FootballGPT: Technical Analysis of an Open-Source Football Match Prediction Engine

FootballGPT is an open-source football prediction system based on XGBoost and LSTM neural networks. By analyzing over 500,000 match event data points and integrating advanced features such as Expected Goals (xG) and PPDA defensive pressure metrics, it achieves statistical prediction of match outcomes across 8 major leagues. The core of the project is to identify mispriced odds by bookmakers, find "value bets", and emphasize transparent open-source practices and strict fund management.

The sports prediction field has many "black box" services with opaque logic that are hard to verify. FootballGPT was initiated by engineers to build a fully open-source, logically transparent system. Its core philosophy is "Football is not luck, but mathematics", and the team focuses on finding statistically valid "value bets" rather than claiming to be prediction experts.

FootballGPT adopts a machine learning pipeline architecture:

• Raw Data Layer: Integrates multi-dimensional data such as historical matches, real-time odds, player injuries, and team form;
• Feature Engineering Layer: Core competitive advantage, including professional features like xG models, PPDA defensive metrics, xG-adjusted ELO ratings, and squad depth;
• Model Integration Layer: Dual models (XGBoost classifier + LSTM neural network) + Bayesian calibrator;
• Output Decision Layer: Outputs match outcome probabilities, confidence levels, edge calculations, and value bet markers (triggered when edge exceeds 6%).

The project emphasizes fund management discipline for long-term profitability:

• Kelly Criterion: Dynamically calculates optimal betting proportions to balance returns and risks;
• Risk Control: Single bet does not exceed 2-3% of total funds, highly selective strategy (only a few edge predictions are released weekly);
• Disciplinary Principle: "No edge = no release", distinguishing professional strategies from amateur guesses.

The system covers 8 major leagues: Premier League, Champions League, La Liga, Bundesliga, Serie A, Ligue 1, Eredivisie, and Africa Cup of Nations. Practical performance is publicly tracked:

Season	Number of Predictions	Win Rate	ROI
2023/24	187	58.3%	+11.2%
2024/25	214	61.2%	+14.7%
2025/26	In Progress	Updating	Updating
The project reminds that "past performance does not guarantee future results" and clearly states the risks of gambling.

FootballGPT is fully open-source (MIT License), providing a complete code repository (data processing, model implementation, fund management, result tracking). The team releases daily predictions and analyses via Telegram channel, all content is free with no VIP or paid subscriptions.

Core logic: "Do not predict winners; find mispriced odds". For example, if the model calculates Manchester City's winning probability as 65% and the bookmaker's implied probability is 57%, there is an 8% edge. Key insight: Continuously finding positive edge bets can lead to long-term profits (mathematical expectation >1). Project insights: Success comes from high-quality feature engineering, strict methodology, transparent open-source practices, and responsible risk management; domain knowledge and ethical responsibility are equally important.