Research / FPL optimization
The best team is not the best strategy
Prediction and optimization across thirty-eight connected gameweeks
A weekly optimizer can choose the best squad it sees and still produce a poor season. FPL charges for transfers, rewards captaincy, changes player prices, and exposes teams to future fixtures. This thesis joins machine-learning forecasts with a mixed-integer model built for repeated weekly decisions.
Its main contribution is the bridge between prediction and action. Its main warning is the large gap between beating a simple baseline and approaching elite human performance.
The weekly loop
The system begins with historical player data and engineers features for the next gameweek. Several forecasting methods estimate future returns. A mixed-integer program then selects a squad under budget, position, club, captaincy, and transfer constraints.
After the gameweek, prices and results update. The model runs again from its new squad rather than rebuilding from nothing.
- 01Update player data
- 02Forecast the next returns
- 03Optimize squad and captain
- 04Carry the new state forward
Optimization beats greed
The thesis reports that the combined pipeline improves final scores compared with heuristic and greedy approaches. This is the result we should expect when transfers connect the weeks. A greedy manager takes the largest immediate gain. An optimizer can reject it when the later cost is larger.
Captaincy increases the effect because the model must distinguish between a player worth owning and a player worth doubling. Market updates add another reason to preserve a flexible squad.
The best immediate move is only best when tomorrow has no cost.
nil nil interpretation
The benchmark keeps the claim honest
The reported optimized season still sits far below the real 2023/24 winner. That gap is more informative than a victory over a weak greedy rule. It shows where practical FPL systems fail: minutes, injuries, transfers, fixture changes, and extreme player returns remain hard to forecast.
Optimization can reduce decision waste. It cannot create information that the prediction stage does not contain.
A better use for the solver
Chasing the global winner is a poor standard for a probabilistic system. The winner benefits from skill, variance, and a path that cannot be known in advance. A stronger evaluation would compare thousands of simulated or historical decision points with realistic information cutoffs.
The solver is most useful as a transparent planning aid. It can show the cost of a move, the future weeks supporting it, and the alternative path it rejected.
A useful system with a weak absolute score
This 2025 master's thesis compares its pipeline with simple decision rules, but its season total does not approach the best human results. The exact benchmark depends on the data available to the model and how future information was separated. The work supports the value of joined prediction and optimization, not a claim that the resulting manager is competitive at the top of FPL.
Read the original research ↗