Nylon Calculus: Game Control and a new take on team metrics

Back on Nov. 7, 2017, the Nets and Nuggets played a game in Denver. Though the Nets kept it close through the middle of the second quarter, the Nuggets soon pulled away. Denver held a 13-point lead at halftime which swelled to 25 by the middle of the third quarter. By the fourth quarter the Nugget had the game comfortably in hand, took their foot off the gas and ultimately won by a modest eight points, a margin which did not accurately represent the blowout nature of the game.

Most NBA fans know that point differential and Net Rating are extremely useful metrics and are better predictors of a team’s success than win percentage. Typically, when a team whose point differential is drastically different than what their record would suggest, it indicates that team had some good or bad luck in close games and is due for a correction in the win or loss column. Early on in the season, we can usually pinpoint teams who are likely to improve or regress due to their point differential offering underlying information on their performance. However, Net Rating isn’t perfect and there is a fundamental flaw with looking at it as an outcome metric.

Put simply, point differential isn’t the true outcome of a basketball game. Teams play to win the game, not maximize their margin of victory, and at the end of the day a one point victory counts the same in the standings as a blowout. That’s why the Nuggets were content to let the Nets cut into their lead in a meaningless fourth quarter, and why Stephen Curry regularly sits out for the entire last frame. In both of those games, the final score didn’t do justice to how dominating the winner was throughout the course of the game. It works in reverse, too – on Nov. 16, the Wolves beat the Mavericks by 24 even though Dallas was up for most of the first half and the teams were tied early in the final frame. The game wasn’t really that out of hand, the wheels just fell off for the Mavs at the end.

That’s why I’ve been working on a new team metric called Game Control, inspired by the ESPN stat for college football. Game Control is designed to quantify a team’s probability of winning at every point in the game, with the idea being that it’s more in line with a team’s intentions than Net Rating. It’s all based on win probability, and it’s pretty simple in concept. As is appropriate for this website, Game Control is just a team’s win probability* integrated over the course of a game. The score is then adjusted to have a theoretical maximum of 100 (if a team had a win probability of 100% throughout the entire game) and theoretical minimum of -100 (if a team had a win probability of 0% throughout the entire game). In practice, for single games, Game Control is distributed as such:

*To delve a bit into the technical details, I computed a team’s win probability by using a logistic regression on point differential, time remaining, possession and home court status to predict game outcome (it is not informed by team quality). My model is trained on regular season play-by-play data since the 2008 season.

In that November game against the Nets the Nuggets had a Game Control score of 69.4, which falls in the 92nd percentile in my dataset. The biggest blowout this year belongs to the Philadelphia 76ers, who jumped out to a 15-2 lead en route to a 36-point victory over the Pistons on Jan. 5. The 76ers had a Game Control score of 89.7 for that game:

So Game Control does a pretty good job of crediting teams like the Nuggets, who jump out to a lead and maintain it, but it is far from perfect itself. Its major flaw is how it handles comebacks — a team can still have a positive Game Control score in a loss, if they held the lead for most of the game and then lost it. For an extreme example, the Rockets had an impressive score of 68.0 in their December loss to the Celtics, when they blew a 26 point lead:

I also found Game Control to be less predictive of future wins than Net Rating at the team level, so until further research is done this metric can be resigned to “kind of interesting, but not particularly useful” status for now.

Despite its shortcomings, I thought it would be interesting to aggregate these scores by team and see how they stack up. The range for teams is around -35 to 35, and teams with an aggregated total Game Control Score above 20 are generally elite. Houston is the surprising leader in total Game Control thus far in 2018 (as of 01/27/2018), likely due in part to it getting far too much credit for the Boston collapse. Next is Golden State, Toronto, Philly(!) and OKC, with Phoenix and Sacramento bringing up the rear. The below chart shows all total Game Control scores for 2018, compared to Net Rating and win percentage:

Looking back since 2013, the Warriors dominate on top with the highest aggregated Game Control scores while the 2016 Lakers bring up the rear. The 73-9 Warriors are first in this metric despite being second in Net Rating behind the Spurs in 2016:

Although there remains much more work to be done on the concept of using win probability to improve team metrics, Game Control does a decent job of addressing the flaws of Net Rating and I hope to improve it in the future.