EPV (Expected Possession Value) is a real-time prediction of a possession's final outcome given its spatial history. Like a stock ticker for an NBA possession, EPV provides an instantaneous summary of a possession's value based on all available information---the players on the court, their positions, and unique offensive skills.
EPV fills the gap between the coarse, play-by-play and box score data that reduce possessions to one or two numbers, and the dynamic ecosystem of NBA basketball. EPV reacts to every on-court movement and action, allowing us to quantify basketball features and events that impact possession outcomes, but fall through the cracks of other advanced metrics.
For instance, instead of just seeing a Jason Terry 3 pointer with an assist from Andray Blatche, we see that Terry took a shot worth a bit over 1.25 points. This was a smart move, as before shooting, we expected only 0.9 points from this position (it was closer to 1.1, but that was before Ray Allen closed to defend Terry). Still, it took a smart pass (increasing EPV by 0.15) by Blatche to reach Terry, after Livingston did nothing useful (EPV flat at around 0.85) during the first part of the possession.
We've written up further descriptions and examples of EPV, including all the gory details of the modeling and computation. There's even some code available for anyone to generate their own GIFs like the one above (for now, from the same game), as well as many other metrics and visuals.
Moving forward, we hope to
- Develop visualizations that identify key actions and players using EPV.
- Make our EPV model faster to compute, and updated regularly throughout the season.
- Incorporate more features into the EPV, so that we learn more about the impact of defensive matchups, set plays, and off-ball events. More long term, we'd love to have EPV for other free-flowing sports, such as soccer, hockey, and football (mid-play).