Fantrax player points forecasting tools: xG / xA
Training your critical eye and using xFpts as a data to judge player value
We continue our tour of Player Forecasting methods here with a discussion xG / xA. If you missed the previous article in the series, we talked about age and the impact of age on draft decisions in a previous article. You can find that here.
In a few of my articles, I’ve made reference to xG, npxG, xA and, most importantly, xFpts. So it’s probably about time I talked about it a bit more… And this one’s a bit more complicated than the last one. In it’s simplest form, xFpts (as explained and made available by the excellent @overthnkfootball which I’ve linked below) looks to account for the difference in what did happen in an event (a historical event reported, in this case whether a shot went in or a pass led to an assist) and what we would expect to happen as a result of that shot or pass based on analysing event data across every event of that type (useful as a forecast).
Watching fans are often not good at judging the value of a shot, whether we want to admit it or not. For example, we often hear a one v one chance spoken about in binary terms “the striker has to score that”. Or people judge what are perceived as “sure thing” chances to score as if they’re a certainty, sorry Raheem but I’m going there again, when they’re not. They probably go in most of the time, in Sterling’s case probably above 85% of the time, but they’re not a dead certainty. xG is not a perfect data point, but it is a data point that helps us to fix that flaw in how we process events that have occurred. In an xG world, “The striker has to score that” might become “the average striker had a 30% chance of putting that away”. The point is: precision matters (though I admit it isn’t great for pub chat). It is very important for Fantrax though.
Stats companies such as Statsbomb have driven major advances in this field since the idea was first floated and Statsbomb now supply the excellent FBRef website, which is a great resource for would-be football analysts and Fantrax managers. xG as a metric also led to a lot of discussion around “finishing skill” and repeatability of xG over-performance, given that it measures what the “average” striker would do. As the data has improved and the amount of variables used to forecast has grown, it seems clear enough that there are examples of players who can over-perform xG on a consistent basis. And this makes sense, super-stars should probably be performing better than the average player. But we should be careful applying a blanket rule, as there is not a totally one-to-one relationship between players that finish the season with the most goals over an extended period of time, as we might expect there to be.
Case in point: Robert Lewandowski’s last three seasons have seen goals returns of 29, 22 and 34. And yet, excluding penalties, his xG says that he actually scored 2.6 fewer goals than he should have over that period. Similarly, despite 78 goals in the past three years, Cristiano Ronaldo only has a 1.9 over-performance of npxG. What both of these players have in common, however, is incredibly good shot generating skill. Lewandowski takes 4.5 shots per 90 and generates 0.2 npxG per shot. Ronaldo averages a quite frankly obscene 6.09 shots per 90 with a 0.11 npxG per shot rating. So it’s not hard to see three things:
These players are doing something repeatable and consistent that get shots off at a high rate per game.
The volume, coupled with the shot selection quality mean that they’re scoring at rates close to a goal a game.
They can therefore score an elite number of goals, despite not necessarily having to be elite “finishers” (yes they are both still good finishers)
So am I saying that elite finishers don’t exist if those two aren’t examples of it? Nope. If you’ve got the time, go look up the npxG performance of this man…
Anyway, as a Fantrax manager what we want to try to understand is:
Who over / under-performed their xG / xA numbers last season and what was the impact on their total points tally as a result?
Do we have any data to suggest that they are a “plus-finisher” and that they can continue to over-perform xG over a long period?
And lastly, if the under / over-performance regresses to league average, what does it do to the player’s value?
(Note: xA is harder to assess than xG because Fantrax and other Fantasy games use a different definition of assist than most stats tracking websites, such as FBRef. As such, we’ll see big gaps from xA on FBRef to Assists on Fantrax. I’ve typically adjusted for this by pulling the assist values that FBRef list and comparing xA to that, then adding back on the difference between Fantrax assists and FBRef assists - not scientific, but the only fair way to look at the data that I could think of)
Let’s look at Pierre-Emerick Aubameyang, as a case study. Potentially the red half of North London’s favourite striker since a certain Frenchman left town (sorry Olivier, I meant Thierry), a regression to what his npxG numbers suggest for him (14.2 goals) would cost Aubameyang 69.3 points, moving him from the 15th best player last year in points scored to the 25th. For a player that currently has an ADP on the turn of Rounds 1 and 2, this is something we really need to have a view on, if we’re even thinking of buying in at this price. Especially as he’s pretty goal-reliant when it comes to his points:
So do we have reason to worry about Auba? Well looking at his time in the Premier League with Arsenal, we see a 9.2 npxG total over-performance, across three seasons, which is a pretty impressive number (though this year was the biggest over-performance so far, which is notable, with 5.8 of the 9 coming last season). He’s taking 2.7 shots per 90 and averaging 0.2 xG per shot. These numbers are pretty good, though they’re not Ronaldo or peak-Harry Kane level. Importantly, when making this assessment, his over-performance is not so extreme that it’s entirely out of the realms of rational explanation. We should probably expect it to drop a bit closer to the average of what he’s done over the last three years, but there’s every reason to expect him to be able to remain at least a goal or two ahead of his npxG, based on the numbers we’ve looked at here.
On our last test, as we saw from @drafterthoughts tweet above, Aubameyang is mostly reliant on goals for his points. So if he did lose some of the goals, it’s unlikely that his ghost points would make up for the difference. Overall though, when trying to forecast for next year’s performance, Auba looks like he’s got every chance to continue to put up good npxG numbers, and probably to outperform them. If I was making a prediction based on his historical values, I would probably consider something in the region of a 2-3 goal over-performance to be a relatively solid forecast and that still makes him very draftable. If the bump in performance against npxG is something that is a result of Arteta’s system or something in how Aubameyang has been coached, you might even see more than that. Considering all of this, I don’t knock Auba’s value too highly, based on xG but, at the same time, I’m aware that I might be paying a price for a value that he might not be able to necessarily deliver in full.
Before we conclude, it’s worth saying that there were a few other names that it’s worth taking a look at yourself with regard to xG performance and what you expect to happen next season, including:
Gabriel Jesus
Anthony Martial
Danny Ings
Roberto Firmino
Harvey Barnes
Summarising: when you’re looking at 2019/20 points to try to forecast what you should expect from a player in 2020/21, you need to be looking at a lot of data points. @overthnkfootball has provided an incredible resource to critically analyse the total points a player has scored and to make your own judgements based on the data whether you think the performance last year is repeatable (and there are tonnes of factors you could also choose to apply on top of what I have looked at above: team style, manager changes, age etc.). A little time looking at this kind of info can help you to further minimise your risk of draft busts, find players with depressed value, or to find hidden gems (though beware very small sample sizes, or you’ll end up taking Luke Thomas in the 1st). Most importantly, it’ll help you assess real player value and could put you at a real advantage in your league.
(Note: I don’t talk about penalties in this article. Penalties are viewed as chances that massively favour the taker with estimates of somewhere between 75-80% of all penalties taken going in. For the purposes of this article, I’m just interested in npxG.)