Sometimes it’s good to zig when everybody expects you to zag. This is true in the NFL: If a coach calls for a pass when his opponent is expecting a run, he could be in store for a big play. But which head coaches are the biggest play calling contrarians?
I trained a set of machine-learning classifiers1 to estimate the probability of a coach calling a pass, rush, punt, or field goal, on every play in the NFL from 2000-2013, based on several parameters: down, distance, field position, time remaining, score, and how many passes or rushes were called in the game prior to that play. This is similar to what @FantasyDouche did to create his Play Calling Index.
This gave me a baseline of what to expect an average coach to call in a given situation defined by these parameters. Then, I compared the predictions from my classifiers with what actually happened. For the 2013 season, the predictions were accurate 84.2 percent of the time across all plays, 68.8 percent on passes and rushes.
Coaching Contrarian Rate (CCR)
Importantly, the accuracy of the predictions varied across coaches. Because our predictions represent what an average coach would do in a given situation, this variability represents how often a coach deviated from the norm in terms of play calling. Since statistics always seem to need snazzy names, let’s call this the Coaching Contrarian Rate.2Stats with snazzy names need snazzy acronyms too, so we’ll call it CCR from here on out.3
CCR values can theoretically range from 0-1; lower values mean the coach rarely deviates from the play calling norm in a given situation, and higher values mean that the coach goes against the grain more often. In the figure below, I plotted the CCRs for all head coaches in the NFL from the 2013 season, across all play types and game situations.
Coaches with reputations as innovators or offensive gurus, like Chip Kelly and Marc Trestman, tended to have higher CCRs, and coaches with reputations as conservative play callers or who are more defensive minded, like Mike Smith and Wade Phillips, tended to have lower CCRs. This is what we’d expect.
Expected Passing Situations
CCR can also be calculated for specific situations. In the next figure, I plotted the CCRs for plays that the classifier predicted would be passes (i.e. when an average coach would call a pass), for all of the head coaches in the NFL from the 2013 season.
Pete Carroll, Jim Harbaugh, and Chip Kelly had the highest CCRs on expected passes in 2013, deviating from the norm more than 15 percent of the time. Given that the Eagles, 49ers, and Seahawks finished first, third, and fourth in the NFL in yards rushing last year, this is not surprising.
Expected Field Goal Situations
In this last figure, I plotted the CCRs for expected field goal attempts. Some coaches, most notably Bill Belichick, have earned notoriety for electing to eschew field goal attempts in favor of going for it on fourth down.
The classifier did a really good job of predicting when a coach would attempt a field goal: In 2013, it was 99.1 percent accurate in those situations. So it isn’t surprising that the range of CCRs encompasses less than one percent. Still, some differences between coaches emerged. It looks like Bellichick has earned his reputation as a coach who goes against the grain in situations where an average coach would attempt a field goal. I was surprised to see Kelly near the bottom of the list, but his sample size is quite small, and again, the difference between Kelly and the coach with the highest CCR in this situation, John Fox, was less than 0.9 percent.
CCR certainly isn’t perfect. The baseline predictions have room for improvement, particularly on called rushes and passes. There are many additional game- and season-level predictors that could be used to improve the classifiers, too. For example, coaches probably adjust their play calls based on the success and failure of previous plays. They may also be influenced by more subtle factors, like temperature, wind speed, playing surface, etc.
It is also important to keep in mind that head coaches don’t always directly call every (or any) individual play in a given game, though many of them do. However, the head coach is still the person who bears the ultimate responsibility for the outcome of each play, and he is also the person who has final say on which plays are run during practice and which plays are included in the weekly game plan. Even if the head coach isn’t directly instructing his quarterback to drop back on a particular play, he is still playing a major role in the QB’s decision to drop back.
CCR is a useful statistic for measuring a coach’s tendency to go against the grain with his play calling. Anyone interested in predicting the outcomes of NFL games could benefit from CCR, or something like it: Professional bettors may choose to avoid betting on teams with coaches who are less predictable; fantasy football owners may choose to draft wide receivers whose coaches tend to pass more often than average, etc. It isn’t clear at this point if CCR is a good predictor of a coach’s success in the NFL, but it certainly could be, and it would be worth investigating further.
- Specifically, the decision tree algorithms available in the JMP software package; hit me up in the comments or @jimkloet for more details. (back)
- Not to be confused with the @ff_contrarian Sean Siegele, though you should check out his work here at RotoViz and on his own site. (back)
- Not to be confused with this CCR; don’t sue me, John Fogerty. (back)