Rating the Ratings
1995
There are a number of different statistics people use on the rec.sport.basketball.pro
newsgroup to support all the woofing that goes on. I'm not so
much interested in the woofing, but in these statistics and
whether they make any sense. At all. So I'm going to look
at a few and discuss some of their features and what they contribute
to our knowledge of the game. I will do this in no particular
order, just so you don't think I'm playing favorites....
Team Points per Possession -- Tony Minkoff posts this
every so often for teams, using Doug Steele's raw stats.
All it does is divide the team's points scored and allowed
by the number of possessions used by the team. Fortunately,
Minkoff uses a good definition of possessions. I believe
his formula is FGA - OR + TO + 0.4 * FTA. Basically, this
means that a team has possession until the other team has
possession -- this is so obvious it's hard to say. Other
definitions do not subtract the offensive rebounds. This
implies that a team shooting the ball, then getting the offensive
board has two possessions. By defining possessions the way
he does, Minkoff essentially says two opposing teams in a
game have an equal number of possessions with which to win
the game, just as each baseball team has 27 outs with which
to win the game. Defining possessions the other way doesn't
give this nice feature. Anyway, this rating is simple and
essentially states a fact about a team, removing the illusion
caused by fast vs slow paces. It is a fundamental tool
for evaluating teams' abilities to score or prevent scoring.
Minkoff Player Rating -- My understanding of this stat is that
it takes a linear regression of minutes per game against each
per minute stat, then uses those coefficients to compute the
number of minutes "earned" through a players statistical performance.
This may be modified by weighting it with actual minutes per
game. This is one of the most interesting player stats I have
seen and I have seen a lot. The concept of rating players by
how much time they deserve is very innovative and conceptually
simple. I am not convinced that linear regression is a great
tool for a game as complex as basketball (I know, I know, Dickie
V. calls it a simple game.) I think this may be why the stat
is fudged by weighting the actual minutes per game along with it.
For what it is, this rating is quite good, but it does have
the limitation that it serves no other purpose other than to
woof. It is a rating. It ends all debate by hiding it behind
one number.
Tendex (and all linear weights) -- Tendex was created by Dave
Heeren, a real sportswriter in Florida (for what that is worth).
It basically consists of adding a player's positive stats
and subtracting his negative stats, sometimes weighting
each of these by some amount. Doug Steele has his version.
This, like the Minkoff Player Rating, is pretty much just a
rating, ending all debate. It has an analogous form in baseball
analysis, called the Linear Weights method. Bill James argued
against it a lot because he had an alternative method and,
well, there tends to be a proliferation of people in this world
who think you can always add things up and get the true answer.
Tendex has the weakness in that there is no good theory
for what the weights on all the stats should be. Some people
say that offensive rebounds should be weighted heavily, some say
no. Some say they should be weighted the same as defensive
rebounds, some say no. Some say assists are the same as
blocks, some say no. There is no good theory to guide us on
what the weights should be because a linear weight model
of basketball is not a good model of the game.
Norris Power Ratings -- This is essentially a filter, as known
to statisticians or mathematicians.
Kalman filters are common
if anyone out there has heard of them. It essentially takes the
score of games, the location, and tries to "learn" how good a
team is based pretty much on these two facts. Traditional filters
also have information like how much variability a team has. Norris'
does things a different way, but uses pretty much the same
information, making me think he has a filter in a different
formality. My congratulations to him if it is a filter.
These have proven successful in predicting physical processes,
even complex ones. I believe Norris' does pretty well, too.
Basically a lot of information is wrapped up in the
scores of games and the home-road thing. It ignores the
details, which is a problem when, say Michael Jordan returns to
the Bulls in midseason, but otherwise it does pretty well.
This is probably a good thing for all you gamblers out there,
but don't blame it on me if you lose to Vegas -- they use
the best filter of all.
Jon Moser's Player Efficiency Rating -- I have only seen
this a few times, but it appears to divide a player's points
scored by the number of shots taken, essentially. I may be
mistaken on this. Obviously it ignores rebounds and assists
and maybe turnovers. It does what it says it does and doesn't
say anything more. I don't think he claims it says more.
That is good.
Finally, I want to add that Doug Steele's stats, which are the
basis for a lot of the numbers we see posted are deficient in
one slight area. Most public sources of stats do not include
team turnovers, incurred on 5 second calls or 24 second calls,
usually. This is an increasing source of error as 24 second calls
continue to rise with a slower paced game. It is still minor
and Doug sometimes gets them, but I have compared his numbers
with official NBA stats and his always have too few team turnovers.
This, I will reiterate, is no fault of Doug's. Doug does
good work. I just wish he would show us his individual defensive
stats, too.