Accuracy record · 2026-27 · settled

Gameweek 3

In GW3 the model missed by 2.30 points a player across the 199 likely starters it scored, and the real team scored 56.

There’s one of these for every settled GW, good or bad, and I don’t touch them once they’re up. The misses are at the top, and everything they’re worked out from is further down.

Where Gameweek 3 went wrong

Its ordering of the players barely matched the week's: rank correlation 0.16, where 1.00 would be a perfect order and 0 no relationship at all.
Spotting 10+ hauls was no better than a coin toss this week. It came out at 64% and the interval runs from 46% to 80%, and takes in 50%.
The team finished 1.3 behind a sample of top-10k managers (average 57.3, 170 sampled). This early, the sample is chosen on the very weeks it is scored on, so the gap is overstated.
The best team the model could legally have bought scored 51, 40% of the 127 a perfect-hindsight XI would have made.

What went right

Better than usual this week - the model was out by 2.30 points a player, against 2.44 across the season so far.
The model was closer than last season's points per game on points per player: 2.40 to 2.48 on the 156 players both scored.
The model was closer than last season's points per game on the big misses: 3.08 to 3.25 on the 156 players both scored.
The model ordered the players better than last season's points per game: 0.20 to 0.07 on the 156 players both scored.
The model was better than last season's points per game at spotting 6+ hauls: 68% to 56% on the 156 players both scored.
The model was better than last season's points per game at spotting 10+ hauls: 64% to 43% on the 156 players both scored.
The model ordered the players better than top-10k effective ownership: 0.16 to 0.09 on the 199 players both scored.
The model was better than top-10k effective ownership at spotting 6+ hauls: 69% to 55% on the 199 players both scored.
The model was better than top-10k effective ownership at spotting 10+ hauls: 64% to 49% on the 199 players both scored.
The team scored 56, 5 above the average manager's 51.

Not available this week

Only 42 players scored 6+ this week, too few to claim a result on the 6+ figure.
Only 12 players scored 10+ this week, too few to claim a result on the 10+ figure.

League-wide accuracy

Scored over predicted E[minutes] > 60 (n=199 of 618 player-fixtures). Every figure below is this gameweek only; the season-to-date line beside each one is the pooled figure across all settled gameweeks.

MAE

2.30

season to date 2.44 · mean absolute error, points per player-fixture

RMSE

2.99

season to date 3.19 · punishes big misses harder

Rank corr.

0.16 (0.03–0.28)

Spearman: 1.00 means it ordered the players how the week turned out, 0 means no relationship. The bracket is a 95% row-bootstrap interval and if anything it's too narrow.

Rows

199

player-fixtures scored in this gameweek's population

Σ projected

681

points the model projected over that population

Σ realised

685

points those players actually scored (the model was −4 on the total)

P(≥6) AUC

0.69 (0.60–0.77)

put a player who went on to haul above one who didn't 69% of the time; 0.50 is a coin flip. 42 hauls, too few to claim a result

P(≥10) AUC

0.64 (0.46–0.80)

put a player who went on to haul above one who didn't 64% of the time; 12 hauls, too few to claim a result

Prediction vintage: 4 Sept, 16:27 UTC run · Predictions are from the last model run before each gameweek's deadline (90 minutes before the first kickoff). A gameweek marked 'seed' is from before I kept the runs, so it's scored against the pre-season predictions I committed.

95% percentile intervals from a row-level bootstrap (1000 resamples, seeded). Player-fixtures in the same GW share fixtures and clean sheets, which makes these intervals understate the true width - read them as a floor on the uncertainty. Fewer than 50 hauls is too few to claim a result from a haul AUC.

The same error on every group of players

I never show the every-player figure on its own (SPEC §1.6), so all four are here. Whichever group you think is the right one, my number for it is on the page - just compare it with other people's on the same rows.

Population	n	MAE	RMSE	Season MAE
Every player-fixture with a prediction	618	1.08	1.90	1.16
Predicted E[minutes] > 0	518	1.29	2.07	1.35
Played at least one minute (selected on the outcome)hindsight	290	1.91	2.65	2.00
Predicted E[minutes] > 60headline	199	2.30	2.99	2.44

Reliability of the stated probabilities

In each bin of predicted probability, what fraction actually happened, with a Wilson 95% interval. A calibrated forecaster's observed column tracks its predicted column. The minutes model's P(60+) is read on every gameweek, the haul probabilities on the gameweeks whose run retained them.

