Our team of Hall of Famers and guest writers are offering regular contributions throughout the 2023/24 Fantasy Premier League (FPL) campaign. Here, former champion Simon March looks at whether or not we can really predict future returns from a negative expected goals (xG) delta.
Among FPL managers, a negative xG delta (the difference between the number of goals a player has scored and the number of goals their xG statistic suggests they should have scored) is often invoked as a sort of statistical manifestation of the age-old footballing concept of a player being ‘due a goal’.
The theory behind it is simple: if a player has scored fewer goals than expected, we should therefore expect him to overperform in the future in order to compensate for this underperformance.
But how reliable is this theory when applied in practice? Can we really use a negative xG delta to predict the future? These questions will be the focus of this week’s article.
“Negative xG delta, what’s the deal with that?”
In an episode of the sitcom Seinfeld, titled ‘The Opposite’, Jerry reaches the conclusion that everything somehow ends up even for him. For example, he has a gig cancelled, but then he gets another booking five minutes later. He then breaks even playing poker and so on, always ending up even. Elaine tests the theory by throwing a $20 note out of the window and, sure enough, Jerry immediately finds a forgotten $20 note in his pocket. Consequently, Jerry stops worrying about bad things happening to him, like his girlfriend breaking up with him, because he expects things to just even out for him.
Not only does this episode provide a humorous illustration of the statistical phenomenon of ‘regression toward the mean’ but it also demonstrates two cognitive biases that are also useful to this discussion: the gambler’s fallacy and belief in the law of small numbers.
A gambler’s fallacy is essentially the fallacious belief that, if one random event occurs, it makes the opposite random event more likely. For example, a person might believe that, if they flip a coin and get heads, they are more likely to get tails on their next flip.
Belief in the law of small numbers, is the also fallacious belief that small samples will be as representative of the larger population from which they are drawn as large samples would be.
So, for example, if you flipped a coin the chances of getting heads or tails is, of course, even. Thus you might reasonably expect that, if you flipped a coin 10 times, you would get five heads and five tails. In practice, however, it could take you many more flips before you achieve a perfectly even split.
In short, luck evens out over enough time or enough attempts, but the time or attempts involved in getting there tend to be greater than we anticipate.
What does this mean for xG deltas?
When a player underperforms their xG, we rightfully attribute a degree of this to misfortune. We might reasonably expect that better luck will eventually compensate for this misfortune.
The problems start, however, when we apply this logic over short periods say, for example, a player’s past three Gameweeks, and assume that we will start to see their luck even out over their next three Gameweeks.
Firstly, the initial sample of three Gameweeks is too small to produce any sort of reliable view of a player’s xG or their performance against it. The potential for outliers to skew the data here is very high.
Secondly, even if we accept that a player’s goal output will, over time, more closely resemble their xG (and this, of course, is a whole other debate), the likelihood that we will see this in a period as short as three Gameweeks is very low. Performance versus xG could take many matches, perhaps even several seasons to normalise, if indeed it ever does.
Finally, while we can expect the luck aspect of goals versus xG to even out over time, there remains the skill aspect of scoring goals. Because xG assumes average ability among players themselves, and because it doesn’t distinguish between player styles, there’s no guarantee that the player in question will ever live up to their xG.
Even if a team creates lots of high xG chances for a player, it doesn’t mean that they are the right type of chances for this particular player, or that this player is equipped to convert them at an equivalent rate, however many opportunities he is given.
Uses for negative xG deltas
While its predictive value might therefore be limited, negative xG deltas are not entirely without value for FPL managers.
Firstly, the volume of xG in the equation can be quite useful indicator for predicting future goal output, particularly if it is very high. While a player might continue to underperform their xG, the conversion rate might not matter much from an FPL perspective, as long as the xG is high enough.
For example, the current biggest negative xG delta for this season is the -5.32 of Dominic Calvert-Lewin. If this belonged to a player with a total xG of 20, they’d have 15 goals and, depending on their price, we may therefore not care what their xG delta was.
Players who get a lot of chances can afford to be a bit profligate, from an FPL perspective at least. The question in this scenario becomes whether this volume of xG is sustainable and thus, as is often the case, it benefits us to look at the team first, and then the player.
Secondly, when it does come to considering players, a negative xG delta could have some predictive value in circumstances where luck’s role in causing it was very significant. For example, if a striker has faced a succession of goalkeepers who have pulled off a string of worldie saves, you could reasonably argue that this player is, in fact, due better luck in front of goal.
This, of course, is still a subjective judgement and it is also a difficult one to measure at scale so, perhaps, a more practical way to estimate how noisy luck is in a negative xG delta might be to consider how well the player has performed against their xG historically.
If, for example, they have a long history of outperforming it while scoring a decent volume of goals, it is more reasonable to assume that they’ve been unfortunate this time around and that this will, eventually, be compensated for.
‘Eventually’, however, is the key word here and, again, we are unlikely to see long or short periods of historical xG underperformance even out over short periods of future Gameweeks.
When we find a player underperforming their xG, we need to decide first, whether this is really important in the wider interests of scoring FPL points and, secondly, we need to look at the broader context. In particular, we need to consider the role luck has played in forming the xG delta and the ability of the player in question to finish the chances he’s getting.
Only then can we determine how useful an xG delta statistic is when it comes to informing our future decision-making.
