THE LEDGER · EVERY NUMBER CHECKED

When we say a number, how close does it land?

Every projection is graded after the final whistle against the points that actually happened — the good weeks and the bad ones, because this page is generated from the grading, not from marketing. The short version: our typical miss is about a goal’s worth of points smaller than FPL’s own projection across a quarter of a million minutes of graded football.

±1.57
pts · our typical miss
per player per gameweek
±1.86
pts · FPL's own projection
same players, same weeks
25,518
predictions graded
the model never saw them in training
76%
of weeks we were closer
than FPL's projection, 25/26 backtest

For the stats people · “typical miss” is mean absolute error (MAE), the standard grade for a points model — every prediction counts, so it cannot be gamed by only calling the easy ones.

Week by week: whose number was closer?

25/26 · 19,421 GRADED
0.71.01.3ONSIDEFPL XPGW5–GW29 · LOWER IS BETTER

Two lines, one question: whose guess landed closer to what actually happened? The engine was closer than FPL’s own projection in 76% of gameweeks, closer than “just follow recent form” in 84%, and closer than a season-average guess in 92%. The weeks it lost are on the chart too — that’s the point of the page.

How close we get, position by position

Goalkeepers0.74
Defenders1.17
Midfielders0.98
Forwards1.04

The bar is our typical miss in points — shorter is better. Keepers are the most predictable players in the game; defenders the least, because their returns arrive in lumps of clean sheets.

When a player actually scored, who saw it coming?

CLOSER IS BETTER

One average buries the question a manager actually cares about: when a player returned, did the model see it coming? So here is the same record cut by what happened on the pitch. Misses are naturally bigger on haul weeks for every model in existence — the column to read is the last one: how much closer we were than FPL’s own projection, in points.

WHAT HAPPENEDPLAYERSOUR MISSFPL'S MISSWE'RE CLOSER BY
Steady returns 3–4 pts
our biggest edge — we see the quiet contributors coming
1,1261.533.13+1.60 pts
Big hauls 5+ pts
the hardest thing in FPL to call, and where we beat FPL furthest
1,6305.967.50+1.54 pts
Blanks 1–2 pts
we flag the weeks a player does nothing
4,4191.191.48+0.29 pts
No-shows didn't play
FPL wins this band by projecting ~0 for everyone doubtful — a safe bet we deliberately don’t make
12,2460.870.34-0.53 pts

Same 19,421 graded predictions as above, cut four ways instead of averaged once. Positive means our number was the closer one.

For the stats people · “miss” here is RMSE (root mean squared error), which weights the big misses hardest — the strictest grading available.

V5 ENGINE · PRE-SEASON · AS OF 10 SEPT · ONSIDEARENA.COM