Gameweek 7 scored: the captain finally landed, and the list lost anyway

The locked file was written at 06:38 UTC on Saturday, 3.4 hours before the deadline, by the same morning job as last week. I did not write a preview for Gameweek 7, so this one is only the table.

All ten games are finished and the bonus points are in. FPL had not marked the week as final when I pulled the numbers on Tuesday morning, so a point could still move. Nothing below depends on one point.

The list lost to the baseline

The twenty players averaged 3.40 points. The 170 players the model expects to play an hour or more averaged 3.66. That is 0.26 points a player below the baseline. Gameweek 4 was +0.36 and Gameweek 6 was +0.30, so this is the first scored week where the list did worse than just picking starters.

Against the 200 players who actually played 60 minutes, who averaged 3.87, it is worse still. Same caveat as before: nobody knew on Saturday who would play 60.

The thing I said to keep an eye on last week got worse. The twenty were projected 4.85 and got 3.40, a gap of 1.45 points a player. It was 1.0 in Gameweek 4 and 1.2 in Gameweek 6. Some of that gap is normal for any top twenty. Three weeks in a row, and growing, is starting to look like the model is too generous at the top, not just unlucky.

The captain pick finally matched the crowd

Top of the locked list was Haaland at 6.45, almost a full point clear of Groß. He scored twice in a 3-2 win over Ipswich, took all three bonus points and got 13. He was also the most captained player in the game, so it is 26 against 26. A draw, not a win.

Last week I wrote that when the top three are within a quarter of a point, the number one is not really a pick. This week the gap was 0.95, the model agreed with the crowd, and both were right. That is the easy version of the question. Across the three scored weeks, before doubling, the top pick has 17 points and the crowd's captain has 28.

The rest of the captain list: Groß 5, Saka 3, De Cuyper 6 and Cherki 0.

Every pick, scored

Four of the twenty got six points or more. Twelve got two or fewer, against ten in Gameweek 6 and eight in Gameweek 4. That number is going the wrong way.

Five players got 12 or more this week: Neto 15, Ayari 13, Haaland 13, Evanilson 12 and Kostoulas 12. Haaland was the only one on the list.

What went right

Brighton. The midweek piece said Gameweeks 6 and 7 were the two good Brighton fixtures. They beat Crystal Palace 5-0. The list had three Brighton players and they got 22 between them. Vuskovic, in 5% of teams, had an assist, a clean sheet and 12 defensive actions for 11. De Cuyper kept a clean sheet for 6. Groß had an assist but went off after 70 minutes and got 5. The two Brighton players who did the most damage, Ayari and Kostoulas, were not on it.

Mbeumo, again. Scored in a 1-1 draw at Leeds and got 8, the same as last week. Bruno, from the same game, got 1, the same as last week.

Defensive contributions, on the defenders who did not concede much. Van Hecke made 11 for Spurs and got 4. Thomas made 10 at Spurs and got 3, even with Coventry conceding three.

What went wrong

City, apart from Haaland. This is the week. The model had five City players in the twenty. Haaland got 13. The other four were projected 19.3 points between them and got 2. Cherki did not get on the pitch, with no injury listed. Gvardiol played 29 minutes, picked up a yellow and got 0. Semenyo played 13 minutes and is now listed with a knee injury. Guéhi played 90 as City conceded twice at home to Ipswich, and got 1.

Two of those are minutes, which the model cannot see coming. But all five came from the same fixture, and a home game against Ipswich looked like the best one of the week. When the model likes a fixture this much it piles in, and one bad afternoon for City's back line takes a quarter of the list with it.

Everton. Three Everton players at home to Chelsea, got 4 between them. Chelsea won 2-0. Tarkowski made 7 defensive actions, not enough for the bonus, then picked up a yellow and finished on 0. Dewsbury-Hall and Barry both played 90 for 2.

Chelsea's attack. The same game, the other side. Rogers, João Pedro and Palmer all played 90 and all got 2. Rogers and Palmer lost their clean sheet point to a yellow card each, and João Pedro had 0.28 xG and nothing else. Neto, in 2.3% of teams, scored both goals and got 15.

Stach against Calafiori, again. The locked file backed Stach, in 2.3%, against the crowd. He got 2. It left out Calafiori, in 53%, at 4.20. Arsenal won 1-0 at Forest and he got 6. Last week it was Calafiori 9 and Ampadu 1. The against-the-crowd pick has now lost both weeks, by 12 points in total.

The record so far

Across the three scored weeks the list is ahead by about 0.13 points a player. That is still positive, but it is a lot less convincing than it looked a week ago. "Likely starters" is still filtered on what the model expects now, not what it expected before the week.

The honest summary

Two weeks of a small edge, then one week of a slightly bigger loss. The captain question went fine because there was no real question. The list lost because it put a quarter of its picks into one fixture, and that fixture went wrong for everyone except Haaland.

What I take from it is the gap between projected and actual: 1.0, 1.2, 1.45. If it is still above a point after Gameweek 10, the model needs fixing, not just explaining.

Gameweek 8's deadline is Friday 23 October at 18:30 UK time. The numbers lock before that, and get scored the same way next week.