new data becomes available.

🛠️ Tech Stack

• Language: Python
• Data Manipulation: pandas, numpy
• Machine Learning: scikit-learn (Logistic Regression, StandardScaler)
• Metrics & Evaluation: Log Loss, Brier Score, Calibration Curves
• Visualization: matplotlib

📊 Data & Feature Engineering

The dataset consists of Premier League matches from seasons 2018/19 to 2023/24, sourced from football-data.co.uk.

🚀 Key Features

• Advanced Feature Engineering: Utilizes Exponential Moving Averages (EMA) with a span of 5 matches to capture recent team form, ensuring zero data leakage through proper time-shifting.
• Rolling Season Training: Simulates real-world deployment by training on all previous seasons to predict the current one.
• Probabilistic Evaluation: Uses Log Loss and Brier Scores to measure the quality of the predicted probabilities.
• Calibration Analysis: Includes Reliability Diagrams (Calibration Curves) to detect and visualize model bias.
• Value Betting Simulator: A backtesting engine that calculates ROI and bankroll evolution based on a "Value" strategy against Bet365 odds.
• Dark-Theme Visualizations: High-quality plots for professional presentation.

📊 Methodology

1. Data Pipeline

Data is fetched from football-data.co.uk, covering multiple Premier League seasons (2018-2024).

• Features: Home/Away Attack Form, Home/Away Defense Form (calculated via EMA).
• Target: Match Result (H: 0, D: 1, A: 2).

2. The Model

A Multinomial Logistic Regression model is employed to provide well-behaved probability estimates across the three possible outcomes.

3. Calibration

Recognizing that models can be over-confident (especially for Away wins), the project includes an optional Isotonic Regression calibration step using CalibratedClassifierCV.

📈 Performance & Backtesting

The model was evaluated on the 2023/2024 season using a 1% value threshold.

• Total Bets Placed: 1,787
• ROI: -4.76%
• Final Bankroll: 95.24 (from initial 100)

Note: The ROI reflects the difficulty of beating the bookmaker's margin (overround). The stability of the bankroll curve suggests the model effectively captures market dynamics but requires further edge refinement.

🔮 Future Improvements

• Integration of Expected Goals (xG) data for better performance tracking.

• Implementing Kelly Criterion for dynamic bet sizing.

• Testing non-linear models like XGBoost or LightGBM.

• Inclusion of Market Closing Odds to measure "Beat the Close" performance.

🚀 How to Run

1. Clone the repository:

   git clone [https://github.com/YOUR_USERNAME/YOUR_REPO_NAME.git](https://github.com/YOUR_USERNAME/YOUR_REPO_NAME.git)

2. Install dependencies:

   pip install pandas numpy scikit-learn matplotlib jupyter

3. Run the analysis: Open and run the Jupyter Notebook: Multiclass_probabilistic_classification_football.ipynb

---

This project was developed as a demonstration of end-to-end Machine Learning workflows, specifically focusing on time-series data and probabilistic evaluation.