How many gameweeks does the first Wildcard need?

Three gameweeks gone, an international break to think in, and the first Wildcard has to be used by the GW19 deadline. The assumption I started with was the sensible-sounding one: every extra gameweek is more evidence, so the longer you wait the better the rebuild, and the only reason to go early is a squad that's on fire. I wanted to put a number on what each week of waiting was actually worth.

The setup is ten seasons of fixture-grain player data, 2016-17 to 2025-26, with xG from 2022-23 onwards. 'After k gameweeks' throughout means the Wildcard is played at the GW k+1 deadline, using GW1 to k as the evidence. The whole thing runs in under a minute, which was a bit annoying given how long I'd spent assuming the opposite.

Who's actually playing

The first thing I looked at was the simplest: how well does a player's points per game through k rank the players for the next 8 gameweeks? At k=1 form is a 0.32 rank correlation with what comes next, worse than just using last season's PPG (0.38). Form draws level at k=3 (0.42 against 0.37), climbs to 0.48 at k=6 and then sits there. From k=7 the gain from one more gameweek is zero, give or take, and the cross-season noise on these numbers is 0.04-0.07, so I wouldn't read anything into a gap under about 0.03.

What came out on top was last season's PPG scaled by how many gameweeks the player has actually turned up for. Not form, and not last season's PPG on its own. I checked that one twice. It's the best predictor at every k from 2 onwards, 0.54 by k=6 against form's 0.48, and it keeps winning all the way to GW18. It's what an experienced manager does anyway, keep last season's view of quality but only for the players who are nailed. So there are two patterns here, settling at different speeds. Who is playing is mostly resolved within 2-3 gameweeks, how well they are playing takes until about 6. That's roughly how my empirical-Bayes prior sits on top of the minutes model, so I'll admit some relief there.

Points form vs xGI

I then asked a harder question of the rates themselves, with no reference to future points: take a regular starter's per-90 rate over GW1 to k, correlate it with the same rate over the next k games, and see how much is signal. Points per 90 is 0.20 at k=3, 0.25 at k=4 and doesn't reach 0.5 until 12-14 gameweeks. xGI per 90 is 0.67 after 2 gameweeks, 0.74 at 4 and 0.79 at 6.

So a Wildcard built at GW4-8 on the FPL form number is chasing something that is 25-40% reliable, and one built on involvement is working with 75-80%. This is why the engine ranks on xP rather than form, but I hadn't seen the gap this big before. Two goals in three games is basically nothing. A run of chances you can actually do something with.

The simulated wildcard

Correlations are one thing, so I tried to turn it into actual points. For every season and every k I ran the same MILP that runs the site's optimiser on what was known at k. It picks a 15-man squad (£100m, 2/5/5/3, max three per club, prices as of GW k), I took the best legal XI out of it, no captain and no bench swaps, and scored that over the next 8 gameweeks as a fraction of the oracle squad picked with hindsight.

A GW1 squad built purely on last season captures 46% of the oracle. After one gameweek it's 55%, after two 58%, after three 61%.

That 46 to 61 step is the biggest anywhere in the first half of the season. Then it just stops: from k=3 to k=16 every number sits between 57 and 62%, and the season-to-season sd is 0.05-0.09, so k=6 and k=12 are indistinguishable. I reran that at a 12-gameweek horizon because I didn't quite believe a curve could be that flat, and got the same shape.

Honestly the worst-case numbers changed my view more than the averages. The worst single season for a last-season-only GW1 squad was 2016-17 at 26% of the oracle. By k=3 the worst season is 52%, and above half in all ten. Three gameweeks is enough to stop you building something horrible, which is probably all 'the roles have settled' ever meant. And the per-season best k, for what it's worth, was 1, 3, 7, 9, 9, 10, 13, 17, 18, 18 (the two 18s are 2017-18 and last season, the 1 is 2021-22, make of that what you will). I stared at that list for a while - if there were a right week to do it you'd expect it to turn up more than twice, and it just doesn't.

Adding it up over GW1-19

The capture numbers only cover the eight weeks after the chip. To see the cost of waiting I also did the crude accounting: hold a GW1 squad built on last season with no transfers at all through k, Wildcard, then hold that squad to GW19. Never wildcarding averages 568 points over the first half. Playing it at k=3 gives 736, the best in the table; k=4 to k=8 all land between 705 and 727; k=12 is 664 and k=18 is 581. A stale squad scores about 33 a week and a fresh one 38-40, so each week of delay costs 5-7 points here.

A caveat on that one - it makes going early look better than it is, and I know it. A real GW1 squad uses pre-season and price information (this baseline ignores new signings entirely), and real managers have a free transfer each week to patch the worst leak. Both push the optimum later, toward k=6-8, which is where the capture test goes flat. What does hold up is that k=3 to k=8 are all close and anything after GW10 is plainly worse.

What I'd do in 2026/27

GW4 is the next deadline and it's the first one the data actually backs. Most GW4 wildcards get played because GW1-3 went badly, which isn't the same thing. From here to GW8 it's a toss-up as far as the numbers go, so it comes down to fixtures and who's injured rather than how much you've seen.

• Inside GW4-8 it comes down to fixture swings, injuries, price drift and the international breaks - there's one just gone and another after GW7, which is the one I'd aim at if the squad's not on fire, since a break gives you a week to sit with the rebuild and nothing much changes in the meantime. Waiting for form to settle isn't a reason, it has settled as much as it is going to by GW6-7.
• Don't wait for the table either - I tested it and last season's goal difference beats the current table until about GW6-7 (0.66 against 0.47 after one gameweek). Last season plus the bookmakers is enough to judge a fixture swing on.
• If your squad is leaking points then just go, each week of a stale squad in the accounting above cost 5-7 points and the eight-week capture at k=3 is inside a standard deviation of anything later.
• Rank the rebuild on xP. Points per 90 is 25-40% reliable at this point in the season and xGI per 90 is 75-80%, so whatever tool you use, make sure it's looking at chances and minutes and not the last three scores.
• Keep last season's quality in the picture for players who are nailed. Last season's PPG times this season's availability beat every purely in-season number through GW18.

Genuine question for anyone with a season in mind where waiting until GW12 clearly paid off - I couldn't find it in ten seasons, there are individual years where waiting worked but no pattern I'd bet on. If someone's got one, I'd like to see it.

Ten seasons of fixture-level player data (2016-17 to 2025-26) from the FPL API and the vaastav archive; xG from 2022-23. Every table and chart is reproduced by scripts/wildcard_timing.py in the project repo.