Ten seasons of captaincy rules, and the simple one that beats the template
I tested 20 ways of picking a captain over 339 gameweeks. The one that won is pretty basic, and my own model didn't beat it.
Photo: Erling Haaland, 2023 — Jacek Stanislawek / Wikimedia Commons (CC BY-SA 4.0)
The template captain is whichever midfielder or forward is in the most squads that week, and it's been Haaland or Salah in 186 of the last 339 gameweeks I looked at. Most of us just give him the armband. I wanted to know whether that actually gets you points, or whether we all do it because nobody gets punished for it, so I tested 20 other ways of choosing.
I had each rule pick one attacker a week from GW5 to GW38 across ten seasons (2016/17 to 2025/26, 339 gameweeks), only using roughly what you'd know before the deadline - ownership, price, points so far and the market odds. I doubled whatever the pick actually scored and compared that with the template's captain the same week. I got each team's expected goals from the match odds (the result price and the over/under 2.5 line) that football-data collects on the Friday afternoon, or the Tuesday for midweek games.
The rule that won
Take the three most owned attackers and captain whichever one's team is expected to score the most goals.
It only picks someone different in 162 of the 339 weeks. When it does, it comes out ahead 61% of the time and behind 29%. Doing it with the top five instead of the top three is about the same (+2.1).
With 20 rules I was worried I'd just found the luckiest one, so I split the ten seasons in half. It was the best rule on 2016-21 (+2.0) and then did +2.2 on 2021-26. The 8 out of 10 is softer than it looks though. It was only just ahead in 2017/18, and fitting the expected goals to the odds a slightly different way flips one pick that season and makes it 7 of 10. Either way it's still about +2 a week, which is the bit I actually care about.
All twenty rules
| Rule | Captain pts/GW | vs template | 95% range | Seasons ahead | Blanks |
|---|---|---|---|---|---|
| Top-5 owned: most team expected goals | 14.6 | +2.1 | +1.0 to +3.3 | 8/10 | 28% |
| Top-3 owned: most team expected goals | 14.5 | +2.1 | +1.1 to +3.0 | 8/10 | 27% |
| Top-10 owned: points per game × team goals | 14.1 | +1.7 | +0.6 to +2.8 | 8/10 | 27% |
| Sideline model: top-3 owned, highest xP² | 13.9 | +1.6 | −0.2 to +3.5 | 3/3 | 30% |
| Top-3 owned: team most likely to win | 14.0 | +1.6 | +0.7 to +2.5 | 8/10 | 29% |
| Sideline model: highest xP² | 13.5 | +1.2 | −0.8 to +3.3 | 3/3 | 31% |
| Top-10 owned: xGI last 4 × team goals¹ | 13.8 | +1.1 | −0.5 to +2.8 | 4/4 | 34% |
| Most-owned in the highest-scoring team | 13.5 | +1.1 | −0.2 to +2.3 | 7/10 | 31% |
| Most 10+ hauls this season | 13.1 | +0.7 | −0.3 to +1.6 | 6/10 | 32% |
| Most-owned attacker playing at home | 13.1 | +0.6 | −0.4 to +1.7 | 7/10 | 32% |
| Season's top scorer so far | 12.7 | +0.3 | −0.7 to +1.3 | 4/10 | 34% |
| Top-5 owned: most xGI in the last 4 GWs¹ | 12.8 | +0.1 | −1.6 to +1.8 | 2/4 | 42% |
| Most-owned attacker (the template) | 12.4 | – | – | – | 37% |
| Best points per game (4+ apps) | 11.8 | −0.6 | −1.7 to +0.5 | 3/10 | 37% |
| Form: most points in the last 4 GWs | 11.8 | −0.6 | −2.0 to +0.7 | 4/10 | 36% |
| Most expensive attacker | 11.6 | −0.8 | −2.0 to +0.4 | 3/10 | 40% |
| Last season's points per game³ | 11.4 | −1.3 | −2.7 to +0.2 | 3/9 | 42% |
| Most xGI in the last 4 GWs¹ | 11.2 | −1.4 | −3.4 to +0.5 | 1/4 | 44% |
| Random pick from the top-5 owned | 10.9 | −1.6 | −2.6 to −0.5 | 2/10 | 43% |
| Differential: outside the top-15 owned, PPG × team goals | 7.1 | −5.4 | −6.8 to −4.0 | 1/10 | 63% |
Captain points per gameweek (double the pick's actual points), GW5-38 over 2016/17 to 2025/26. 'vs template' is the average weekly gap to captaining the most owned attacker, with its 95% range. Seasons ahead counts seasons where the rule averaged more than the template. Blanks are weeks the captain scored 2 or fewer. ¹ 2022/23 on, when FPL started publishing xG (135 GWs). ² 2023/24 on, my model's walk-forward predictions (102 GWs). ³ No 2016/17, there's no season before it in the archive (305 GWs). Rows marked ¹, ² or ³ cover fewer gameweeks, so compare them on 'vs template' rather than on captain points.