P(plays 60+ minutes)

played 60+ minutes · every projected player-fixture whose run kept this probability (n=618 of 618)

Predicted	Observed	95%	n
0%	0%	0–3	124
1%	0%	0–6	62
1%	2%	0–9	61
5%	2%	0–9	62
15%	18%	10–29	62
43%	46%	34–58	61
74%	87%	77–93	62
89%	89%	78–94	62
94%	94%	85–97	62

P(scores 6+)

scored 6+ points · every projected player-fixture whose run kept this probability (n=618 of 618)

Predicted	Observed	95%	n
0%	0%	0–3	125
0%	0%	0–6	61
1%	0%	0–6	61
2%	2%	0–9	62
5%	3%	1–11	62
9%	3%	1–11	61
15%	13%	7–23	62
23%	24%	15–36	62
35%	32%	22–45	62

P(scores 10+)

scored 10+ points · every projected player-fixture whose run kept this probability (n=618 of 618)

Predicted	Observed	95%	n
0%	0%	0–3	126
0%	0%	0–6	61
0%	0%	0–6	62
0%	0%	0–6	60
1%	0%	0–6	63
2%	3%	1–11	60
4%	3%	1–11	62
5%	3%	1–11	62
11%	11%	6–22	62

Against the comparators

The model's error next to a comparator's, on the same rows. If a comparator can't cover every row the model scored, I drop the rows it's missing and re-score the model on what's left. The model's headline number for the week isn't part of the comparison.

Comparator

Last season's points per game

Each player's total FPL points last season divided by his appearances, which is the number most of us reach for before a ball is kicked.

Different rows. This comparator scored 156 of the model's 199 rows for this GW; 43 of the model’s rows are left out because it has nothing to say about them. Every “model” number in this table is the model re-scored on exactly the 156 rows it covers. Its headline error for the GW (2.30, over 199 rows) is not the one to read against this comparator.

Population: reference arm's predicted E[minutes] > 60, imposed on every arm (n=156 of 401 player-fixtures); players with a prior-season appearance only

Last season's points per game against the model on the 156 rows both cover, Gameweek 3
Metric	Model · these 156 rows	Last season's points per game
MAE model ahead	2.40	2.48
RMSE model ahead	3.08	3.25
Spearman rank correlation model ahead	0.20	0.07
Haul AUC, 6+ points model ahead	0.68	0.56
Haul AUC, 10+ points model ahead	0.64	0.43

A dash means there’s no figure for that metric from this comparator, which isn’t the same as a zero. “No direction stated” means the data doesn’t say which way is better for that metric, and without that I don’t call a winner.

Comparator

Top-10k effective ownership

How much the top-10k sample owned each player at the last capture before the deadline, including captaincy. It doesn't come with a points prediction so I only score it as a rank.

Rows covered. This comparator scored 199 of the model's 199 rows for this GW, with none of the model’s rows excluded. Every “model” number in this table is the model re-scored on exactly the 199 rows it covers, which is the only like-for-like comparison.

Population: reference arm's predicted E[minutes] > 60, imposed on every arm (n=199 of 618 player-fixtures); every model row; double-gameweek rows excluded (ownership is per gameweek, the record is per fixture)

Top-10k effective ownership against the model on the 199 rows both cover, Gameweek 3
Metric	Model · these 199 rows	Top-10k effective ownership
MAE not carried by this arm	—	—
RMSE not carried by this arm	—	—
Spearman rank correlation model ahead	0.16	0.09
Haul AUC, 6+ points model ahead	0.69	0.55
Haul AUC, 10+ points model ahead	0.64	0.49

Comparator

FPL's own projection (ep_next)

The expected-points figure FPL publishes for each player before the GW, captured between the previous deadline and this one. It's a per-GW number and can't tell you anything about the second game in a DGW, so like the ownership figure I only score it as a rank.

Different rows. This comparator scored 0 of the model's 199 rows for this GW; 199 of the model’s rows are left out because it has nothing to say about them. Every “model” number in this table is the model re-scored on exactly the 0 rows it covers. Its headline error for the GW (2.30, over 199 rows) is not the one to read against this comparator.

Population:

FPL's own projection (ep_next) against the model on the 0 rows both cover, Gameweek 3
Metric	Model · these 0 rows	FPL's own projection (ep_next)
MAE not carried by this arm	—	—
RMSE not carried by this arm	—	—
Spearman rank correlation not carried by this arm	—	—
Haul AUC, 6+ points not carried by this arm	—	—
Haul AUC, 10+ points not carried by this arm	—	—

How the comparators are scored Lower is better for MAE and RMSE, and higher is better for Spearman and haul AUC. Which way is better is stored with each metric and the verdicts just follow that. Each baseline is scored on the model's own likely starters (expected minutes over 60) and reports the model's score over those same rows, which keeps the two numbers like for like even when the baseline can't cover every row. Overall Spearman and AUC are averages over the GWs where both sides have a value. I don't pool them across gameweeks, because that would rank a GW1 blank against a GW7 haul.

