Offensive PT-PM: The Next Iteration

Oct 29, 2014; Boston, MA, USA; Brooklyn Nets center Kevin Garnett (2) reacts to being called for a technical foul against the Boston Celtics during the fourth quarter of Boston

In the off season I created a beta version of a player metric using SportVU player tracking stats and traditional box score metrics, you can read up on the methodology here, if you’re into that sort of thing. Then I gave the new metric a test run in creating season projections.

This is a new iteration with a little data clean up and a little rethinking of the original inputs on the offensive side, along with my own commentary on the process.  My expectation is that this will be the last interim iteration before locking down the formula, basically writing this up is a way to think through the process and solicit feedback.

The Formula

Below is the formula for the current version of Offensive Player Tracking Plus Minus (PT-PM), like all plus/minus metrics PT-PM is calibrated around zero as average production that would lead to a .500 record. The player tracking data stats are in italics.

Offensive PT-PM ~ .643 * Points40 + -.54 FGA40 + -1.102 * TOV40 + .15 * Points Created by Assists40 + .8 * PassingEfficiency + .316 *Oreb40 + -.178 * FTA40 + 1.186 * C&S Points Ratio

Points!

To start off, fancy cameras in the rafters or not, the best determinant of a good offensive player, or team, is scoring efficiently. Combined, points per 40 minutes, FGA per 40 and FTA per 40 explain 62% of the model’s power. As we would expect, scoring points is a positive in the model, while taking shots (using possessions) is a negative. Below I have a table that works out the model’s math on scoring, assuming no free throws taken, and serves as something of a logic test.

Basically, volume helps as long the player is able to maintain a decent efficiency and the more efficient the better, naturally.

There is an issue that needs further investigation in terms of the proportion of the penalty for free throw attempts to the field goal attempt penalty. Generally the accepted ratio in basketball analytics for free throws is 0.44 possessions used for every free throw attempt to account for the expect number of “and one” opportunities. In this analysis the penalty for free throw attempts is lower than 44 percent of field goal attempts. Free throw rates are systematically higher among big men, who tend to perform worse in scoring volume and assists, so more analysis will need to be run to decide if that is related to the coefficient in the model.

Passing!

There are two terms in the model related to a player’s ability to pass the ball. One is a stat that I outlined in an earlier post with my first post looking at SportVU stats in a plus/minus framework, Passing Efficiency, scaled as assists per ten passes. Passing Efficiency can best be thought of as passing with a purpose. One interesting thing about passing efficiency is that gap between wing players and points is smaller than for volume assists, with Manu Ginoboli, for example, being one of the league leaders last year.

In addition, I am using points created by assist per 40 minutes rather than simply assists per 40. This measure, which credits three-point assists for the extra point and includes free throw assists, consistently scored better in the modeling process than simple assists, even though the two measures are very highly correlated. Players like Chris Paul, Russell Westbrook and LeBron James, playing in elite offenses were all players showing up on the ‘over’ side of the relationship between the points created by assists and simple assists.

Rebounds

Here’s where SportVU creates a modeling conundrum, it is clear that contested offensive rebounds are more valuable in the model. In fact, uncontested rebounds have no value in any of the regressions. This could be interpreted as uncontested offensive rebounds mostly being the result of luck, or close enough to not matter. Or the effect is hidden based on the context of other variables in the model. For now I have decided to not split the rebound types until I have a better analytic basis for apportioning the value, and am sticking to the plan vanilla offensive rebounds.

Catch and Shoot Ratio (And Stretch)

The catch-and-shoot points ratio (C&S Ratio) measure is calculated as simply the percentage of a player’s points scored as a catch and shoot jumper. The C&S Ratio is pretty well correlated to three-point rate, but with an interesting difference since some front court players that shoot from the mid-range show up decently here.

The value of simply attempting threes is pretty well established in terms of spacing and analytics, but RAPM data and other observations also indicate that the value on offense of front court players with extremely limited range may be overstated if we look at only their true shooting numbers (Cough, DeAndre Jordan, cough). To be clear this was the least important term in the model, and was only narrowly ahead of three-point rate in my cross validation checks. But, I think it may possibly be pointing to something interesting, that spacing doesn’t begin or end at the three point line.

Conclusion

Basically the new version is heavily correlated to the original version released in the off season, with an R2 of .89, which I would expect.  The issues still to be worked out revolve primarily around the value of contested vs. uncontested offensive rebounds and exploring the catch-and-shoot vs. three-point rate factor, plus possible new goodies released this year like shot clock usage for scoring and contested shot rates.