What are the best metrics for picking an FPL captain?

The Fantasy Football Scout captain poll is one of the safest bets when it comes to selecting the best candidate for your armband each week – but it is not the absolute best.

After some extensive research, I have discovered that honour falls to the Rate My Team tool, which has outperformed the captain poll over the last three seasons and is the single best predictor of immediate captaincy points.

This article analyses how successful various captaincy selection criteria have been over that time, comparing captaincy poll winners, highest RMT scores, and various individual and team statistics against actual points scored.

Perhaps fixtures are better than form if you know where to look. Player and team attacking stats produce mixed results, with all but one statistic consistently underperforming the three-season dataset average. And with player big chances being the most unreliable of all the categories studies. 

Meanwhile, targeting opposition stats seems to be a more reliable strategy. The single most successful stat to target seems to be opposition shots in the box conceded. 

DISCLAIMER: I am no statistician. There may well be gaps in my approach, and by picking the single best player to represent a particular statistic in a particular week, I have produced a solid but not massive dataset from which to draw conclusions. Any of the criteria discussed, and the results and conclusions drawn, have the possibility to revert towards some sort of mean as more data becomes available in the future. With all this said, I hope that the data, and my conclusions, are interesting, and I will continue to collect data in the future to try to improve the dataset and produce even more reliable results.

As this article uses extensive levels of Opta data, only those with a valid Fantasy Football Scout Membership can access it in full.

THE BACKGROUND

A couple of months ago I decided to run the numbers for the Fantasy Football Scout captaincy poll winners, each week for the season to see how many points the winners got and whether it would have been worth taking a hit for them every week. 

But that got me thinking too: are the results of the captaincy poll the best predictive indicators of Gameweek performance in terms of points gained? Are there other statistics which work better? Player stats? Team stats? xG/xGI? Opposition defensive stats? 

Is there really “one stat to rule them all” when it comes to predicting the best attacking score in a given Gameweek? Taking three seasons’ worth of data, I decided to find out.

THE METHOD

Taking a combination of captaincy articles and members’ data stats, I decided to identify the ‘best’ single player in every Gameweek between 2017/18 and 2019/20 to represent each of a variety of different predictive categories: 

• popularity (winning the captaincy vote) 
• FFS RMT algorithm 
• a range of player attacking stats
• team attacking stats
• opposition defensive stats

In order to mirror the format of David‘s captaincy articles, the dataset used was ‘last four matches’ for player stats. Because I had to reconstruct a lot of the data from the Premium Members Area, which was often not captured in the older captaincy articles, I had to use last four Gameweeks for team stats. This does create a slight imbalance when it comes to the weeks following Blank and Double Gameweeks but I decided that it didn’t make much difference.

Looking at the data, the good and bad teams tended to be top and bottom of their various statistical categories for many weeks on the trot, smoothing out any slight variations. A few spot-checks revealed only the most minor of differences, with little or nothing to affect player/team rankings.

While Scout’s captaincy picks inevitably take home/away fixtures into account, my statistical approach did not consider this factor. This is because the captain articles earlier in each season were unable to make this distinction (with small sample sizes). 

Comparing home and away data is particularly relevant to opposition defensive stats and in future, I will collect this information. I will, in due course, probably do back through the data as far as I can and populate these fields for past seasons. 

However, as we will see, despite a possible bias towards the captaincy poll in the dataset, particularly at the expense of opposition defensive metrics, the results suggest that we ought to target defensive stats even without taking home/away into account. 

To make the results fair, to model what could realistically have been predicted without hindsight, I decided to limit the list of players to those above a certain popularity threshold. This meant limiting my selections to those players who made it into the top 10 in the captaincy poll in any given Gameweek. There are big scores from the most differential of differentials on occasion, but it cannot realistically have been expected that any serious player would have trusted them with the armband, so these less popular players were excluded from the picks. As it happens, this does not matter much for two reasons: 

1. The attacking stats tend to be dominated by premium picks
2. Ranks 7-10 in the captaincy picks tend to scrape the barrel somewhat anyway, usually getting down to the 1% range.

My criteria for selection were as follows:

• Captaincy poll – This was self-selecting
• RMT – The same applies
• Player attacking stats – The player who topped the attacking stats in each category. If that player didn’t make the top 10 in the captain poll I would move onto the next player and so on. If none of the captain contenders were in the top 10 for a given stat I would leave it blank. Being, for example, the 15th best player for touches in the box is hardly comparable with being the best player for shots in the box, and I didn’t want to skew or dilute the data. The categories are: player touches in the box, player shots, player shots in the box, player shots on target, player big chances, player xG, player xGI.
• Team attacking stats – The player ranked highest in the captaincy poll whose team topped various attacking stats. The categories are: team shots, team shots in the box, team shots on target, team big chances, team xG.
• Opposition defending stats – the player ranked highest in the captaincy poll whose opponent had the worst stats in each of the following categories: opposition shots in the box conceded, opposition big chances conceded, opposition xGC. When the captaincy poll started using opposition shots on target conceded I started adding it as well (past two seasons). 

As there are only 19 possible opponents a player could face, I decided to limit the ‘worst’ teams to the worst five in each category. Expanding further would push opponents into mid-table range, and this could actually be quite respectable, so I didn’t want to skew the data by stretching this too far. Again, if no teams in the bottom five for each of these categories had a match-up against a top-10 captaincy candidate I left this datapoint blank. 

Other things to note:

Blank / Double Gameweeks

I did not collect data for these. This is mainly because captaincy polls recognise players with two fixtures, and will tend to zero in on one or two top picks, whereas stats would not necessarily favour Double Gameweek players, skewing the numbers towards captains who play twice. 

Blank Gameweeks limit the pool of players available to the point where, potentially, the averages are brought down, possible unevenly. 

There was more than enough data over three seasons to be able to pull out hundreds of data points without factoring Double and Blank Gameweeks in.

I am aware that these altered rounds cause slight problems when we take subsequent player and team stats which look back over the last four matches. As stated above, I saw enough in the stats to suggest that players/teams tended to top their categories for many weeks in a row, and so discounted this as a serious concern.

Players tied on a certain stat

I always went for the player with the highest captaincy poll rating as a tie-breaker.

Note that I also made no weighting distinctions between players, in that a runaway leader in any of the categories are treated the same as narrow winners. I will try to think of a way of weighting those categories in the future to see how accurate they are at predicting points.

When I had finished I had something which looked like this (this screenshot is an arbitrary portion of the spreadsheet):

Best players by position:

• Goalkeepers: Budget
• Goalkeepers: Premium
• Defenders: Budget
• Defenders: Mid-Price and Premium
• Midfielders: Budget
• Midfielders: Mid-Price
• Midfielders: Premium
• Forwards: Budget
• Forwards: Mid-Price
• Forwards: Premium