Legal-squad Team of the Week

Picked from the last model run before the deadline, at the prices captured before it, then scored on what actually happened.

Not comparable with the real entry. A fresh £100.0m squad bought from scratch every gameweek - it never pays a transfer hit or banks a price rise and nothing carries over from the week before. Think of it as a weekly check on the model's picks within the real squad rules. Nobody could run a team this way in practice, so it's not comparable with a managed entry.

XI projected

59.3

with captain 66.0

XI realised

with captain 60

Capture

40%

of the 127-point hindsight XI, a ceiling that ignores the budget and the club cap

Spend

£99.8m

of £100.0m · 4-3-3

The legal-squad Team of the Week for Gameweek 3
Player	Club	Pos	Points
Kelleher	Brentford	GK	2
O'Reilly	Man City	DEF	0
Guéhi	Man City	DEF	8
De Cuyper	Brighton	DEF	4
Collins	Brentford	DEF	1
Mbeumo	Man Utd	MID	8
Szoboszlai	Liverpool	MID	3
Saka	Arsenal	MID	2
Haaland(C)	Man City	FWD	9
Isak	Liverpool	FWD	13
Wissa	Newcastle	FWD	1
Bench (does not score)
Dubravka	Spurs	GK	0
Slater	Hull City	MID	3
Egan	Hull City	DEF	6
Yalcouyé	Brighton	MID	2

Against the real entry

TOTW

XI with the captain doubled

Entry

before hits; none were taken this week

Difference

not a like-for-like contest

Don’t read this as a head-to-head. A fresh £100.0m squad bought from scratch every gameweek - it never pays a transfer hit or banks a price rise and nothing carries over from the week before. Think of it as a weekly check on the model's picks within the real squad rules. Nobody could run a team this way in practice, so it's not comparable with a managed entry.

Deadline 4 Sept, 17:30 UTC · prediction 4 Sept, 16:27 UTC (run) · prices 18 Jul, 13:53 UTC–4 Sept, 16:27 UTC (189 from an earlier capture) · pool 626 of 626 predicted

The real entry · Sideline's Team

This is the squad the model actually ran. Unlike the Team of the Week above, it has to live with its transfer hits, its bench and whatever it carried over from the week before.

Points

no hits taken

vs average

field average 51 (FPL publishes it before hits)

vs top 10k

−1.3

sampled average 57.3 from 170 managers, sampled as of the capture

Overall rank

1,684,830

gameweek rank 3,574,059

Projected

59.8

Σ multiplier × xP over 15 of 15 picks, from the last pre-deadline run (4 Sept, 16:27 UTC)

Bench

left on the bench

Transfers

0 paid

Squad value

£99.9m

£0.5m in the bank · captain O'Reilly

The entry’s squad in Gameweek 3
Player	Club	Mins	Bonus	Points
Raya	Arsenal	90	0	3
Gabriel	Arsenal	90	0	2
Virgil	Liverpool	90	0	6
Guéhi(V)	Man City	90	2	8
Calafiori	Arsenal	66	0	2
Tavernier	Bournemouth	88	3	10
B.Fernandes	Man Utd	90	0	2
Szoboszlai	Liverpool	90	0	3
Wirtz	Liverpool	90	0	3
Mbeumo	Man Utd	90	1	8
João Pedro	Chelsea	90	0	1
Phillipsbench	Hull City	0	0	0
O'Reilly(C)bench	Man City	0	0	0
Walle Egelibench	Ipswich Town	0	0	0
Kusi-Asarebench	Fulham	0	0	0

Transfers Tarkowski → Guéhi

Cumulative after this gameweek: 206 · captured 12 Sept, 11:50 UTC · 11 squad captures

Where these numbers come from

Predictions are from the last model run before each gameweek's deadline (90 minutes before the first kickoff). A gameweek marked 'seed' is from before I kept the runs, so it's scored against the pre-season predictions I committed.

Population scored: predicted E[minutes] > 60 (n=199 of 618 player-fixtures).

This page is a slice of the same /performance/live payload that drives the full record, which means a GW page and the season page can’t disagree. The projections themselves are published in full at /data/gw/3.

The exact rows this GW was scored on (the last run before the deadline, as it was served) are at fpl-quant-api.fly.dev/performance/live/gw/3/predictions with sha256 6f8b387e…5a77. The digest is of the rows array alone as compact JSON (no whitespace), rows sorted by player then fixture, and the response says so beside it. You can recompute any number on this page from them.

What this record measures, on which rows, against which comparators and at which checkpoints was fixed in advance: the pre-registration (sha256 27307797…ba4f, checkpoints GW10, GW19, GW38).

Every settled gameweek has a page: GW1, GW2, GW3, GW4, GW5, GW6, GW7, GW8.