FPL Gameweek 3 Review: What Our Picks Got Right (and Wrong)
Gameweek 3 was a mixed one for us. Across 654 players our average error was 0.859 points, with a correlation of 0.737 between projected and actual scores. The model found Bogle (14) and Isak (13), but our captain pick, Bruno Fernandes, was projected 7.34 and scored 2.
The headline numbers
We scored 654 players in Gameweek 3. The average error was 0.859 points per player, the root mean squared error was 1.687, and the correlation between what we projected and what actually happened was 0.737.
That correlation is the number worth sitting with. It says the shape of the gameweek was broadly right — the players we rated highly did, on the whole, score more than the players we rated poorly. It does not say we called individual hauls. Nothing does. The RMSE of 1.687 sitting well above the average error of 0.859 is the mathematical fingerprint of a normal FPL week: most players land close to projection, and a handful blow the doors off or produce absolutely nothing. Those few are what everyone remembers.
Our top 10 contained a top scorer 40% of the time. Which is to say: more often than not, the very best returns of the week were not in our top 10. That is worth being blunt about, because it frames how you should use any projection, including ours. The list is a shortlist, not a lottery ticket.
The captain call, which was bad
We went with Bruno Fernandes, projected 7.34. He played 90 minutes, registered 0 goals, 0 assists, 0 bonus, and scored 2.
There is no clever framing for that. It was our highest projection of the week, we doubled it, and it returned the floor. The hindsight pick was Mitchell, who scored 15 — a defender, and not one that was anywhere near our captain conversation.
What is at least defensible is the process. Fernandes played the full 90, so this was not a minutes miss or a rotation trap we should have seen coming. He was on the pitch for the entire match and the returns simply did not arrive. That is the single most common way a projection fails and the one you can do least about. A midfielder with a high projection is being told: over many repetitions of this fixture, you expect roughly this. Any one repetition can be a blank, and this was a blank.
The honest lesson is not "stop trusting the top projection." It is that a projection of 7.34 is not a promise of 7.34 — it is the centre of a range that includes 2 at one end. If the difference between the top two captain options in your squad is small, the projection is not the thing that should decide it. Fixture, home advantage, and how comfortable you are living with a blank should carry weight too.
Where we were right
Three calls landed well, and one of them landed far beyond what we asked of it.
Bogle was projected 4.92 and scored 14. He played 60 minutes, got 1 goal and picked up 3 bonus. A defender projected under five points is, in our language, a solid-but-unspectacular pick — a clean-sheet punt with a bit of upside. He delivered an attacking return and the bonus that comes with it. We will happily take credit for having him on the list, but we should be clear we were not projecting a haul there; we were projecting a decent floor and got a ceiling.
Isak was projected 5.96 and scored 13, with 2 goals and 3 bonus from 63 minutes. This is closer to a genuine win. The projection said he was among the better forward options, and he repaid it. The 63 minutes are worth noting for future weeks — he came off, and a striker being managed like that caps the top end of what any projection can reasonably give him.
Gakpo was projected 6.71 and scored 11: 90 minutes, 0 goals, 2 assists, 2 bonus. This is the cleanest hit of the three. High projection, full minutes, delivered above it through creativity rather than a fluke finish. If you want an example of a projection working exactly as intended, that is it.
Where we were wrong
Beyond the captaincy, two more premium picks let us down, and in the same way.
Palmer was projected 5.6 and scored 1 from 90 minutes — 0 goals, 0 assists, 0 bonus. João Pedro was projected 5.28 and scored 1, also from a full 90, also with nothing across the board.
Notice the pattern. All three of our biggest misses played the entire match. Not one was an injury we missed, a benching we did not price in, or a late fitness scare. The minutes calls were correct in every case; the returns just did not come. That is a genuinely different kind of error from projecting a player who then plays twenty minutes, and it is worth distinguishing between them when you decide how much to trust next week's list.
If we had been wrong on minutes, that would be a data problem we could fix. Being wrong on whether an attacking midfielder converts in a single match is closer to the irreducible noise of the game. It will happen again. It should happen again — a model that never blanked would be one fitted to the past rather than forecasting the future.
The scorecard
| Player | Projected | Actual | Minutes | Bonus |
|---|---|---|---|---|
| Bogle | 4.92 | 14 | 60 | 3 |
| Isak | 5.96 | 13 | 63 | 3 |
| Gakpo | 6.71 | 11 | 90 | 2 |
| B.Fernandes (C) | 7.34 | 2 | 90 | 0 |
| Palmer | 5.6 | 1 | 90 | 0 |
| João Pedro | 5.28 | 1 | 90 | 0 |
What this means for next week
An average error of 0.859 points across 654 players is a respectable week's work, and a correlation of 0.737 means the rankings held up. But the captain miss and the 40% top-10 hit rate are the two figures that should shape how you read the next post.
Use the projections to build a shortlist and to avoid obvious mistakes — the players with poor fixtures and shaky minutes really do score less, and that is where the model earns its keep across a season. Do not use a single week's top projection as a guarantee. Fernandes at 7.34 was the most confident thing we said all week, and it returned 2.
We will keep publishing this scorecard whether it flatters us or not. A projection you cannot audit is not worth much.
Questions
- How accurate were the coachme projections in Gameweek 3?
- Across 654 scored players, the average error was 0.859 points per player, with a root mean squared error of 1.687 and a correlation of 0.737 between projected and actual points. Our top 10 contained a top scorer 40% of the time. Broadly, the rankings held up, but individual hauls were harder to call.
- Who should the captain have been in Gameweek 3?
- In hindsight, Mitchell, who scored 15. We went with Bruno Fernandes on a projection of 7.34 and he returned 2 from a full 90 minutes with no goals, assists or bonus. It was our highest projection of the week and it produced the floor.
- Why did coachme get Palmer and João Pedro wrong?
- Both played the full 90 minutes and returned 1 point each, with no goals, assists or bonus. Palmer was projected 5.6 and João Pedro 5.28. Neither miss came from a minutes or rotation error — the players were on the pitch throughout, the attacking returns simply did not arrive.
- Should I still trust the projections after a bad captain week?
- Treat them as a shortlist rather than a guarantee. The correlation of 0.737 suggests the overall ordering was sound, and picks like Gakpo (projected 6.71, scored 11) worked as intended. But with a top-10 hit rate of 40%, the very best returns of a gameweek often sit outside our list, so use judgement on fixtures and your own risk tolerance when two options are close.