MLS Player Elo: The How and the Why
I’ve been interested in American Soccer Analysis’ G+ model for a while now, using it to dig around in player value. I’ve worked at isolating single game G+ values and finding ways to present the pieces as they make up the whole in interesting and informative ways. Now, I’m extending the use of the single game G+ measures into an Elo model that should help capture new information about players at moments in time, in a way that hasn’t been done much before now.
How it all works:
Traditionally, Elo rankings are used to compare two like competitors in direct competition with each other, otherwise known as pairwise comparisons. The classic use is for chess — two chess players face off, and the Elo ranking is plugged into a formula that determines a likelihood of winning for each player. The higher ranked player is more likely to win, according to the model, but never 100% likely. That means regardless of outcome, there is movement in the ranking.
Elo has been used in soccer for a while, from World Football Elo rankings to 538’s Global Club Soccer rankings. Elo rank changes are assessed according to results by game, though the latter example takes away from direct comparison and treats attacking and defending as separate entities with separate goals (in this case maximizing xG and minimizing xGA).
This is another extension, moving even further away from direct comparison. On one side, there are the players, playing at one of 6 specific positions. On the other, there are the teams, specifically against one of the 6 positions. The matchup is then Player A (as a striker) vs Team B (against strikers). The outcome is determined by how a player performs in a game according to G+ vs expectation. The expectation is determined both by the position, and whether or not the game is at home or away. Because of the defense vs offense nature of the calculation, the Interrupting subdomain of G+ is omitted.
Here’s how it works on an individual basis. (Note — this was created before a major model update, so this exact situation no longer exists, but the concept is the same.)
The data set starts at 2013, where all players and teams were assigned an Elo score of 1200 at each position. As a new player arrives in the league (or an expansion team joins), they are given scores of 1200 as well. That does mean that there is a period where the model is adjusting to the new player, and it does take time to stabilize. That makes it less useful for players without a lot of games at a position, or just in general.
Why this approach?
It’s important to establish the reasoning behind this approach; G+ is already a robust measure that effectively captures performance across a number of domains. What does this add that isn’t there already?
- The first major contribution of the Elo calculation is that it quantifies the relative strength of opposing defenses. While traditionally G+ does not have any adjustment for the strength of the opponent, the Elo score relies on the opponent strength as the basis for all calculations. At a glance, it can inform the viewer about how strong or weak a specific opponent is against a certain position.
- The second major benefit is that it provides temporal context to a player’s performance. When charted, it is easy to see how a player goes on temporary runs of positive and negative form. While something like a rolling 5 game G+ average could capture something similar, combined with the opponent measure it provides an excellent look at current form.
Understanding the Elo numbers
Every player starts the Elo ranking at 1200. Over the course of time, a player’s performances will move them up or down the Elo ladder. The further up the ladder a player moves, the greater the performance needed to gain additional points.
The full dataset contains roughly 94,000 individual player performances by 1907 players since 2013. Over the course of seasons, there quickly becomes important landmarks to understand player quality.
The Best in the League:
The 1500 mark has only been reached by one player — Zlatan Ibrahimovic. He moved past the mark in a 2–1 playoff victory over Minnesota United. After the loss to LAFC which eliminated the Galaxy from the playoffs, he settled at 1501 before returning to Europe. Zlatan made headlines recently for claiming to be the best player in the history of MLS; according to this measure, he is exactly that.
The 1400 Club:
The next threshold is naturally the 1400s. Just 40 players have reached this mark; the youngest was Eduard Atuesta but the quickest was Anton Tinnerholm, who only took 18 games to reach 1400. On average, it took players 76 games to reach this mark.
Sebastian Blanco is the only player to hold a 1400 Elo rating in two positions at the same time, both as a winger and as an attacking midfielder. He has also had the greatest number of games in the 1400 level, but has never been a contender in MVP voting.
As of current date of publishing, 5/24/2022, there are just 4 players with a current rating of 1400; Adam Buksa, Keaton Parks, Maxi Moralez, and Blanco. With a current Elo rating of 1427, Keaton Parks is the best player in MLS at the current moment. Carles Gil was above 1400 at the beginning of the month, but has since dropped.
Just a handful of the 40 players in the 1400 club are Americans. The aforementioned Parks is one, and he is joined by Michael Bradley, Matt Besler, Dax McCarty, Graham Zusi, Ryan Hollingshead, and Landon Donovan. Clint Dempsey is the closest American striker to 1400, topping out at 1396.
General Observations:
After playing around with the results, some trends stand out. These may confirm things we already know, or bring up new questions for more in depth study.
The first is that in general, teams are getting better at defending. The median Elo across the entire sample for MLS teams against opposition was 1192, suggesting that in general defenses perform worse than expected, but limited to just the last two years, the median Elo was 1204. This is despite the continued increase in spending on players, defenses are gaining ground.
Another observation is that players with particularly low Elo ratings don’t stay on their teams for long, typically being removed from the league or traded after continued underperformance. This suggests that the measure is capturing something that front offices have identified as valuable.
Something else that jumps out with visualization is how regularly players go through spells of particularly good or bad form. Even a player who is normally good can go through a 3 or 4 game stretch where they underperform expectations and the ratings drop sharply. Youth players especially are prone to swings.
The Data:
All data for this project can be found in the following interactive vizzes
Go here for the full Elo viz
Go here for Player Elo separated by teams (easier to see what’s going on)
More vizzes are coming!
