Gameweek 4 Scored: The Model Beat the Field and Lost the Captaincy

A projection you can revise after the fact is worth nothing. So before the Gameweek 4 deadline the model's top twenty were captured to a file, timestamped, and not touched again. This is the first time that record has been scored.

The list beat the field

The twenty averaged 4.20 points. For comparison, every player who appeared in Gameweek 4 — all 307 of them — averaged 3.00. But that is a soft benchmark, because it includes substitutes who played eleven minutes.

The honest comparison is against players who were likely to start, and there the average was 3.84. So the edge was 0.36 points a player, about 9%. Real, and small.

Picking players who actually start is most of what a projection does. Beating that baseline by 9% across twenty names is the sort of edge that compounds over a season — if it holds. One gameweek does not tell you whether it holds.

The captain pick did not

The model's top captain pick was Isak, projected 6.03. Liverpool beat Forest 4–1. Isak scored 2 points.

The most-captained player in the game was Haaland, who scored 9. Doubled, that is 4 against 18 — a clear loss on the one decision that matters most in any given week.

The mitigating detail, and it is only a detail: Haaland was fourth on the same list at 5.34, a quarter of a point behind Isak. The ranking was close to right. The call was wrong. That distinction matters for diagnosing the model, and not at all to anyone who took the captaincy.

Every pick, scored

#	Player	Team	xP	Actual
1	Isak	LIV	6.03	2
2	João Pedro	CHE	5.74	12
3	Muharemović	LEE	5.48	7
4	Haaland	MCI	5.34	9
5	Scott	BOU	5.32	4
6	Szoboszlai	LIV	5.31	3
7	Saka	ARS	5.26	8
8	Tavernier	BOU	5.26	8
9	Rogers	CHE	5.26	8
10	B.Fernandes	MUN	5.19	2
11	Thomas	COV	5.17	0
12	Palmer	CHE	5.17	5
13	Stach	LEE	5.14	3
14	Mendy	HUL	5.13	1
15	Nketiah	CRY	5.11	0
16	Ødegaard	ARS	4.92	3
17	Dewsbury-Hall	EVE	4.82	2
18	Lewis-Potter	BRE	4.82	1
19	Justin	LEE	4.82	2
20	M.Sangaré	BRE	4.80	4
	Mean		5.21	4.20

Six of the twenty returned six points or more. Eight returned two or fewer. That spread is not a flaw in the model — it is the actual shape of Fantasy Premier League. Points arrive in lumps, and a list that is right on average is still wrong about most individual players in any single week. Anyone selling you weekly certainty is selling you something else.

Nketiah is worth singling out, because it is the least defensible entry. He was projected 5.11 and did not feature at all. He is scored as zero here, not quietly removed — if you had picked him on this list you would have had nothing.

The Bruno test

Before Gameweek 4 the model was changed in one specific way, and this was its first public outing.

Bruno Fernandes had scored 2, 23 and 2 in the opening three gameweeks. On season totals he looked like one of the best attacking midfielders in the game — and the model rated him top of the list at 7.55, recommending him as captain against Manchester City. That was obviously wrong to anyone who had watched the three matches, and it was wrong for a specific reason: 90% of his expected goals came from a single afternoon.

The model now reads every gameweek separately instead of the season total, and measures how concentrated a player's output is. Bruno fell from 7.55 to 5.19 and from first to tenth. He scored 2.

One gameweek does not prove that fix. It would have looked just as good by luck. But it did not contradict it, and the reasoning stands on its own regardless of what happened on the day.

What changed for Gameweek 5

Gameweek 4 exposed something the scoring did not: FPL rated Chelsea at home to Hull and Liverpool at home to Forest identically, both a 2. The model had no way to tell them apart on the attacking side, even though its own numbers had Forest among the meanest defences in the division and Hull at league average.

That inconsistency has been fixed. The attacking adjustment is now a 50/50 blend of FPL's difficulty rating and what the opponent's defence is actually conceding. The clean-sheet side of the model always worked that way; the attacking side simply did not. The full method is on the how it works page.

It would not have saved Gameweek 4. Isak scored 2 despite Liverpool winning 4–1, which is variance at the player level, not a misread fixture. The change is right because the logic was inconsistent, not because it wins last week over again.

The honest summary

One gameweek is an anecdote. The list beat the starter baseline, the captain pick lost to the crowd, and both of those facts have a sample size of one. By Gameweek 10 there will be something worth arguing about.