The Definitive PER Criticism

The goal of PER is very noble.

It wants to tell us who the best players in basketball are.

Does it actually do this?

Not really.

The PER formula is a trailblazer of sort (pun intended) in the world of analytics. It was practically the first formula available for public consumption based on the notion that numbers can help us better understand the game of basketball.

So in that sense it is, in fact, a success, and it helped pave the way for this new age of analytics that we are enjoying.

However, in fulfilling its actual goal (which is telling us who the best players are) it fails woefully.

Before I start writing this piece, I just wanted to say that I am a huge fan of analytics and advanced statistics in general. This blog is called “The Z Score” for a reason (my name is Zein, so that worked out well). I study Economics at university, so we learn about a lot of the tools used in constructing these formulas, so I have a good enough grasp on this issue to write about it.

The Philosophical problem with PER:

The problem with PER and some of the similar metrics of its ilk is that they are based on a flawed rationale. People who use this stat often talk about how a box score is not  a proper assessment of a player’s ability.

Fair enough.

That isn’t the problem.

The problem is when they say a metric like PER is a much better way to judge a player. All PER does is take all the counting stats you see in a box score, attribute different weights to them,  and scale them with the league average. You aren’t getting that much more insight concerning a player’s skill level than you would have by just looking at the box score. This is where advanced statistics such as ESPN real plus-minus have a decisive edge over their contemporaries.

Does it provide some insight if done right?

Yes, it does.

But, I personally believe PER does not do it right.

For instance, one of the many criticisms of the model is the lack of a “scientific approach” taken by John Hollinger. He defends PER by stating that the rankings it spits out correlate heavily with what people think they should be.

Doing that is highly inappropriate, as Dave Berri (author of the wages of win) notes far more eloquently than I can:

In other words, his metrics fit what he believed about the players before he started.

Unfortunately, this is not the way science works. We do not begin with our beliefs, play with the numbers until our beliefs are confirmed, and then call it a day. Models are not evaluated in terms of whether they are consistent with what we believe, but in terms of their ability to explain what we purport to explain (and furthermore, provide predictive power).

The proper way to this kind of work whether it is in Economics or Basketball is to build a model which attempts to predict how many points, assists, rebounds, etc., a player would log, then regress each input.

Doing that is extremely challenging which is why we don’t see many people doing it.

The Physical problem with PER:

Redefining the metric:

Now that we’ve debated the philosophical issues with PER, it’s time to get into the nitty-gritty.

PER claims to be a holistic measure of a player’s performance, meaning that it takes into account the offensive and defensive performance of the player. However, as we all know it’s more of an offensive statistic because it only takes into account steals, blocks and defensive rebounds.

So why not do away with all those different inputs and claim that you’re an offensive only statistic?

To use a real-life example, let’s look at Al Horford for a second. Clearly, he is an above average NBA player and even deserving of his all-star selection this year. But, the moment we compare him to the MVP candidates he doesn’t look so good.

So taking this analogy back to PER. The metric is trying to do way too much and not succeeding in doing it. So why not scale back on the inputs and improve the integrity of the metric.

Positional Bias:

PER also has a bias regarding what kind of player it favours over others. PER values big men who play a limited amount of minutes, rebound well, have a low turnover rate and shoot a high percentage. The best example of this is Enes Kanter, who, according to PER, is the 14th best player in the NBA, which he clearly isn’t.

You might be thinking to yourself, the player I just described sounds like an excellent player. But there is a reason why he is playing such limited minutes: he isn’t that good. There is a reason why he shoots at a high percentage: he creates none of his shots. Finally, the reason he has a low turnover rate is because he doesn’t have the ability to warrant having the ball in his hands.

The PER metric is far less favourable to guards. If you look at the top 20, all time PER leaders only 5 of them are guards (Jordan, Paul, Wade, Magic and Robertson) as opposed to the eight you would expect if this was a normal distribution.

So, to fix this positional bias, it would be more optimal to scale each input by the positional average and not the league average. This would lead to a more predictive statistic because it is more appropriate to compare a point guard’s assists with the average for point guards and not that of the entire league.

Similarly, rebounding and low turnovers are significant components of this statistic, and clearly big men are going to be better performers than their contemporaries.

I understand that we are moving into an era of position-less basketball spearheaded by the Warriors. Having said that, I still feel like scaling each input with that of their peers is a better way to evaluate a player’s performance.

Modifying the inputs:

Some of the inputs used in the PER formula are very superficial and aren’t indicative of all the nuances of an NBA game.

Take turnovers for instance. In real life, there is a very real difference between a dead ball and a live turnover. For obvious reasons, in the case of former, although you missed, your defence still has a chance to set up and the possession from that point can basically be viewed as a normal possession. In the case of the later, it’s usually a fast break where you don’t have numbers and are very likely to be scored on.

Although the ratio of live ball to dead ball turnovers is pretty even split, with 51.4% being live ball in 2011. You would expect the distribution to be different when comparing individual players. Therefore, by distinguishing between the two, you would get a better idea of how damaging a turnover is and how efficient a player is.

Another input which could use some tweaking is assists. One of the major flaws with PER is that it does not take into account the differing abilities of NBA players.

What do I mean by that?

If you have better teammates, a higher percentage of your passes will be assists. The opposite is true with bad teammates.

A great example of this is John Wall (who ranks 42nd by PER) who according to NBA stats and info is averaging 19.4 potential assists with only 9 of them actually translating into field goals made.

Amin El Hassan (who is probably the best twitter follow there is) said that when he was working in Phoenix, the Suns wanted to draft Steph Curry. One of their reasons behind this was that Steph had bad teammates, so his assists numbers were very low. So they tracked his potential assists which they call “Lovedals” (after one of his teammates).

So, a better picture of a player’s passing ability would be to take into account the actual context of a team’s ability and not judge a player like he is in a vacuum.

I could go on and on about my problems with PER, but no one wants to read a long article about this. So I might do a part 2 depending on how this goes.


Hope you guys learned a thing or two

Shami out

Facebook Page: