xP & POE – Serve (2017)

To break down serving, we must first identify the expectations and then figure out how to quantify the results. Some may immediately think that because it’s a closed loop skill, that the expectations are simply 0. We have no expectations, there’s no other force acting on your ability to hit the serve. But this isn’t really true. Much like ranking free throw shooters, the NBA average isn’t 0, it’s about 79%. Meaning that if you make 79% of your free throws, you are average. 79% is the expectation, based on the standard of the league’s performance. The same is true for serving.

So to calculate this for serving expectations, we simply use the receiving team’s FBSO efficiency. FBSO efficiency is just (won points minus lost points) divided by total serves, with the condition that the point is won or lost within two possessions (possession 1 is the serve, possession 2 is the reception, set, and first ball attack). If the play is not terminated on possession 1 or 2, then for our calculations, a 0 is assigned. Missed serves do count here as an FBSO eff of 1.000 as the receiving team wins the rally 100% of the time when the serving team misses. After looking at aces and errors, it’s just the attack efficiency by the receiving team on that first ball. Here’s what that looks like by league:

Screen Shot 2018-10-26 at 8.45.02 AM.pngKeep in mind that these are the expectations for the serving team. They are all negative because you are always at a disadvantage when you serve at this level. The top five conferences are women, the bottom five are men. Using these values as our expectation, our baseline, we can then look at how the results are above or below what we think they should be.

In order to accurately describe the results of the serve, we must choose metrics which correlate highly to the likelihood of winning and losing the point. If we judged serving by where the serve landed or by if the passer played it with her platform or hands, this is certainly a valid method, but it doesn’t really tell us anything about whether the serve helped or hurt the serving team’s chances of winning the point. To make sure we’re looking at the right things, we must create a model to validate the parameters of our results.

After some tinkering, you find that simply using the R (receive) code as coded by VolleyMetrics, is the best way to evaluate the serve. This, in combination with a coefficient for gender, yields the best model.


There’s no need to break results up by conference, just accounting for gender yields an excellent model. At the top, you see the “eff ~ OutputContact + gender” this is the model where we are trying to guess the FBSO eff using the grade of the pass and the gender of the teams. In a practical sense, you start with -0.119 (the intercept) and then if the contact was one of the contacts listed, you add or subtract that coefficient’s estimate. If it was a women’s match, you add 0.0345 to our overall result prediction.

For example, let’s say you’re the Stanford women. You serve into a perfect pass (R#). Well, you start with -0.119 because that’s the intercept, it’s the default value if all other variables are 0. But it’s a perfect pass, so you subtract 0.249 from -0.119. This is -.368. But it’s a women’s team, so you add 0.0345, resulting in a result prediction of -0.334.

This model is excellent, with an adjusted R-squared of 0.9929, meaning that over 99% of the variance in the data is explained with this model. We can be confident that the combination of Receive codes and gender have a very strong correlation to FBSO eff.

The final step is creating Points Over Expectation (POE). This is just looking at the difference between your results and the expectation. Results minus expectation gives us this value, which may often be negative if you are performing worse than average. Here are the 2017 Big Ten and Pac 12 teams ranked by POE in their serving performance on the season overall.


You’ll see the super technical teams of Minnesota (McCutcheon), Washington (K. Cook), and Nebraska (J. Cook) at the top of this list. Minnesota being 0.030 POE is impressive. For context, we know the Big Ten was -.222 in expectation, but this means that Minnesota on average was up at -.190. That may not seem like a big deal, but that’s massively better than everyone else in the league. This is likely because Minnesota is 4.4% service error while being 5.3% service ace. Keep in mind that because they start at -.222, a service error is a POE of -.778 whereas a service ace has a POE of 1.222 since you’re going the other way.

Anyway, there’s your breakdown of xP and POE for Serving. This will be more interesting when I finally get around to snagging some 2018 women’s data from this fall.