Research / FPL prediction
Forecast first. Select second.
Why predicting players and building an FPL squad are different problems
A dream team is easy to find after the season. Add the highest scorers, ignore the decisions that were impossible at the time, and call the result optimal. This paper attempts the harder version: forecast player points from past seasons, then construct a legal team before the future is known.
It combines two kinds of time-series model with linear programming. One forecasts. The other assembles.
Players are moving sequences
Recent points do not arrive independently. Minutes, role, team, price, and form develop across time. The study represents each player's history as a sequence and tests two forecasting families.
ARIMA describes a series through its previous values and errors. A recurrent neural network learns a more flexible relationship across the sequence. Their predictions are combined into an ensemble rather than forcing one model to carry the entire decision.
A forecast is not a squad
Even perfect player rankings would not answer the selection problem. FPL requires exact positional counts, a fixed budget, and a limit on players from one club. Selecting one expensive player removes several other combinations.
Linear programming expresses those restrictions directly. Each player becomes a yes-or-no decision. The model maximizes predicted points while rejecting every illegal team.
- 01Forecast each player's points
- 02Attach price and position
- 03Enforce the FPL constraints
- 04Maximize the combined return
The highest predicted players do not necessarily make the highest predicted legal squad.
nil nil interpretation
The ensemble reduces commitment
An ensemble can be valuable even when neither member is consistently superior. If two models make different errors, their average may be more stable than either forecast alone. That is the paper's reason for combining the statistical and neural approaches.
But stability is not the same as calibration. An FPL forecast should say whether a predicted six points behaves like six points across many future cases. Team selection can magnify small biases by repeatedly choosing players whose estimates are too high.
The real object is a decision chain
The paper solves a static squad problem after forecasting future points. Actual FPL adds transfers, captaincy, benches, chips, price changes, injuries, blanks, and doubles. Each extra rule connects the weeks more tightly.
The study remains useful because it separates two stages that are often confused. First ask whether the numbers predict. Then ask whether the squad builder uses those numbers well.
Promising architecture, limited validation
The paper uses three past seasons to predict another and reports that selected players performed as expected. It does not provide the kind of detailed prospective benchmark needed to compare the full system with strong human managers. The method also predates several changes to FPL and modern public expected-points models.
Read the original research ↗