Methodology

The model establishes a baseline for a player’s style of play and also get an idea of the variation around that baseline the player exhibits. The baseline value and variation is calculated for each variable. So, for instance, passing accuracy, shots per 90, and successful dribbles will each have their own baseline value and variation, as will every variable.

Here is a classy 2X2 matrix to explain the utility of the baseline value and the variation.

	HighBaselineValue	LowBaselineValue
HighVariation	Strengths of this player Not critical to his output	Something the player can do Not critical to his output
LowVariation	Strengths of this player Critical to his output	Something the player doesn’t do

Comparison

We now compare each variable between the PoI and all other players who play in similar roles. The comparison is made with the PoI’s baseline and weighed by the variation. High baseline value, low variation variables dictate a majority of the comparison and low baseline value, high variation variables have almost no effect on the comparison.

Shortlisting

A shortlist is drawn based on the worst difference amongst the variables from the PoI. This is so that we find players that don’t impose a significant compromise on any of the individual variables. For eg. let’s take a player who has lower passing accuracy, lower shots per 90, and lower successful dribbles compared to the PoI, but higher goals per 90. Based on the baseline and the variation if lower passing accuracy is the worst compared to the PoI then the other three variables don’t affect the shortlisting.

Order of suitability

The order of suitability is then decided based on the total shortcomings of the shortlisted players. In this step, the lower shots per 90 and lower successful dribbles also affect the ranking while the higher goals per 90 is still not incorporated. This is again keeping in mind that we want to find players that offer the least compromise on the skills of the PoI. Any variable where the comparison with the PoI is in favour of the other player is a bonus buth expectaion from the player being compared is just to meet the PoI’s baseline.

FAQ

Why are other players above the PoI? They are better

More details about the methodology

It's June 2022 and I get paid to do this stuff now so if it helps you, here are some more details about the model I had. It's been too long and this is as much as I can recollect without going through code.

Split the data by each player and any additional qualifiers like season, role, etc. so long as you don't spread your data too thin. Each split is basically one of the entities which you want to compare to another entitiy to find similarities. Let's say you want to find players similar to Messi's 2019 performance so maybe you split your data by player-season combinations amongst which one of the splits would be Messi-2019.

From each split, take random samples with replacement of, say, 5 matches. This gives you a distribution from which you can calculate a mean/median, and a variance in the value of each stat. Think of these values as what you could expect to see if you were to randomly observe 5 matches they played in, sort of like a scout maybe. We are in central limit theorem territory so we can retain the mean, median, and variance and discard the rest of the data.

As an explanation of the philosophy behind the model, let us club the values we see for a particular feature into four categories - high mean, high coeff of variance ( std dev / mean ), low mean, high coeff of variance, high mean, low coeff of variance, low mean, low coeff of variance. If we were to divide the mean and std dev values you see for non-penalty goals per 90 for all the players amongst these four groups, you could interpret it as follows - players with high mean - low coeff of variance do well at this feature - such as strikers with high consistent output. Players with HM-HCV are the ones probably good at it but are either inconsistent or the team doesn't need them to be consistent at - maybe a midfielder who sometimes has more attacking responsibilities and scores often at those times but not at other times. Players with LM-LCV and LM-HCV may or may not have the skill but don't need to be performing on that feature at all - maybe centre backs.

For each feature, we now compute the Z score for all other entities based on the feature means and std devs of Messi-2019. So if Messi in 2019 scored 0.6 goals per 90 with a variance of 0.25, and another player-season scored 0.55 goals per 90, the second player would have a Z score on this feature of ( 0.55 - 0.6 ) / 0.25 = -0.2. We can similarly calculate a Z score for all the other features for all other player-seasons.

We now have a numerical representation of each player wrt to Messi-2019. We can now do a bunch of things rank and filter them. We could rank them by the number of features a player has a negative Z score on, or the average Z score across all features, or something more nuanced such as the average Z score but only onthe subset of feature that Messi-2019 had a relatively low CV on.

As with most algorithms, your choice of splits, your choice of features, your choice of ranking logic, and many other choices that you're making will eventually decide the quality of your results. What works for one player may not necessarily work for another. Given how long it has been since I used this model, you're probably better off experimenting yourself and finding what works for you. Also don't get too hung up on the ranks - you'll have more suitable players towards the top of the rankings and less suitable players towards the bottom but I wouldn't fuss too much about rank 1 being much better than rank 5.

For an idea of the results you can expect, some other places I used this:

• A Twitter thread with a bunch of results from back in summer 2020 - https://twitter.com/thecomeonman/status/1306654311661760512 with, I think, some default settings.
• A post about finding replacements for Messi and Suarez from summer 2020 - https://github.com/thecomeonman/Replacing-Suarez-and-Messi. This is a more agressive way of using this model where I run it multiple times over different seasons and picking those who appear towards the top more often.