Read on Twitter
Last week, @LukeBornn, @dcervone0, and I uploaded to ArXiv the full technical version of our expected possession value (EPV) approach for soccer, using tracking-data.
I start a thread highlighting the most relevant parts of the paper.
https://arxiv.org/abs/2011.09426
I start a thread highlighting the most relevant parts of the paper.
https://arxiv.org/abs/2011.09426
In this paper, we present in full detail on how to estimate each of the components of our "decomposed EPV" proposal. We also show that each of the models produces calibrated estimations. The good news is: now we have a set of useful models linked with overall EPV.
Why is this useful? Now we can inspect specific situations and say: here, the EPV is low because of the pressure on the ball carrier. There the EPV is high because of the space being created by these 3 players. Or: this player should not take this shot and pass back instead
All the pass components are estimated using low-level spatiotemporal data and exploiting our SoccerMap architecture to its best. The result is fully descriptive probability surfaces for better visual inspection. For the rest of the components, we use soccer-specific features.
When analyzing the distribution of the EPVadded for a set of different pass types, we observe some interesting behavior such as the significant risk of crossing, the high value of through balls, that you might earn EPV when losing the ball, or that many times passing back is good
We also explore a series of practical applications for off-ball performance. For example, what would be the best way for defending/pressing buildups against B. Rodgers' Liverpool? Here we show where spaces are created and where they are exploited against different formations.
Another application, "standing on the shoulders of David Silva," explores which players from Man City's 2014 squad were creating more space for Silva and receiving more value from his passes.
Most importantly, this work wouldn't be possible without the remarkable contributions to action-value models and EPV we have been seeing in the last years. With special mention to @dcervone0 (et al.) and @srudd_ok first approaches in basketball and soccer, and so many others.