APBRmetrics

Stumbled upon a few sources and many technical terms that I know little about. Wondering if anyone can help me understand if they would be useful for basketball analysis. I'll try to look them up myself with time, but would appreciate any quick explanations if you find any to be relevant.

Thurstone / Shapley analysis
ARCH and GARCH models
adaptive drift modeling
multi-dimensional scaling
VAR and VECM
zero-inflated count models
genetic search
jackknife techniques
conjoint analysis
choice modeling
survival analysis
bayesian discriminant analysis

https://books.google.com/books?id=eExA0 ... CA0Q6AEwAQ

There are 3 books referenced on sports and financial forecasting. Anyone read these or similar? Want to... and share thoughts?

I don't know your background, but I'd think that basic statistics would be a good thing to start with. I also think logistic regression is a good thing to learn about for anything where you want to estimate probabilities.

Your advice lays out appropriate skill development stepping stones. But I am interested right now in what I asked. Not for me to do at this time but to identify tools that others might be able to bring to bear, if they thought about it, and to see if they might have been used on NBA in academia and not yet found.

I'll knock one off the list for you: survival analysis deals with, well, the death rates for a population. Death can mean many things, like a failure. It deals a lot with a question like, How much of the population will survive after T amount of time? I'm not well-versed in this field, however. I think it can be applicable to the NBA for something like measuring how long players play in the NBA and when they retire. And it looks like I was right....

If you google NBA and survival analysis articles do come up, like this one:
http://repositories.lib.utexas.edu/hand ... 12-05-5631

"Bayesian hierarchical parametric survival analysis for NBA career longevity"

Thanks. This is exactly how I thought this technique could / would be used. It is typical to hear survival mentioned as an issue, but then nothing more. But there is more and you found it. That was my intention.

Conjoint analysis is commonly used in market research, you've surely encountered it at one time or another. It's applied to figure out why a person prefers something, and that something is often a particular product. If you wanted to use conjoint analysis to help you build the most desirable car, you first split car into attributes (price, features, size, speed, etc.), and then apply multiple levels to each attribute (so for price: $20k, $30k, $40k). Then you would survey people on a certain amount of combinations of these attributes and levels to determine what it is they value most. The combinations a respondent is asked to choose from, or rank, at anyone time are usually fairly similar which ensures that their answer will make their preference clear. Obviously this is very simplified example, but I think it gives a good idea.

I don't think conjoint analysis is very applicable to on court NBA analysis, although I am not a very creative mind!

I think it could be for ranking preferences for players with different skill levels on certain things, ages, sizes, contract dimensions, role abilities / expectations etc. Especially if you resist roll-up metrics, this could help with decision-making using the separate considerations perhaps without having to roll them up or decide a priori how much of one is another worth. If players are bundles, evaluate the bundle.

Maybe a list of 13 techniques is a net-casting at possible targets. We've discussed 2 briefly and hopefully more.

Simulated annealing and genetic search are generic optimization techniques. That means that they can be applied to almost any type of regression, but they're particularly useful for sitations where the model is not tractable by analytic methods.

Zero inflated models are pretty much just what the name says. Poisson and binomial would be examples; if you had a process that you would expect to be Poisson but think there might be more 0s than expected due to some additional factor, you would model it with a zero inflated Poisson (ZIP) instead of a regular Poisson. If you were modeling NBA games at the possession level, I guess something could be zero inflated. Rebounds relative to player height (shorter line ups have more 0 board possessions?) maybe?

Thanks for the feedback.

Thurstone Shapley analysis apparently has to do with scaling and specifically unidimensional scaling. I saw this http://www.socialresearchmethods.net/kb/scalthur.php and wondered if it could be done for scout speak about players, professional or pro supplemented by internet amateurs. There can be wisdom in words as well as numbers. Probably some unmined wisdom in specific evaluations and in the aggregate. This technique might be of some organizing and evaluative use, especially if you organize and evaluate the folks saying these things, what they "know", what they value most / least, what they are inadequately aware or knowledgeable about, etc.

AcrossTheCourt wrote:I'll knock one off the list for you: survival analysis deals with, well, the death rates for a population. Death can mean many things, like a failure. It deals a lot with a question like, How much of the population will survive after T amount of time? I'm not well-versed in this field, however. I think it can be applicable to the NBA for something like measuring how long players play in the NBA and when they retire. And it looks like I was right....

If you google NBA and survival analysis articles do come up, like this one:
http://repositories.lib.utexas.edu/hand ... 12-05-5631

"Bayesian hierarchical parametric survival analysis for NBA career longevity"

I had done some very basic survival analysis stuff last April-ish that I recently found:
https://twitter.com/knarsu3/status/572442903735377920
https://twitter.com/knarsu3/status/572443694588157952

As Evan Z pointed out, it should actually read: "probability for possessions not ending in shot clock violations or turnovers". And shot clock violations really should be included- thought I had, will have to go back and look.

But this is definitely a tehnique/analysis I want to go back and re-explore with the SportVu data being made public. I had originally done this with the Vantage data but with a larger dataset from SportVu, would definitely make sense to go back and look at this again. Basically what I did above was a starting point. This is definitely applicable to the shot clock though. The 0 left on the shot clock would be "death".