How accurate is our FPL points model?

Every forecast is published before the deadline and graded after the football. Mean absolute error across 1543 player-gameweeks with minutes: 1.96 points a player. Negative bias means the model forecast too few.

Gameweek by gameweek

Against FPL's own predictions

Both forecasts frozen at the same moment before each deadline and graded on the same 616 player-gameweeks: our error 2.02, FPL's 2.25 points a player. Lower is better.

Gameweek 1, backtested

The model did not exist for Gameweek 1. Re-running it against only the information available before that round — no results, no ratings, no minutes — gives a mean error of 2.53 points a player across 310 players. That is worse than the published weeks, which is what a model with no history should look like. It is shown for completeness and kept out of the record above.

What is being measured

Only the first forecast for a gameweek counts — the one published before the deadline, not anything recomputed afterwards. Only players who actually got minutes are scored, because forecasting a zero for an unused substitute is not a skill. Mean absolute error is the average distance between forecast and actual, in points.

Gameweek by gameweek, in full

Gameweek 1 · Gameweek 2 · Gameweek 3 · Gameweek 4 · Gameweek 5 · Gameweek 6

Individual records are on every player page, for example Haaland. The current forecasts are in the All players table.