A few of these surprised me, team goals beating team win probability most of all (+2.1 vs +1.6), though it makes sense when you think about it, a scrappy 1-0 doesn't do much for your forwards. Nothing that only looks at the player beat the template by more than noise, and form (-0.6), price (-0.8), last season's points (-1.3) and recent xGI (-1.4) all came out behind it. Put the fixture back in, points per game times team goals among the top 10 owned, and it's +1.7.
The crowd is also better than I gave it credit for. A random pick from the top five owned loses 1.6 a week, and going properly differential (the best points per game times team goals outside the 15 most owned attackers) loses 5.4. So I'd stay with the template guys and just choose between them on the fixture.
Fixtures
The template captain's own scores follow his team's chances pretty closely.
| His team's chance of winning | Gameweeks | Avg points | 10+ hauls |
|---|---|---|---|
| Under 40% | 42 | 3.7 | 17% |
| 40-55% | 51 | 6.1 | 24% |
| 55-70% | 96 | 5.8 | 21% |
| Over 70% | 130 | 6.5 | 22% |
The template captain's raw points (not doubled) by his team's market chance of winning, single-fixture gameweeks only.
That's 3.7 points on average in the 42 weeks his team was under 40% to win, against 6.5 in the 130 weeks they were over 70%.
Did my model do better?
No, not on captaincy, which I wasn't expecting. I've only got walk-forward predictions for the last three seasons, so it's 102 gameweeks rather than 339, but on those same weeks:
| Rule | Captain pts/GW |
|---|---|
| Top-3 owned: most team expected goals | 14.5 |
| My model: top-3 owned, highest xP | 13.9 |
| My model: highest xP | 13.5 |
| Most-owned attacker (the template) | 12.2 |
Captain points per gameweek on the 102 gameweeks of 2023/24 to 2025/26 that my model has walk-forward predictions for.
The simple rule is 1.1 a week ahead of my model's top pick (give or take 1.8 at 95%) and 0.6 ahead of it limited to the top three owned (give or take 1.3). Statistically you can't separate them, but I'm not going to pretend it goes my way.
Where I think the model does earn its keep is minutes, which a fixture rule can't see at all. I checked its start probabilities against what actually happened and they hold up across the range:
| Model said | Average prediction | Actually started | Player-fixtures |
|---|---|---|---|
| Under 10% | 2% | 2% | 47,698 |
| 10-30% | 18% | 18% | 9,059 |
| 30-50% | 40% | 38% | 5,011 |
| 50-70% | 60% | 59% | 5,019 |
| 70-90% | 82% | 82% | 10,269 |
| Over 90% | 94% | 93% | 9,699 |
My model's start probabilities against what happened, every player-fixture in the 2023/24 to 2025/26 walk-forward backtest.
GW6
| Player | Owned | Fixture | Team xG | Chance of starting | xP |
|---|---|---|---|---|---|
| Haaland | 74% | @ LIV | 1.57 | 96% | 5.28 |
| João Pedro | 65% | v BOU | 2.10 | 61% | 3.48 |
| Rogers | 41% | v BOU | 2.10 | 96% | 4.68 |
| B.Fernandes | 38% | v TOT | 2.10 | 96% | 5.53 |
| Szoboszlai | 32% | v MCI | 1.54 | 93% | 4.20 |
The five most owned attackers this week, the shortlist the top-3 and top-5 rules choose from.
| Rule | Pick | Fixture | Team xG | Chance of starting | xP |
|---|---|---|---|---|---|
| Template | Haaland | @ LIV | 1.57 | 96% | 5.28 |
| Top-3 owned, most team goals | João Pedro | v BOU | 2.10 | 61% | 3.48 |
| Top-5 owned, most team goals | João Pedro | v BOU | 2.10 | 61% | 3.48 |
| Top-3 owned, most likely to win | João Pedro | v BOU | 2.10 | 61% | 3.48 |
| My model, any attacker | Mbeumo | v TOT | 2.10 | 94% | 5.91 |
| My model, top-3 owned | Haaland | @ LIV | 1.57 | 96% | 5.28 |
GW6, from my model's run of 8 October 2026. Team expected goals here are the model's, which are calibrated to the odds.
So the rule moves the armband off Haaland, away at Anfield with City at about 39% to win, onto Pedro, who my model only has at 61% to start. The rule only picks him over Rogers because he's more owned - same Chelsea attack, same 2.10 expected goals - and Rogers is 96% to start. If Chelsea's presser makes Pedro look less nailed, that's the obvious switch, and Bruno is basically level on team goals at home to Spurs (2.10) too. In fairness to Haaland, in games where City were under 50% to win he's averaged 4.9 points and hauled twice in 15, and my model gives him 14% for 10+ this week.
Genuine question for anyone who's watched more Chelsea than I have - is Pedro actually nailed right now or is 61% about right? If he starts and blanks I'm sure I'll hear about it.
Ten seasons of fixture-level player data (2016-17 to 2025-26) from the FPL API and the vaastav archive; match odds from football-data.co.uk. Every number and table is reproduced by scripts/captaincy_rules.py in the project repo.