Every expected points total we publish rests on a guess about how long someone will be on the pitch. This page is that guess measured against what happened.
The test
Build the minutes prior from 24/25, use it to predict the first 6 gameweeks of 25/26, then compare it to the minutes those players actually played. Both ends are FPL's own published data: last season's minutes in, this season's first 6 gameweeks out. The prior is a decay-weighted average of a player's own past rounds, half-life 10 gameweeks, and nothing in it sees team news. 415 players had at least five rounds behind them and were still in the game the next season, so they qualified.
Where the error is
| Group | Players | Average error, minutes | Bias, minutes |
|---|---|---|---|
| Every player in the test | 415 | 21.5 | +4.3 |
| Played at least one minute | 303 | 21.8 | -1.8 |
| Never appeared | 112 | 20.7 | +20.7 |
| We expected 60+ minutes | 124 | 25.3 | +16.1 |
| We expected 60+ and they played | 111 | 19.7 | +9.4 |
| prior >=80 AND played | 41 | 12.4 | +9.3 |
Bias is how far we were above what happened. The column is close to unbiased for players who got on the pitch at all, and the whole overshoot sits with the players who never appeared.
After the squad constraint
The builder normalises each club to one keeper and ten outfield players before the numbers reach a page, so these are the same groups measured after that pass.
| Group | Players | Average error, minutes | Bias, minutes |
|---|---|---|---|
| Every player in the test | 415 | 21.3 | +3.2 |
| Played at least one minute | 303 | 22.1 | -2.6 |
| Never appeared | 112 | 19.1 | +19.1 |
| We expected 60+ minutes | 124 | 25.4 | +15.8 |
| We expected 60+ and they played | 111 | 20.2 | +9.5 |
| prior >=80 AND played | 41 | 12.7 | +10.2 |
Three summers, not one
One summer could be luck, so the same test runs on three.
| Prior to season | Players | Error, live setting | Error, best setting | Overshoot, 60+ starters | Overshoot, 80+ starters |
|---|---|---|---|---|---|
| 22/23 to 23/24 | 350 | 20.95 | 20.95 | +7.5 | +12.2 |
| 23/24 to 24/25 | 351 | 21.23 | 21.19 | +5.4 | +9.8 |
| 24/25 to 25/26 | 415 | 21.46 | 21.33 | +9.4 | +9.3 |
Tuning doesn't fix it
The prior weights recent rounds more heavily than old ones. Here is the whole range, from very sharp to no decay at all.
| Setting | Average error, minutes | Bias |
|---|---|---|
| Half-life 2 | 22.871 | +3.66 |
| Half-life 4 | 21.648 | +3.95 |
| Half-life 6 | 21.332 | +4.11 |
| Half-life 8 | 21.364 | +4.21 |
| Half-life 10 (live) | 21.457 | +4.27 |
| Half-life 12 | 21.551 | +4.32 |
| Half-life 16 | 21.699 | +4.38 |
| Half-life 24 | 21.868 | +4.44 |
| Flat | 22.291 | +4.57 |
The best setting scores 21.332 and the one we run scores 21.457, which is 0.13 of a minute apart. Flat weighting, which is where most people start, scores 22.291.
What to do with this
Treat our expected minutes as a good estimate for players you're already sure are starting, and as close to no information for players you're not.
The write-up of this measurement ▸ · The projections it feeds ▸
Measured 2026-08-19. These numbers are written to this page out of the run that produced them, so the page cannot drift from the measurement. GoalIQ model predictions are statistical estimates for fun and analysis, not betting advice, and not a gambling service.