Useful numbers.
Honest uncertainty.

One snapshot, before the deadline

Player prices, teams, availability and the fixture schedule come from one snapshot taken before the gameweek deadline. Every input is fingerprinted and stored together, and the model is trained only on data from before the target week. If any input changes, or two inputs come from different snapshots, the forecast is not published. History uses only matches that kicked off before the fixture, and team news counts only if it was published, and collected by us, before kickoff.

The freeze time appears on every page. An old forecast stays an old forecast: late news does not silently rewrite it.

Starts, then minutes, then points

1. StartsThe chance a player starts each fixture.
2. MinutesExpected minutes, from starts and substitute appearances.
3. PointsFPL scoring events, averaged into expected points.

Start chances combine football history, figures we compute and pre-match availability news. Judgements read from team news inform the model; our own code does all the arithmetic, dates and counting. We project minutes and turn expected appearances, goals, assists, clean sheets and other FPL scoring events into expected points.

A start chance is the chance of starting one particular fixture. It is different from the chance of playing 60 minutes. Expected points (shortened to xPts in tables) are an average across possible outcomes, before the captain’s double; one match can finish far above or below that average.

Double weeks, blanks and missing data

Points and minutes add across fixtures in a double gameweek. The shortlist’s start chance is the largest single-fixture chance; the player page shows each one. It is not the chance of starting both. A confirmed blank has zero points and minutes. Missing forecasts are marked unavailable, and a forecast that does not cover every week is never shown as a full points total.

How we judge calibration

Calibration means the chances match reality: of the player-fixtures we give a 70% chance, roughly 70% should start over time. After the matches we score frozen forecasts with the Brier score (the average squared gap between the chance we gave and what happened, where a start counts as 1 and a non-start as 0) and a reliability chart that groups forecasts into five ranges.

For expected minutes and points, we look at the size of the errors and at bias: whether forecasts run consistently too high or too low. We also check the players managers choose between, and groups such as positions. One gameweek is a small sample; only many gameweeks can show reliable calibration.

We publish only our own forecasts against what happened, never comparisons with data suppliers. Demo numbers are made up and have no measured performance.

What the recommendations mean

Captain picks rank expected points for the target week as they are. They do not yet plan for a captain who might not play, or weigh risk against reward. Fixture ratings are our own 1–5 scale, built mostly for midfielders, and shown for context only; they do not change the forecast.

“My team” joins your last published FPL picks to this frozen forecast. It cannot see transfers or changes you have made privately for the upcoming deadline. Transfer plans and chip timing are in the full planner, which needs an account and a subscription.