Statistical Test: How Well Does Net Rating Correlate With Wins?

There is a maxim that this site lives by, and most of you who follow me on Twitter (@RealFoxD) know what I’m about to say:

“Great teams win big and lose close.”

Which, in layman’s terms, ultimately reduces to the fact that the higher a team’s point differential over the course of a season, the more games they win. It makes logical, intuitive, downright obvious sense.

But the bigger question at work here is “just how much is, say, an extra point per game worth over the course of an 82-game season?”

Since this site’s inception, I’ve used an assumption originally posited by Dean Oliver, the father of modern basketball analysis, that states “30 points is worth a win.”

Indeed, most conversions of things, like VORP to Estimated Wins Added, would have you multiply by 2.73 (82/30) to convert one to the other.

There’s just one problem with this: It’s pace-blind. It might work in one season but not in others depending on if the game is slower or faster (let’s face it, 30 points is a lot different in a mid-90s Cavaliers-Pistons game (average pace: 85) compared to a 2019 game between the Bucks and Rockets (average pace: 105.)

Now, I’m a guy with a nine-to-five who has better things to do than dump every season in NBA history into a spreadsheet, but what I do have is time to dump data from every season since 2013 (per Basketball Reference’s season summaries), roughly dovetailing with an increase in pace and scoring efficiency and comprising a 210-team sample size, into Alcula’s Linear Regression calculator (http://www.alcula.com/calculators/statistics/linear-regression/ )

I’m working with two simple variables. On the X axis, net rating, from the putrescent 2014 Sixers (-10.5) to the sublime 2017 Warriors (11.6, which was actually better than the 10.7 they put up in the 73-win season the year before, one reason they nearly ran the table in the playoffs.)

On the Y axis, win total, from the 10-win 2016 Sixers to the 73-win 2016 Warriors.

Then it was just a matter of letting the computer crunch those 210 pairs of data and give me a result.

Let’s start with the little bit of weirdness. You’d expect, through logical sense, that the mean would be 41 for wins and 0.00 for Net Rating, since every game has a winner and a loser and the winning margins cancel each other out, right?

Well…OK, we got super-duper close, but there is just a tiny bit of rounding weirdness insofar as the average win total was 40.995 and the average Net Rating was 0.004. But considering rounding integers and numbers to one decimal place never comes out quite right, I’m just sticking it there for the sticklers.

The regression line equation? To two decimal places on wins and three on Net Rating, Wins = 40.99+2.515x (where X is NetRtg.)

Now here’s where things REALLY get screwy.

You’d think that margin of victory would be a better predictor of wins than derived stats like Win Shares or Value Over Replacement Player (both of which have been the subject of their own Statistical Tests on this here site), right? After all, at the end of the game, one team scores more points than the other team regardless of nuts-and-bolts things like rebounds or turnovers or whatever, right?

The r² on this is still pretty strong at .944…

But that isn’t as good as Win Shares (.963) or VORP (.953). You can actually do a better job of predicting a team’s win totals by the output of their players than you can by how many points they win or don’t win by.

I have a couple of hypotheses as to why.

For one thing, coaches don’t run up the score (although this is starting to change as “no 30-point lead is safe.”) There’s a natural brake on Net Rating when a team gets up 20, puts the scrubs in midway through the fourth quarter, and doesn’t sweat the late run that makes a closer-than-the-score game.

For another thing, there’s “load management” to consider. A team that on paper should beat another team by 20 might only win by 5 because they sat a star or two. Likewise, an injury stretch where they win a lot of games that are a lot closer than they’d be if the star was healthy is going to skew a team’s net rating.

WS and VORP are designed to take a holistic player contribution from game to game relative to other players on his team into account and assign credit for a win (or blame for a loss) in a way that smooths out over the course of a season.

Point differential is too subject to the human vagaries of the sport to do that.

And let’s circle back to one other thought from earlier. There’s a big difference between a game with 85 possessions in the NBA and one with 105 possessions. In theory, the more possessions, the more chances the better team has to beat the worse team; it’s why the Knicks could beat the Lakers on any given day in 2020 but the chance the Knicks have of winning a seven-game series against the Lakers is so close to zero that unless half the Lakers players got hurt, it’d be a foregone conclusion.

So what happens when we regress the 2018-19 data (full season) and the 1996-97 data (the slowest non-lockout season in NBA history) and then throw in the 1982-83 season (the fastest season in the 3-point era but a season where teams tended to count two at a time)?

For one thing, way-out-there Net Ratings weren’t really a thing in 1983; the 76ers, who won the title, had a 7.4 NetRtg, far lower than even your average 3 or 4 seed today. But beyond that…

Well, here are your r² for all three:
2018-19: .960
1982-83: .957
1996-97: .941.

Bingo. Fast-paced seasons, even independent of the wacky variance that the broad use of the 3-pointer has introduced into the league, have stronger correlation between Net Rating and wins than do slow seasons (yes, I know that’s a 3-season sample size.)

Which makes sense. In college or the pros, how do bad teams beat good teams? They slow the clock down, jump out to a lead, and run out the clock before the law of averages takes over.

This also redeems Net Rating as a stat compared to WS and VORP, as the 2018-19 data, comparing one season’s apples to apples, lines up nicely with the WS and VORP correlations, falling in between those two numbers.