Research / FPL strategy
Can AI learn when to transfer?
A reinforcement-learning agent plays seven historical FPL seasons
Most FPL tools score players. This thesis tries to score decisions. It trains an agent inside a simulated season, lets the agent transfer players, and rewards only the points that arrive after those choices.
The distinction matters. A player can be a good pick and a bad transfer at the same time. Price, timing, squad structure, and the alternatives already owned determine the value of the move.
FPL as an environment
The thesis models FPL as a Markov decision process. The state describes the current squad and relevant game conditions. The action changes the team within a limited set of legal moves. The reward is the realized gameweek score.
Training repeats that cycle across historical seasons. Instead of writing a fixed rule such as buy the player with the highest expected points, the system learns which actions tend to improve long-run return.
- 01Observe squad and gameweek
- 02Choose a legal action
- 03Receive realized points
- 04Update the decision policy
The unit of FPL skill is not the player pick. It is the decision made from a particular squad.
nil nil interpretation
Seven seasons of repetition
The project combines historical FPL and Understat data from 2018/19 through 2024/25. Those seasons provide training, validation, and test environments for the agent.
The author reports that the learned policy develops general patterns and consistently beats average FPL managers, even when the action space is deliberately restricted. That suggests the agent found useful structure rather than memorizing one exact sequence of transfers.
A restricted agent can still teach
Limiting the action space makes the problem tractable but changes the game. The agent may not see the creative transfer that a human would choose. It may also avoid the combinatorial explosion created by every possible hit, chip, captain, and wildcard.
That trade can still produce insight. If a restricted policy performs well, its recurring choices can be inspected. Does it preserve transfers? Buy for several weeks? Avoid hits? Concentrate captaincy? The policy becomes a source of hypotheses about strategy.
From research to decision support
A useful FPL assistant should not present reinforcement learning as magic. It should expose the state, candidate actions, expected consequences, and confidence. The manager can then understand why waiting may beat an apparently attractive transfer.
The best product opportunity is a comparison between policies: aggressive, patient, value-building, and rank-protecting. The research provides a base for that idea, but not yet a complete public system.
A recent thesis, not a final benchmark
This 2026 master's thesis is academically supervised but not a peer-reviewed journal or conference paper. Its abstract reports performance against average managers and heuristic baselines, not elite managers or the best public forecasting systems. The restricted action space and simulated historical environment may make the task easier than live FPL.
Read the original research ↗