Nerdnumbers avatar

NBA Minutes are Hard to Predict

Intro

Per-minute performance is one of the most consistent things in basketball. However, guessing how many minutes players will play from season to season is a much more daunting task. We do a quick breakdown about how minutes change season to season in the NBA.

The NBA season approaches and we'll venture again into prediction territory. As a reminder, we do not find predictions to be a good test for production metrics. The key reason is that while it is fairly easy to predict how good a player is, it is hard to know how much they will play, and at what position. Injuries, coaching, and other factors can cause large swings in how minutes in the NBA are dolled out. I decided to do a quick examination of NBA history to show how minutes change year to year.

Returning Players

The first group of players I looked at was those that returned. From year to year, the majority of minutes are handed out to veterans. However, how those minutes are handed out varies. I took the difference in minutes for every season they played and looked at how different that total was year to year.* Here's what we got.

Yes, the returning set of players gets the lion share of the minutes from the previous season. Except, these are allocated much differently. For over the last decade the difference in minutes from returning players has accounted for almost one-third of the total minutes played! Players that are entering their prime start to get more minutes. Aging players start going to the bench. Random coaches decide various role players deserve more or fewer minutes than they were getting. In the end, even for returning vets it's hard to pin down exact minutes.

We're not done yet either!

New Players

Every season the NBA holds a draft that brings in a crop of new players. Some teams find talent from other leagues. And sometimes players that were sidelined for entire seasons return. And this means every year there's an influx of new minutes. I broke down how that looked as well.

This is not as severe as our returning players. And in recent years this has dropped. However, almost 10% of the minutes in a given NBA season are by players that weren't on the court the prior year. This is important for a key reason. While there are factors about rookies that translate to NBA performance, the majority of what a rookie will look like in the NBA is unknown. Even for good NCAA models, only 40% of a player's variance can be explained by their college numbers. That means every season a tenth of the league is new minutes from players that are highly unpredictable (rookies and players returning from long term injuries)

We have one last group to examine.

Exiting Players

Of course, if new players come in, and existing players get different minutes, that means some minutes are lost. Players leaving to retirement, permanent injury, or that can't find an NBA team that wants them drop out. Here's a look at how many minutes are lost season to season in the NBA.

This is not as severe as our previous swings. However, you can expect somewhere between five and ten percent of the NBA to "disappear" season to season.

All Together

Between our three swings -- 30%+ minute changes, ~10% to new players, and 5-10% lost -- in any given NBA season almost half of the minutes are different than the previous season. Yes, per-minute production is important for determining success. But the minute part of that equation should make it clear that knowing who will be on the court is also important. One need only look at the Miami Heat with a healthy Dwyane Wade playing a full season versus an injured Dwyane Wade playing limited minutes to know that. Many analysts will guess how the upcoming NBA season will look. Most of them will know who the good players and bad players are. However, none of us will have a clear picture of how the minute landscape in the NBA will shake up. That is, of course, if history repeats itself.


*The season differences are normalized. That means the minutes were adjusted so the minutes in both seasons were the same and then a difference was taken using the current year's minutes. This was to help for things like the lockout seasons and changes in the NBA schedule length.

This makes college basketball predictions even more laughable. Especially with "impact"(the NBA's equivalent to lottery picks) freshman, transfer(get players from other leagues, injured players, and players still developing. There are lots of parallels between all sports. All the biases and fallacies, we see in the NBA are very similar. I've bugged Berri with emails about this a lot. Predictions and rankings tend to be popular because of self fulfilling prophecy and they get page vve bugged Berri about this a lot.
"NBA Minutes are Hard to Predict"

Come on now. You haven't even tried! I think I deserve some analytical work with a bunch of regressions and stuff. After all I've paid a lot money to read your blog - this is a very expensive subscription.
AL_S,
Of course Arturo tries this yearly :) I will just say the one two punch of not knowing who will be injured, and not knowing who the coach will play instead, is killer.
Andrew,
Yup! I mean if you see my views, I've progressively gotten less and less excited on trying to "predict" how 400+ moving parts work is just silly. And the fact that some use this to "prove" a metric is even sillier.

It's fun to argue about these things, and say when you were right. But if you don't add the why, it doesn't matter. I mean Cleveland has an easy if/then. IF Kevin Love is healthy THEN they will win many more games. Easy to check at season's end.
The Waiters problem is going to be interesting. How long does he last until James and Co. are hating him and his antics?
Looking at the Vegas sport book, i'm not trolling when I say this but I like Denver and Philadelphia with the over this season. That's about as risky as I would be willing to get.
Phi: 15.5
Den:40.5
I remember Patrick saying in one of the podcasts that bsg website will have a tool which would enable users to play with data for ex. Dre thinks that faried will only get 1400 minutes which I why denver will only win 50 but if we disagree with it we can change the minutes to 2000 and denver will be projected to win 53 games instead, my question is will wp still be used or will they use the new model that Arturo is developing?
Why not do it retroactively then? You can assign a wins produced to each player before the season (since it's "easy to predict how good a player will be"), then at the end of the season, use the actual minutes played to determine how many games that team should have won.

The differences between the actual and theoretical wins (or point differential) can help you gauge how effective the metric is, and what (if any) its shortcomings are. For example, if this model is consistently underrating/overrating star-studded teams like the Cavaliers, then that will give you valuable insight on how to improve the metric.
Shrinidhi,
A fun point but still not a good test. Pick up the Wages of Wins and Stumbling on Wins, in it we accept that aging, coaching, and changing teams matter for performance. Most performance models are supposed to explain player performance. We know Wins Produced does this and the consistency implies it is skill, not chance. But if we want a "predictive" model to explain next season performance, it has to include the factors that could impact this. I stress, this was not the intent behind Wins Produced (or most performance metrics.) If this is the desired test, then a model that includes those factors is required.
Mostly the way to turn it into a better predictive model is to regress to the mean. WP might correctly identify someone as having just had a very good season, but if it was anomalously good for that player then it's not necessarily a good predictor for the next season.
creedofhubris,
Yup. There are tons of things I'd do to actually make a predictive model. Regression to mean, injuries, age, expected position. The thing is most metrics (including WP) are descriptive metrics. Their goal is to explain what happened. The predictive nature is a byproduct that helps say "These things are likely skill." but not the ultimate test to see if the metric is effective.

Sign in to write a comment.