Orlando, DC, and MLS' Latest Strategic Fashions / by Kevin Minkus

By Kevin Minkus (@kevinminkus)

The press (whether high, counter, or other) is in vogue in MLS. MLS teams are, on average, they pressiest they’ve ever been. The Red Bulls, NYCFC, Atlanta, and New England all primarily defend in some form of press. A handful of other teams - Sporting Kansas City and LAFC most prominently - go to it on occasion. Orlando City began the season trying to play a higher pressure defense:

 The stat used here - passes allowed in the opponent’s half per defensive action in the opponent’s half - is a good, rough metric for measuring the intensity of a team’s defense higher up the field. The higher the number, the more passes a team allows between tackles and interceptions, so a lower number corresponds to more pressing.

The stat used here - passes allowed in the opponent’s half per defensive action in the opponent’s half - is a good, rough metric for measuring the intensity of a team’s defense higher up the field. The higher the number, the more passes a team allows between tackles and interceptions, so a lower number corresponds to more pressing.

 

DC United look like they will try to finish it that way:

dcu.png

Given this landscape, an ability to play the ball out of pressure, and in transition out of the back, is imperative, especially among defenders. I put together a model to measure that ability.

The model

To attempt to focus specifically on high pressure and transition scenarios, I built the model only on passes attempted within five seconds of a team winning the ball back in its own third, from the 2018 season. This is roughly the first three passes a team attempts after the opposition turns it over.

The first component of the model is a Gaussian process classifier that predicts, given the defending team and starting and ending locations of the pass, the probability of it having been completed.  The goal for this piece was to handle the spatial aspect of the prediction. The output of that classifier is then used as an offset for a logistic regression that includes a flag for if the pass originated within 8 yards of the touchline, a game state variable, and, most importantly, a player-specific coefficient. It’s that coefficient that we’ll interpret as a player’s skill at completing these passes. To easily get uncertainty estimates on our skill coefficients, I built a Bayesian model using PyMC3. The priors are Normal(0, 1).

The results

Here are the top ten players with 25 or more such passes attempted. The value plotted is the probability of the player completing a pass of “average” difficulty, and the bars denote 90% credible intervals:

t10.png

While it shouldn’t be surprising to anyone who has watched much of the Union this season, the otherwise slept-on Alejandro Bedoya shows up at the top of the model. He’s incredibly effective at moving the ball from back to front, and that skill has played a big role in the Union’s success this season. The Union’s 19 year old center back Mark McKenzie is up there with him, in 5th. McKenzie is improving here as the season has progressed - he was 14th in this model built on data six weeks ago.

The Impact, in Saphir Taider and Alejandro Silva, also have two players excellent at playing quickly out of the back. This is a pretty vital skill based on the counter attacking play Remi Garde has employed in Montreal. It makes sense, and speaks positively of their recruitment strategy, that two of their offseason signings show so well.

I’ll call out two other players here - Sacha Kljestan, who is excellent at making every other type of pass, can also apparently hit passes in these scenarios. He mostly plays in an attacking role since moving back to MLS, but he did play as a number eight with Anderlecht. And 18-year old Chris Durkin, who gets lauded for his vision and passing ability, but can’t currently get any run in DC.

Here’s the bottom ten (minimum 25 attempts):

t10_2.png

Brandon Vincent finds himself at the very bottom here. Vincent is naturally a left-back, but he’s been forced into a lot of time at center back this season, and it’s likely his distribution suffers for it.

In general I’m okay with the model, but results here do suggest some of these effects are system-dependent: Alphonso Davies, Marcel de Jong, and Jakob Nerwinski are all wide players on Vancouver, for example. Understanding the difference between a player’s ability and what his team asks of him is an incredibly difficult question to parse in soccer analytics, and it’s something I’m especially interested in with passing models, so look out for more on that.

It does say something interesting of Carl Robinson’s system that three wide Vancouver players look so poor under the model. Davies and Nerwinski especially are frequently used as outlets to start a counterattack, usually by carrying the ball into space. This suggests that, when they don’t have that space and are instead forced to make passes out of pressure, they struggle to complete them.

Here’s every player that made 25 or more qualifying passes. There’s a lot of overlap in the middle in our credible intervals. Really the specific ordering is not so precise - it’s more about the general region a player falls in:

pasted image 0.png