How the model is doing, misses included
Loading the record…
The full scorecard and the evidence
Candidate modelsThe control and two candidates, on the players who played 60+ minutes
Loading the model league…
Recalibration historyEvery time the model was offered a change, and what it decided
Loading the recalibration record…
What the measures meanPlain-language definitions, including how correlation reads
Average prediction error (mean absolute error, MAE). How many points off the forecast is for a typical player. Lower is better.
Large errors (root-mean-square error, RMSE). Like average error, but big misses count extra, so it catches a model that is usually close but occasionally wild. Lower is better.
Ranking quality (correlation). Whether the forecast puts players in the right order, which is what you use when choosing between two players. It is Pearson correlation between the forecast and the points actually scored, across the players in one gameweek, and it runs from −1 to +1: +1 means the same order as the results, 0 means no relationship, −1 means the opposite order. Higher is better. A model can be badly calibrated and still rank well.
Bias. Whether forecasts run high or low. Above 0 the model over-projects, below 0 it under-projects. Closer to 0 is better. A model can win on average error simply by projecting everything low, which is why bias is published beside it.
The yardstick is FPL's own forecast. FPL publish an expected-points number of their own, and it is free, so it is the honest thing to be measured against. It is scored on exactly the same measures, over exactly the same players and frozen pre-deadline snapshots.
Which forecasts countOnly forecasts frozen before the deadline
A forecast only counts if it was captured and frozen before the gameweek deadline. A row marked as not counted was captured at or after it, which means it could have seen information the deadline had already closed. Those rows stay in the history, shaded and labelled, and never appear in a headline.
A dash means a number could not be computed, for example a correlation when the projections carry no variance. It is never zero.
What happens when the model is wrong
Weights are re-searched against the frozen forecasts of past gameweeks. A candidate set only replaces the live one if it beats it on every replayed gameweek — not on average, which would let a candidate buy a win in one week with a loss in another. Most candidates fail that bar. The recalibration history above shows every attempt and its verdict.