Fantasy Premier League is a sequential decision-making problem in which a user must manage a limited budget, squad composition and starting eleven, captain selection, transfers, and special chips over 38 gameweeks. This thesis presents a transparent software solution that makes these decisions autonomously in historical simulations and generates decision recommendations for the current FPL season. The data is prepared from pre-gameweek snapshots of the official FPL API and historical FPL data. Footballer projections are expressed as expected FPL points, making them directly comparable across different components of the solution. A mixed-integer linear programming model is used for initial squad selection and transfers, while the starting lineup, bench order, captain, automatic substitutions, and chips are determined by separate decision-making procedures. In addition to independently generating recommendations for the current FPL season, the solution also supports the simulation of different decision strategies using historical FPL data.
In the 2025/26 season, the baseline random strategy achieved an average of 1,721 points across fifty reproducible runs. The deterministic ep_next strategy, which considers only the expected FPL points of individual footballers obtained through the official FPL API, achieved 1,947 points. The full optimiser achieved 2,336 points in the 2024/25 season and 2,310 points in the 2025/26 season. The 2025/26 season is deliberately evaluated in greater detail using several comparative strategies, while the 2024/25 season is used primarily as an additional test of transferability to a different dataset. The results confirm the correct functioning of the complete decision-making cycle and demonstrate the advantage of the full optimisation approach within the evaluated setting; however, they do not constitute a guarantee of performance in future seasons.
|
