Three FPL decisions, three time horizons: captaincy, transfers and squad building

For gameweek 6, Sunderland’s Isidor (£5.5m) is projected for 4.8 points, more than Leeds’ Calvert-Lewin (£6.0m) at 3.6. Sort a table by next week’s points and Isidor wins. Over the next ten gameweeks Calvert-Lewin is projected for 49.0, Isidor for 14.7.

The reason is in the start chances. Isidor has played 117 minutes this season, all from the bench; he starts the next two games only because Brobbey is out with a hamstring injury until late October. Calvert-Lewin starts every week.

Neither projection is wrong. They answer different questions. Who scores most next week is a captaincy question. Who adds most to your team over the time you will own him is a transfer question. Which fifteen players, together, score most under the rules is a squad question. Using one question’s answer for another is where a lot of FPL points go missing.

The armband doubles one player’s points for one gameweek, so the question is narrow: whose expected points are highest this week? Everything that matters for that week is inside a good projection: how many goals his team should score against this opponent, home or away, his share of the goals and assists, penalties and set pieces, clean-sheet and defensive points where they apply, bonus, and whether he will play.

Two things that feel relevant mostly aren’t. Recent points are a noisy guide to next week’s (see the “due” test below). And a player’s reliability over the season doesn’t matter for the armband; only his chance of playing this match does, and it is already in his projection. A player who is 75% to start has a quarter of a blank built in. In a double gameweek his two fixtures add up, which is why doubles dominate captaincy.

Saka edges it on assists and an Arsenal clean-sheet chance that Haaland, away at Liverpool, doesn’t have; Haaland still has the highest goal threat. The right pick is the highest projection. But a gap of 0.4 points is small: over four seasons, when the top two captain options were within half a point, the top one won the head-to-head about as often as he lost it (49% against 46%) (our captaincy test). Close calls are close.

A transfer is a trade, so its value is a difference, summed over the weeks you will own the new player:

Take two regular starters at about the same price: Wissa (£6.2m) and Calvert-Lewin (£6.0m). If you own Wissa, swapping him for Calvert-Lewin loses 1.0 points in gameweek 6. It breaks even by gameweek 8, is worth +2.4 after five weeks and +8.2 after ten, and frees £0.2m.

The simple sum leaves out five things that change real decisions:

Pick the highest-projected players for each slot (two goalkeepers, five defenders, five midfielders, three forwards) and you get 734.2 projected points over ten weeks, for £116.9m, with 4 Arsenal players. Over budget and over the three-per-club limit. The real problem is to find the best squad within the rules, which is an integer programme (Ramezani and Dinh, 2025, describe the standard form).

Solved exactly for gameweeks 6 to 15, the best £100m squad projects 644.1 points, costs £99.5m, and has about £78.9m of it on the pitch in a typical week. It buys a £4.0m goalkeeper it never plays: bench points only count when someone drops out, so the cheapest legal bench frees money for the XI.

Points per £m divides by the whole price, as if a player bought for £0m scored nothing. But you can always fill a slot cheaply with someone who plays: the best player within £1m of the cheapest at each position projects 33.0–39.8 points over the ten weeks. The useful measure is what a player adds above that replacement, per extra £1m spent:

Where the marginal points are decides where the money goes. Right now the best value is in a few cheap, nailed attackers (Barry, Tavernier, Calvert-Lewin); the rest goes on the players who hold the most points per place, Haaland and Saka. That will change as prices and fixtures move, which is why the squad problem has to be re-solved, not remembered.

The three questions overlap in one place: a second premium can be worth more than his own points because he can take the armband in your first premium’s bad weeks. We tested that in One captain or two?. In short: on these projections the best squad holds Haaland and Saka (644.1 points), against 626.0 with no £9m+ partner beside Haaland. But only 2.2 of those 18.1 points is captaincy; the rest is Saka’s own points. Allowing transfers doesn’t change that. Two premiums pay when the second is a good player in his own right, not as captaincy insurance.

Every number above is an average over the ways a gameweek can go. Real gameweeks are noisy: in our simulation of gameweek 6, any XI’s score varied by about ±15 points from its average, while the fifty best XIs were separated by only 0.46 projected points. So a good decision often loses, and a bad one often wins. Judge decisions on what was known before the deadline, over many weeks, not on one result.

After three blanks in a row, regular starters scored +0.1 points above their usual in the next game: no different from any other game. After three returns in a row, +0.3: no hot streak either. A player’s next game is just his next game.

One trap is worth knowing about, because we fell into it first. Compare the next game with the player’s season average and he looks “due”: 5.1 against 4.6. But his season average includes the three blanks you picked him for. Against his other games the effect disappears. Miller and Sanjurjo (2018) show how easily streak comparisons mislead.

A forecast is useful if it is calibrated: things it calls 30% happen about 30% of the time (Gneiting, Balabdaoui and Raftery, 2007). We can check this for start chances: on gameweeks 3 to 5 (1,761 player-matches), the average predicted start chance was 33% and 34% of those players actually started: close, slightly cautious. We can’t yet do the same for points: until gameweek 6 the projections were overwritten after kick-off, so there is no honest record of what they said beforehand. That check comes once a few of the kept gameweeks are played.

Maximising expected points maximises your expected total. It doesn’t maximise your chance of winning a mini-league or reaching the top 1%, which depends on what everyone else does. In the same simulation, a template XI of popular picks projected 56.3 points and a differential XI 54.0, but the differential XI had about twice the chance of a top-1% week (12% against 5.7%, with a simulated field that overstates both). Chasing a rank you need, late in a season, can justify giving up expected points. Most of the season, it doesn’t.

Picking the best transfer each week isn’t the same as planning the best sequence: a transfer now changes what you can do next week. Multi-week solvers plan the whole sequence, but they are only as good as their assumptions. Ours, left free, sold Haaland for one bad fixture and bought him back the week after, which only works because it held prices fixed; in the real game that round trip loses money (details).