Developer Builds 227,000-Match Football Odds Analysis Pipeline Using LLM Architecture

A developer spent six months building an open-source football data analysis pipeline called PitchQuant, processing over 227,000 historical matches to study what signals exist within betting odds. The system combines de-vigging, ELO ratings, Dixon-Coles Poisson modelling, and Kelly criterion calculations, with an LLM coordinating a strict multi-step inference process governed by 34 knowledge files and 238 automated checks. By combining individually weak statistical signals with calibrated weighting and rigorous backtesting, the pipeline improved directional accuracy from 48% to roughly 55%. Key findings include that Asian handicap movement is a stronger indicator than European odds alone, and that low-odds favourites with shallow handicap shifts win far less often than assumed. The project, released under an MIT licence on GitHub, is intended solely for academic research and explicitly does not claim to predict match outcomes.