[ESPN] Pelton Mailbag: all-time defense team and player rating stats

Kevin Pelton provides a really nice breakdown of the various “all-in-one” player rating stats that are publicly available towards the end of this mailbag (the first topic addressed of all-time defensive team is also pretty interesting).

http://www.espn.com/nba/story/_/id/...uding-all-defensive-team-last-quarter-century

He makes a convincing case that looking at BPM, RPM, and his WARP metric is the best way of gauging a player’s overall impact statistically. I like his last sentence quite a bit too.

First, let's note one key difference: Some stats (PER here) rate players on a per-minute basis, while others rate the value they've provided, factoring in playing time (win shares) and replacement level (VORP). Oftentimes, you'll see the same stat used in both ways: Win shares per 48 minutes is a per-minute stat, as is box plus-minus (used to create VORP). Meanwhile, PER can be translated into EWA (estimated wins added), a value stat.

In addition to the three you mentioned, all available on Basketball-Reference.com, there are two other all-in-one metrics I consider: my own wins above replacement player (WARP) metric and ESPN's real plus-minus (RPM), which is used to calculate RPM WAR.

RPM is unique among these stats because it incorporates plus-minus data, adjusted for teammates and opponents, while also utilizing box-score stats for stability. Even with them, RPM tends to be noisier than box-score stats, so it can fluctuate from season to season. (The version of RPM on ESPN.com incorporates only data from the current season; there's also a multiyear version that I use for projections that is not quite as noisy but also tells us less about what has happened this season.) The upside is that RPM can capture skills that aren't tracked in the box score, particularly on the defensive end of the court.

Box plus-minus was designed to replicate RPM using only box-score stats. (Specifically, it's a regression that's built to value component stats like offensive and defensive rebound percentages by how well they predict a player's regularized adjusted plus-minus, the predecessor of RPM that uses only plus-minus data and no box-score stats.) Therefore, the weights on different stats are well-calibrated in box plus-minus, though the use of interaction terms -- assists are multiplied by defensive rebounds, for example -- can create problems when players like Russell Westbrook go beyond the historical norms for these stats.

WARP is probably most similar to box plus-minus. While it wasn't built off a regression, I've also used adjusted plus-minus to calibrate WARP, adding value to 3-point attempts to account for how players with many of them tended to outperform their WARP in terms of team impact. The key difference is WARP doesn't have any of the interaction terms in box plus-minus.

Those three are the stats I utilize the most because I think they're the best at isolating player value. Because win shares attributes all of a team's defensive performance to individuals, it tends to be more sensitive to team success than other all-in-one stats. On the other hand, PER puts relatively low value on defensive stats and does not account for team defense whatsoever, meaning it primarily captures offensive value. PER also tends to overvalue players with high usage rates at the expense of those who are more efficient.

Because of the varying strengths and weaknesses I've laid out, I think it's useful to consider multiple all-in-one stats while using the eye test to help explain the differences between them and where they might not fully capture a player value. Do be careful, however, of looking at various stats until you find the one that confirms the conclusion you already wanted to draw.​