Learn

How this thing works, what the numbers mean, and whether it has been any good.

Start here

• How it works — how the model decides, in plain English, what makes it different, and the published track record including the misses.
• The track record — every graded gameweek, by model version, with mean error, root-mean-square error, correlation and bias side by side, and FPL's own forecast as the yardstick.

The words we use

Projected points (xPts) — the model's expected points for a player in one gameweek. An expectation, not a prediction of a single outcome: a projection of 5.0 does not mean five points, it means that if that same situation happened many times the average would be about five.

Bias — the average of projected minus actual. Positive means the model projects too high, negative too low. A model can have a small average error and still be biased, which is why both are published.

Correlation — how well the ranking holds up, on a scale from −1 (opposite order) through 0 (no relationship) to +1 (same order). A model can be badly calibrated and still rank players correctly, which is what matters most when you are choosing between two players.

Hold — the advice to make no transfer. The net gain of a move has to clear the four-point hit before it is worth taking, so "hold" is a real answer and not a failure to have an opinion.

Risk — how exposed a projection is to a player not starting or not finishing.

Still being built

The deeper explainers — expected points in detail, captain picks, transfer hits and how the profile shapes advice — are being written. They are not here yet rather than here and thin.