Finding Sports Analytics’ Next Revelation

By Stephen Shea, Ph.D.

November 21, 2017

Sports analytics begins with an investigation into the player or team actions that lead to wins. The quantifications of those activities are called “descriptive stats.”

Understanding what happened in the past is valuable, but we might also want to predict what will happen in the future. And, not all descriptive stats have great predictive power. For example, we might say that team A won last night, in part, because they shot 65% on 3s. It’s unlikely team A will repeat that level of efficiency from behind the arc in their next game, and so, it doesn’t help a great deal in predicting the outcome of team A’s next game.

If instead, we found that team A were constructing lineups with excellent 3-point shooters and was executing an offensive strategy to get their best shooters open looks on the 3-point line, we’d have information that is more predictive of future outcomes. It’s more predictive because it gets at the team’s process as opposed to just the outcomes (shooting percentages), and processes like lineups and offensive schemes are reproducible when shooting percentages can be inconsistent.

Predicting future success serves a purpose beyond informing evening parlays. The predictive stats can serve as blueprints for team success. It wouldn’t do much good to tell Team B to shoot 65% from the 3-point line tonight, but if we can show Team B the ways in which Team A is generating open looks for elite shooters, that’s an activity Team B can work to emulate. Predictive stats can get beyond what happened and attempt to describe how or why it occurred, and it’s with an understanding of the how and why that teams and players improve.

Hockey is a low-scoring game, and as such, desperately needs good measures of process. Andi Duroux recently wrote in an introduction to hockey analytics, “The vast majority of hockey analytics boils down to figuring out how to turn descriptive stats into predictive ones.” A major step in this direction was the realization that—while goals and shooting or save percentages can vary wildly across small samples—a team’s shots for and against can be relatively consistent. Historically, shots (or Corsi, as it’s often called) have been a good indicator of team play. In other words, teams that got more shots tended to score more goals. The predictive nature is why Ari Yanover recently wrote, “Corsi is the public face of hockey’s advanced stats movement.”

There are similar efforts in basketball analytics. We’ve all heard the basketball coach say, “We played well, but the shots didn’t fall.” Teams can use an expected points model to determine if they are getting and giving up the shots they want. (See here for details.)

For good reasons, sports analytics is focused on predictive stats. But, what if the next great revelation hides not in the predictive, but in what could be? Should we study more deeply the activities that drive success and where teams are inconsistent? Can we identify the actions where team performance varies wildly, not because it’s random, but because the team hasn’t prioritized the pursuit or hasn’t figured out how to maintain high levels of production?

We shouldn’t be so quick to dismiss the stats that aren’t reproducible. Inconsistency in the past doesn’t preclude consistency in future.

There is a recent example in the NBA of this strategy paying dividends. The NBA is in the midst of a 3-point revolution. Much of the discussion has centered around teams sorting out the best offensive strategies, but what about the defense? If teams are shooting threes at record rates, certainly an ability to defend them would be beneficial.

In a recent article, Kevin Pelton pointed out that the 2016-17 Celtics “benefited from opponents shooting 33.2 percent from 3-point range, the league’s second-lowest mark.” He then suggested that this statistic tends “to regress heavily to the mean.” He is correct that, historically, teams have lacked consistency in defending the 3-point shot.

Can teams reverse that trend? A closer look at the Celtics reveals an organization that has been strategizing to optimize their 3-point defense for more than a decade (back to the days when Tom Thibodeau had a heavy influence on their defensive schemes). Those efforts are culminating in a roster filled with long positionless defenders and a defensive strategy that involves switching screens on the perimeter, obstructing 3-point attempts, and in a best case, running players off the line and into mid-range.

It wasn’t a fluke that Boston held opponents to a low 3-point percentage in 2016-17. It was the 10th straight season the team was in the top 5 in opponent 3-point percentage. This season, they are third and holding opponents to 32.1%, an even lower efficiency than last season.

(The Warriors have also succeeded in this regard and were in the top 5 in each of the last 4 seasons.)

The holy grail of sports analytics is to arrive at conclusions that change behavior. The next great discovery may require more than weeding out the predictive stats among the descriptive. It may come from quite the opposite, a thorough investigation into what isn’t predictive.

You can read more here about our thoughts on this challenge for hockey, but this is a discussion that’s relevant in all sports.

Leave a comment

1 Comment

  1. Applied Sports Science newsletter – November 23, 2017 |

Leave a Reply

Your email address will not be published.