Model card
Forecast next-gameweek player points, then expose the evidence.
FPL Copilot uses a compact Ridge regression to forecast next-gameweek player points from recent performance, minutes, and fixture difficulty.
Live FPL data is unavailable while the weekly refresh is offline.
Prediction target
Next-gameweek FPL points
The scoring pipeline produces one forecast per available player for the next target gameweek. The historical training rows are fixture-level, so double-gameweek aggregation remains a known limitation.
Main model
Ridge regression
A regularized linear model turns the three inputs into a transparent next-gameweek forecast.
Baseline
Recent-points average
The model is compared with the average points from each player's latest five appearances.
Feature construction
What the trained pipeline actually uses
Rolling performance values are shifted before training so each row uses earlier appearances. Numeric features are imputed and scaled; position is one-hot encoded.
- Rolling points over the latest five appearances
- Rolling minutes over the latest five appearances
- Next-opponent fixture difficulty
Time-based validation
Latest evaluation
The pipeline holds out later gameweeks rather than randomly mixing future rows into training. MAE is the average absolute player-point error; RMSE penalizes larger misses more heavily. Lower is better.
Ridge baseline
RMSE unavailable
Unavailable
Ridge regression
RMSE unavailable
Unavailable
Holdout gameweeks
Row counts unavailable
Unavailable
Evaluation values come from the latest training run recorded by the live model service.
Live serving status
Current data, current forecast
Forecast rows come from the live Supabase database. The weekly refresh ingests official FPL data, retrains the model, and writes the next target gameweek back to the database.
Current model context
How to read the evaluation
- Historical rows combine player-gameweek history with current player context fields, with point-in-time handling as an active engineering focus.
- The recorded holdout covers five later gameweeks and gives the product a current player-point evaluation reference.
- The scoring workflow incorporates player availability, recent minutes, and fixture difficulty into the next-gameweek forecast.
- The optimizer compares legal 15-player squads and transfer choices using listed prices and transfer-hit costs.
- The product presents model status, prediction provenance, target gameweek, and generation time alongside the forecast.