Research / FPL prediction
How much recent form matters
A convolutional model looks for patterns inside a player's latest matches
FPL managers constantly argue about form. The phrase usually compresses several questions into one: how many matches matter, whether the order matters, and whether a recent return changes the next expectation.
This technical report treats a player's recent performances as a one-dimensional signal. A convolutional neural network scans that sequence for local patterns, then predicts the next gameweek's points.
A moving window
Every training example contains a fixed window of past player statistics arranged in time. The model applies small filters across adjacent matches. One filter might respond to increasing minutes. Another might respond to a cluster of shots or points.
The same filter is reused across the window. That lets the model recognize a local pattern whether it appears near the start or end of the sequence.
Three models see time differently
The report compares the convolutional network with a multilayer perceptron and a long short-term memory network. The perceptron sees a set of inputs without a strong built-in idea of locality. The recurrent model carries information through the sequence. The convolutional model searches for repeated short patterns.
In the reported tests, the convolutional approach produced the lowest prediction error. The author argues that the time-series representation and local filters account for the advantage.
- 01Build recent-match windows
- 02Train CNN, LSTM, and MLP models
- 03Predict next-week points
- 04Compare mean squared error
Form is not a label. It is a choice about which history the model is allowed to remember.
nil nil interpretation
Lower error may not improve rank
Mean squared error rewards forecasts that stay close to realized points across all players. FPL decisions care more about the ordering near the top of the list and the difference between realistic alternatives.
A model can lower average error by becoming conservative while missing the explosive returns that decide captaincy. The next test should therefore evaluate transfers, captain choices, and full-team scores, not prediction error alone.
What recent form cannot contain
A player's recent line may not show a new injury, role change, transfer, suspension, opponent, or tactical shift. The sequence is useful because it is compact. It is limited for the same reason.
The strongest practical model would combine the learned recent pattern with forward information: expected minutes, fixture context, team strength, and current role.
A technical report with limited data
This 2022 technical report was not presented as a peer-reviewed journal article. The author identifies data availability as the main limitation and evaluates model error rather than a complete live FPL season. Its conclusion supports further testing of convolutions on recent player sequences, not a general rule that CNNs are the best FPL predictor.
Read the original research ↗