How the projections work

Every projection on this site comes from one function, lib/predict/score.ts. Captain picks, differentials, transfer suggestions and the squad page are all the same number sorted differently.

The blend

A player's base rate per appearance mixes three things:

• Form: recent points per game. Fast to react, very noisy early.
• Season points per game: steadier, but slow to notice a change.
• An underlying model: expected goals and assists per 90 converted to points, plus Poisson clean-sheet odds for keepers and defenders.

That base rate is then multiplied by expected minutes and by fixture difficulty, and summed across the fixtures in a gameweek. A blank gameweek scores zero; a double counts both matches.

Why early-season numbers are shrunk

After one match, "form" is that single match. Left alone, a £4.0 defender who hauled once out-projects every premium in the game. That is exactly what happened the first time this ran.

So history is weighted by how many gameweeks have been played and shrinks toward the model below six gameweeks. Before anyone has played at all, even the expected-goals numbers are zero, so FPL's own ep_next carries the cold start.

Where it lies

• Minutes are a prior, not knowledge. Before a deadline the model assumes an available player probably starts. It cannot know a manager's rotation plan, so it over-rates genuine bench players for the first few weeks.
• Early projections run low. ep_next already prices in minutes and gets multiplied by the minutes factor again. Rankings are unaffected; absolute totals are conservative.
• No fixture congestion, no European football, no derbies. Only FPL's own difficulty rating.

Treat the ordering as the signal and the absolute number as a rough guide.