Skip to main content
Menu
What’s New In Our NBA Player Projections For 2017-18

Welcome to the 2017-18 edition of CARMELO,1 FiveThirtyEight’s NBA projection system.

Or to be more precise, welcome to the initial 2017-18 edition of CARMELO. We have a lot of basketball-related projects on tap for the summer, which vary from exploring what makes teams succeed in the playoffs when they struggle in the regular season to trying to develop our own defensive metrics based on player-tracking data. So there’s a chance that we’ll give CARMELO and our other NBA products a more serious overhaul before the NBA regular season starts in October. Be prepared for these numbers to change, in other words.

In the meantime, the initial adjustments from last year’s edition of CARMELO are relatively minor. For a basic outline of the system, which projects player performance by identifying similar players since the NBA-ABA merger in 1976, please see our explanation from two years ago. Here’s what’s different this year:

BPM vs. RPM, revisited, again

We were fairly pleased with how our CARMELO player projections performed last year, with the system identifying breakout stars such as Giannis Antetokounmpo, Nikola Jokic and Rudy Gobert. But we also use CARMELO to make team projections, and we weren’t all that happy with those. In fact, according to the APBRmetrics message board, which tracks various projection systems, CARMELO went from having some of the most accurate team forecasts in its debut season in 2015-16 to some of the least accurate last year. What happened?

In 2015-16, our projections were based on a combination of Real Plus-Minus (RPM), a system that rates each player primarily based on how his team performs when that player is on or off the court, and Box Plus/Minus (BPM), a system that rates players using traditional box-score statistics such as assists and steals. Last year, we switched to using BPM only. Why? There are a lot of things to like about BPM, including that it’s considerably more transparent than RPM, and it can be calculated going back several decades, making for easier historical comparisons.

But as a predictive tool, BPM does not appear to be as accurate as RPM. Instead, BPM has trouble picking up on factors such as defense and team cohesion. That led CARMELO to overrate teams such as the Minnesota Timberwolves and underrate more defensive teams such as the San Antonio Spurs last year. If we’d run the numbers using RPM instead of BPM in 2016-17, our projections would have been above-average again as compared with the projection systems that APBRmetrics tracks, we discovered.

All of this stuff gets complicated, and discussions can quickly devolve into alphabet soup. But for better or worse, the choice of metric matters quite a lot. According to BPM, Russell Westbrook’s 2016-17 season was easily the greatest in NBA history. According to RPM, he was only the ninth-best player in the league last season.

The upshot is that in the short run, we’re using a blend of two-thirds RPM and one-third BPM for this edition of the CARMELO projections. In the long run, we’re interested in developing our own plus-minus stat (but no promises about that quite yet).

Projections for international players

In past years, we’d published CARMELOs for NBA veterans such as Paul George based on their NBA statistics and for rookies such as Lonzo Ball based on their NCAA statistics.2 But we didn’t have projections for rookies such as the Knicks’ Frank Ntilikina, who played in international leagues rather than the NCAA. (The Belgian-born Ntilikina grew up in France and played for the French professional team SIG Strasbourg this year.)

So this year, we’ve introduced simple projections for European draftees based on biographical data: specifically, their age, draft position, height, weight, position and home country.3 CARMELO identifies players such as Ricky Rubio, Evan Fournier, Dennis Schroeder and Tony Parker as being similar to Ntilikina, for instance.

‘Gap year’ and ‘draft-and-stash’ players

Another tricky case involves rookies who missed the entire NBA season, or almost the entire season, on account of injury or other factors in their first year. The Sixers’ Ben Simmons, the No. 1 overall draft pick in 2016, didn’t play at all last season because of a foot injury, for example.

We’ve improved how the system handles these “gap year” players.4 They are now compared only to other players who also sat out what would have been their first season, such as Blake Griffin (who missed the entire 2009-10 season) or Julius Randle (who played just one game in 2014-15 before getting hurt). This is a small sample of cases, so the projections for these players can be somewhat noisy. Still, we think this is better than essentially just ignoring the injuries, as we’d been doing before.

European and other international players, meanwhile, often play an additional season or two abroad even after they’re drafted by an NBA team. (The NBA team who drafts these “draft-and-stash” prospects can retain their rights for a couple of seasons under most circumstances.) Draft-and-stash players such as the Celtics’ Guerschon Yabusele are projected based on comparisons to other draft-and-stash players.

Better biographical data

Finally, we’ve switched to using NBA.com, instead of Basketball-Reference.com, as our primary source for player heights and weights. Players’ listed weights — and sometimes even their official heights — can change from season to season. Basketball-Reference.com generally lists players by the heights and weights they were listed under as rookies, whereas NBA.com keeps more up-to-date with the changes.

We’re also using more detailed data on players’ positions than in the past. While a player’s primary position is determined by Basketball-Reference.com, we’re now also accounting for players’ secondary positions based on additional positions listed at ESPN Fantasy Basketball. Thus, a power forward who also plays center will be treated slightly differently than a power forward who also plays small forward.

We hope you’ll enjoy this year’s edition of CARMELO, and we’ll notify you of further changes as we make them throughout the offseason.

Footnotes

  1. CARMELO stands for “Career-Arc Regression Model Estimator with Local Optimization,” but really it’s a shout-out to the New York Knicks’ Carmelo Anthony.

  2. NCAA statistics are adjusted for pace and opponent quality by ESPN Stats & Information Group.

  3. The projections do not use statistics from European leagues; they’re solely based on these biographical categories. Home country is determined based on where the players were born, rather than on where they played professionally. Thus, Ntilikina is classified as Belgian rather than French, for example. Players born in the same country or from nearby countries get a higher similarity score, although this is a fairly minor factor compared to other variables such as draft position.

  4. Specifically, players who played less than 200 minutes in their first NBA season.

Nate Silver is the founder and editor in chief of FiveThirtyEight.

Comments