Introducing RAPTOR, Our New Metric For The Modern NBA

At FiveThirtyEight, we’ve been running NBA predictions since 2015. We started with Elo ratings before introducing our CARMELO player projection system, which we then incorporated into our “CARM-Elo” season prediction model. We tested and tweaked the prediction model over the years, but it was always powered by metrics from other sources, such as Box Plus/Minus (BPM) and Real Plus-Minus (RPM).

But that changes this year. RAPTOR, which stands for Robust Algorithm (using) Player Tracking (and) On/Off Ratings, is FiveThirtyEight’s new NBA statistic. We’re pretty excited about it. In addition to being a statistic that we bake in house, RAPTOR fulfills two long-standing goals of ours:

• First, we wanted to create a publicly available statistic that takes advantage of modern NBA data, specifically player tracking and play-by-play data that isn’t available in traditional box scores.
• Second, and relatedly, we wanted a statistic that better reflects how modern NBA teams actually evaluate players.

NBA teams highly value floor spacing, defense and shot creation, and they place relatively little value on traditional big-man skills. RAPTOR likewise values these things — not because we made any deliberate attempt to design the system that way but because the importance of those skills emerges naturally from the data. RAPTOR thinks ball-dominant players such as James Harden and Steph Curry are phenomenally good. It highly values two-way wings such as Kawhi Leonard and Paul George. It can have a love-hate relationship with centers, who are sometimes overvalued in other statistical systems. But it appreciates modern centers such as Nikola Jokić and Joel Embiid, as well as defensive stalwarts like Rudy Gobert.

Our sports podcast, Hot Takedown, discusses RAPTOR

This article will go over some of the highlights of how RAPTOR works. For a much deeper and more technical description, you can find our methodological explainer here. But these are the highlights:

• Like BPM and RPM, RAPTOR is a plus-minus statistic that measures the number of points a player contributes to his team’s offense and defense per 100 possessions, relative to a league-average player. For instance, a player with an offensive RAPTOR rating of +2.1 boosts his team’s performance by 2.1 points per 100 offensive possessions while he is on the floor. Likewise, a player with a defensive RAPTOR of +3.4 would improve his team’s defensive performance¹ by 3.4 points per 100 possessions while he’s on the court.

• Plus-minus statistics have certain inherent limitations, and RAPTOR is subject to those, too. Namely, these statistics assume that player performance is largely linear and additive, that is, that you can roughly add up the ratings from individual players to project team performance. In other words, RAPTOR does not account for coaching, systems or synergies between teammates.

• RAPTOR consists of two major components that are blended together to rate players: a “box” (as in “box score”) component, which uses individual statistics (including statistics derived from player tracking and play-by-play data), and an “on-off” component, which evaluates a team’s performance when the player and various combinations of his teammates are on or off the floor.
• The variables included in “box” RAPTOR were chosen by evaluating how they did in predicting long-term RAPM (real adjusted plus-minus), which is a measure of how a team’s performance changes when the player is on and off the floor. Essentially, this is the same technique that BPM used, only RAPTOR uses play-by-play and player tracking statistics in addition to traditional ones. On offense, for instance, in addition to using traditional statistics like points, RAPTOR accounts for factors such as how many of a player’s field goals were assisted and how valuable these assists were,² the value of different types of offensive rebounds, time of possession and various measures of floor spacing, such as the number of contested 3-pointers that the player took. On defense, it looks at factors like how often the player was the nearest defender on an opponents’ shot and how often those shots went in,³ how many points and rebounds were scored by opponents at the defender’s position, and how often the player induced offensive fouls.

• The “on-off” element of RAPTOR evaluates how a player’s team performed while he was on the floor, how the player’s courtmates (the teammates that the player most often shared the court with) performed while they were on the floor without the player, and, finally, how those courtmates’ other courtmates performed when they were on the floor without the player’s courtmates, all adjusted for the strength of competition they were facing. We know it might sound a little goofy. But it’s relatively simple to calculate. And it correlates very well with RAPM, while stabilizing a lot faster than RAPM, which can take years’ worth of data to estimate reliably.
• Overall, however, RAPTOR weights the “box” component more highly than the “on-off” component. In testing RAPTOR on out-of-sample data, we found that while on-court/off-court stats provide useful information, they’re nonetheless quite noisy as compared with individual measures of player value that are used in the “box” part of RAPTOR.

• When applied to past data — for instance, in evaluating who the best players were in the 2018-19 season — RAPTOR is a descriptive statistic. Descriptive RAPTOR is based solely on a player’s on-court performance and the performance of the player’s teammates, as described above. It does not use priors based on a player’s height, weight, age or any other factor.
• However, RAPTOR can also be used to make team and player predictions, and indeed our NBA predictions are now fueled by RAPTOR. (We are retiring the CARMELO brand name from our previous projection system, although much of the code for RAPTOR projections is borrowed from CARMELO.) RAPTOR-driven predictions do use qualities such as height, age and draft position, and even whether a player recently appeared on an All-NBA team — that data improves the performance of the predictive measure. Predictions also weight variables slightly differently than descriptive RAPTOR does, as certain statistics are more subject to luck than others. We refer to this predictive version of RAPTOR as PREDATOR (PREDictive rApTOR).

• RAPTOR is based exclusively on publicly available data. There are other player-tracking statistics we believe could be highly helpful to RAPTOR, especially more detailed measures of on-ball defense, so we hope to be able to revisit RAPTOR as additional data becomes available.
• RAPTOR’s name (in addition to being a whimsical backronym in the tradition of CARMELO and DRAYMOND) honors the 2018-19 Toronto Raptors, which FiveThirtyEight’s previous projection system correctly predicted had an edge over the Golden State Warriors (even though we didn’t fully believe the projection ourselves at the time).

The full-fledged version of RAPTOR is available for the 2013-14 season onward, as that’s when the NBA’s player-tracking data came on line. We also have a historical version of RAPTOR called Approximate RAPTOR dating back to 1976-1977, the first season after the ABA-NBA merger, but that uses a far more limited range of data.

RAPTOR ratings for players with at least 1,000 minutes played⁴ in a season since 2013-14 can be found in the table below. As you can see, RAPTOR generally loves perimeter players and wings, such as Curry, Harden, Leonard and Chris Paul, although some frontcourt players like Jokic, Anthony Davis and Draymond Green are also rated highly by the system. For more detail on past RAPTORs, including the breakdown of box and on-off components, you can download files that list the regular season and playoffs separately, or a version that combines a player’s appearances over the course of the entire season⁵ into one file.