FPL Gameweek 2 Review: What Our Picks Got Right (and Wrong)

Gameweek 2 went well at the top and less well in the middle. Our captain pick was projected 9.66 and returned 23 — the best captain option in hindsight. Across 626 players scored, our average error was 0.777 points per player with a 0.845 correlation between projected and actual, but three picks flopped badly.

The headline: the captain call landed

We projected our captain at 9.66 points. He scored 23. That is the kind of week that makes the whole exercise look easy, and it is worth being clear that it also happened to be the best captain option in hindsight — Player #426 was the top return available, and he was the name we put the armband on.

Before anyone gets carried away, including us: a projection of 9.66 is not a prediction of 23. It never is. A projected point total is the middle of a range of outcomes, and 90 minutes with three goals, an assist and 3 bonus points sits a long way out on the good tail of that range. We were right about who to back. We were not clairvoyant about how much he would score, and no honest projection ever claims to be. The value of the pick was that it pointed at the player most likely to have a big day, not that it forecast the exact size of the day.

The rest of the good side

Two more calls worked. Player #399 was projected 7.93 and returned 14 from 81 minutes, two goals and 3 bonus. Player #411 was projected 6.96 and returned 13 from a full 90, again two goals and 3 bonus. In both cases the pattern is the same as the captain: the projection identified a player with genuine minutes security and attacking involvement, and the ceiling did the rest.

That is the honest way to read a good week. The projections got the shortlist right. The scoreline was generous.

Where we got it wrong

Now the part that matters more.

Player #115 was projected 5.75 and scored 0. He played the full 90 minutes. There was no injury, no early hook, no rotation excuse — he was on the pitch for the entire match and returned nothing at all. Zero from 90 minutes is the purest form of projection failure, because the one thing we did get right (he'd play) made the miss worse rather than better. A blank from a player who played every minute is just an attacking return that didn't arrive.

Player #586 was projected 5.3 and scored 1, having played 58 minutes. This one is a different flavour of wrong. Fewer minutes than a full match materially caps the ceiling, and our projection was built on a fuller shift. Minutes risk is the single most common way a projection goes wrong, and this is a clean example of it.

Player #565 was projected 6.24 and scored 4 from 90 minutes with no goals, no assists and no bonus. This is the least dramatic miss of the three and, frankly, the most normal. A player plays, does his job, picks up appearance points and nothing else. Over a season this outcome will happen constantly to players we like. It is not evidence the projection was broken; it is evidence that a mid-six projection contains plenty of four-point weeks inside it.

---

The numbers on our accuracy

Here is what those actually mean if you are deciding whether to trust next week's pick.

Average error of 0.777 points per player. Across all 626 players we scored, the typical gap between what we said and what happened was under a point. That is the number to hold in your head when you are agonising over two players separated by a few tenths in the projections: the gap between them is smaller than our own typical error. If the projections say one player is a fraction ahead of another, that is not a decision — that is a coin flip with a nice hat on. Use fixtures, minutes security or your own gut to break the tie and don't lose sleep.

RMSE of 1.543. RMSE punishes big misses far more harshly than small ones, which is why it sits well above the average error. The gap between those two figures is the story of Player #115 and the 23-point captain: most players land close, a handful land nowhere near, and those few outliers dominate. That is the shape of football. It is also why a projection is best used to pick a pool of players, not to predict a specific scoreline.

Correlation of 0.845. This is the one we care about most. It says the ordering held up — players we projected higher did, broadly, score higher. That is the entire job. You are not asking us for a scoreline, you are asking us who to pick, and the ranking is what answers that.

Top 10 contained a top scorer 60% of the time. Useful and a long way from perfect. It means our shortlist catches a genuine haul more often than not, and also that roughly two weeks in five the big score comes from somewhere we didn't flag. If you want a guarantee, FPL is the wrong hobby.

What to take into Gameweek 3

Trust the ordering, not the decimal. Back the top of the list with confidence, treat anything separated by less than a point as effectively tied, and remember that the misses in this review — a blank from 90 minutes, a substitution at 58 minutes, a quiet four-pointer — are all completely ordinary events that will happen again next week. A good projection week looks like this one. So does a bad one. The difference is mostly luck, and the accuracy numbers are how you tell them apart over time.