Apologies for missing last week's newsletter. After my Wednesday became incredibly busy and I ran out of time to work on PMR, it really was my intention to get the newsletter out on Thursday instead. Then, of course, as things do, my Thursday got just as busy and that was the end of it. Fear not! I am back this week with another edition to keep the newsletter going. Can you believe it's nearly been two months since I started this? (And nearly two months since the World Cup ended?) Nor can I. Anyway, onto the updates.

The rundown

(in no particular order)

  1. 1.

    In the 10th minute of Bay FC's match against Kansas City Current this past weekend, Izzy Rodriguez scored a textbook Olimpico, curling the ball up and over the keeper's head and slotting it into the far side netting. Opta rated it as a 0.01 xG chance, while ASA's model returned 0.94 xG for the same shot. Why the difference? The discrepancy comes down to what happens to an xG model after it's constructed. Most models these days will use an XGBoost predictor, since it tends to have the highest accuracy and is pretty simple to spin up. But often, the conditions of the data going into the model and the manual adjustments to these predictions is more important than the method of prediction itself. Rodriguez's olimpico reminded me of a section of the StatsBomb xG whitepaper where they talk about observed goal rate recalibration on outlier shots and how it creates more stable predictions that are in line with our intuitions.

  1. 2.

    Brian McKay wrote an interesting piece in Sounder at Heart both analyzing the Dejan Joveljic to Seattle trade and g+ itself. It's a very good piece that perfectly illustrates the uses and limitations of g+, especially when analyzing strikers. One part I'm interested to learn more about is team impact on player g+. Is there a measurable effect of a team's quality on the value of it's players? Does that persist across bad teams and across seasons? Is there a difference between positions, and at what point do we say that the player cause the team to be bad, not the other way around? I think there's a lot to be dug into here, and we've hardly scratched the surface1.

  2. 3.

    I spent a few days analyzing financial data from the MLS summer transfer window for The Guardian and found some fascinating things. First, though, I think we need to talk about how hard it is to find this data in the first place. MLS and NWSL are notoriously terrible about sharing and are afraid of giving away some competitive advantage if their books were to be made public. For that reason, accurate record-keeping relies on dedicated individuals (read: sickos) who thanklessly keep track of these things based on what information is public. A lot of the data I used for the article came from GAMisfun, a website that displays data collected by Arman Kafai as well as Marc Machado, who helped in the early days of the site. Without the site, I would not have had accurate GAM or Cash for Player totals. The second part of the data came from transfermarkt, and the cleaning, joining, and aggregation of all this data took a while. My ability to know where to find this data would not have been possible without the groups of people keeping track of these things, and to them I am indebted2.

  3. 4.

    SCOUTED, in partnership with SkillCorner, recently released a piece on the growing quantity of talent emerging from South Africa, and in it, Jake Entwistle defines five categories of physicality and ranks every league based on them. The categories are Coverage, Agility, Speed, Explosiveness, and Intensity, and thoughI'd love to learn more about how they were calculated, I find the results very interesting. Perhaps unsurprisingly, the Premier League is ranked first "S" rankings in every category except for coverage. MLS has "A"s in every category except for coverage and intensity, where they are assigned "B"s. Funnily, Liga MX rank near the bottom of the list, with ratings of D, D, C, B, and C, respectively. In fact, most of the Latin American leagues live in the bottom half of the rankings. I think these types of grades are quite interesting, especially when they are purely stylistic and don't try to measure performance, too.

  4. 5.

    The Reep Register, an API that aims to link data providers using universal IDs, signed a partnership with SkillCorner that integrates the data provider into the ID mapping service. Other attempts at this kind of service have been made, most notably Glass Onion, an open-source tool constructed by Akshay Easwaran and U.S. Soccer. It's cool to see more of these projects entering the public sphere, now I just wish there were more data in the public sphere too.

A closer look: model considerations

Over in the Futi discord, John Muller sent a link to a KU Leuven research paper that aims to answer questions regarding how certain look-ahead periods in possession value models affect the output of those models. A look-ahead period refers to how far out each model "looks" to see if the current possession ends in a goal ā€” this is the dependent variable in the equation. Functionally, what each possession value model is predicting is the probability that the current possession will end in a goal, and the look-ahead window helps us define that. First, the researchers broke down each type of possession value model based on its look-ahead period. g+, for example, uses the extended possession-based framework, while VAEP is action based, and some EPV models are time-based. This chart, which shows the mean number of actions per framework as well as the respective share of game states in each bucket, is pretty cool:

There's a ton of fascinating stuff in this paper, and one of the main takeaways is just how massive this decision is when building a model. The differences between frameworks are like canyons, with players highly rated by one model and entirely disregarded by another, simply based on the number of actions. The chart below seems to show something interesting, which is that while action-based windows seem to preferable for single (attacking) possession value models, dual possession models have less extreme distributions when using an extended possession-based window.

That being said, when the researchers analyzed the possession value contributions by action type, they found that passes tend to be overall negative in the extended possession-based window, while shots are much more highly rated. In the action-based window, shots, passes, and other actions are approximately evenly weighted.

From my own experience, I find this result to be very interesting. Having explored quite a few of these models, I found Atomic VAEP to be severely lacking and significantly over-valuing shooting and, consequently, forwards, when compared to g+. My main inkling for why this may have been happening is that Atomic VAEP removes the outcome of every action and rates them separately from the actions themselves. This means that any value accrued by a high-quality pass is attributed to the reception rather than the pass itself, since most high-value passes are also risky ones. Likewise, shooting becomes far more valued, since taking a chance at goal will always be better than not taking a chance at goal ā€” xG far exceeds the Atomic VAEP value of retaining possession.

This brings me to the same conclusion the paper reaches: look-ahead window is but one of many considerations when building a possession value model. Each of the factors that goes into the model must be analyzed with the same level of rigor and interrogation that is presented in this paper. Only then will we begin to understand how each minute action that is taken in the process of building our model actually impacts the output. That is to say, what is the expected value of picking an action based look-ahead window? What is the value of separating out outcome from action? The first person to make a possession value model for decision-making in the building of possession value models will have solved soccer analytics. (jk.... unless šŸ‘€)