The basics of CARMELO are the same as last year. We’ve run projections for 485 veterans and 82 rookies. The system identifies historical comparables since the ABA-NBA merger in 1976 to project the career path of today’s players. LeBron James’s top comparable is Larry Bird, for instance, while the system likens Lakers rookie Brandon Ingram to Andrew Wiggins.
Like any sophomore, CARMELO has undergone a few changes around the margin — more about those in a moment. First, though, a quick review of how the system performed last year. Analyzing the projections for individual players is hard, because there aren’t really a lot of publicly available projection systems to compare CARMELO against. The list of players the system liked the most holds up pretty well, however, having been a little ahead of the curve in identifying the value of Kawhi Leonard and Draymond Green, and the breakout of Giannis Antetokounmpo, among other successes. On the flip side, there were CARMELO’s inexplicable crushes on Marcus Smart and Elfrid Payton, who will have to show significant improvement this season to live up to the system’s lofty expectations.
Another way to evaluate CARMELO’s performance is through its team-by-team projections, for which we can make some direct comparisons — namely, to projected team win totals from Las Vegas before last season. That comparison makes CARMELO look good. If you’d been betting on its projections, you’d have gone 18-11 against the Las Vegas spreads (skipping a bet on the Memphis Grizzlies because the CARMELO and Vegas projections were identical), or 13-4 if you restricted yourself to cases where the CARMELO and Vegas lines differed by at least two wins. CARMELO also had a higher correlation with actual win totals, and a lower root-mean-squared error, than Vegas did.
|Los Angeles Clippers||56.1||56.5||53|
|New York Knicks||24.8||31.5||32|
|Los Angeles Lakers||22.1||29.5||17|
|New Orleans Pelicans||44.9||47.5||30|
|San Antonio Spurs||57.1||58.5||67|
|Oklahoma City Thunder||58.0||57.5||55|
|Portland Trail Blazers||36.2||26.5||44|
|Golden State Warriors||61.3||60.5||73|
|Correlation with actual||.82||.76||—|
And CARMELO correctly identified the best teams in each conference, the Cleveland Cavaliers and the Golden State Warriors. It was somewhat too optimistic about the Cavaliers’ regular-season win total and too pessimistic about the Warriors’ but found some redemption when the Cavs beat the Warriors in the NBA Finals.
The most noticeable addition to this year’s version provides some context for all the record-setting contracts being signed under the league’s new salary cap. We’ve added a category called market value, which translates wins above replacement into a dollar figure for each player. (See here for a description of the thought process behind this calculation.) For the upcoming season — after a massive jump in the salary cap — we estimate that each win is worth about $5.2 million in market value. Next season, that dollars-per-win exchange rate will increase to $5.6 million.
Be aware, however, that CARMELO regards the best players in the league and the best rookies as being massively underpaid, and therefore most other players will appear to be overpaid. Essentially, these numbers reflect what salaries might look like if there were a team salary cap but no individual maximum salary. It projects stars like James to be worth $60 million a year or more, far more than the max.
Other changes are relatively technical. First, whereas last year CARMELO projections were based on a combination of Box Plus/Minus and Real Plus-Minus, they’re now based on BPM only. One problem with RPM is that it’s only available for recent seasons, whereas BPM can be calculated with standard player and team statistics dating to the 1970s. That poses a problem for a system that relies heavily on making historical comparisons. Although there are workarounds, we’ve decided — having had a year to review the system’s performance — that a BPM-only approach strikes a better balance between simplicity and accuracy. Frankly, we have designs on our own plus-minus metric that would eventually displace both BPM and RPM, but that’s something that will have to wait for a future incarnation of CARMELO.
The next change is even more technical. CARMELO projections are generated through a two-step process. First, the system produces a baseline projection for each player based on regression analysis. Then, it adjusts the forecast and generates a distribution of possible outcomes based on the comparable players. Although the second part of this process is almost exactly the same as last year, we’ve put more work into the first step, the baseline projection. The system is now a bit smarter about handling players with limited playing time. It also recognizes that different statistics have different amounts of predictive value. For instance, because shooting can be streaky, players who generated strong performances on the basis of good shooting seasons are more apt to regress to the mean than others. By contrast, generating shots, drawing fouls and taking 3-pointers are correlated with improved performance in future seasons. Rebounding, blocking shots and, especially, accumulating steals are also correlated with stronger future performance.2
All in all, this year’s projections should be a little better than last year’s, and last year’s did OK for themselves.