Nba projections model: some questions

[patron]

Hey everyone. I am new to the forum and sort of new also to APBR metrics. I've been looking at them to try and predict scores and matchup advantages since last year and I am getting pretty familiar mostly with the basic concepts like the four factors. What brings me here is that, with some help, I have decided recently to try and create a model of my own to predict scores. My main objective is to reflect the match-up influence and specially the four factors. I have already done lots of tweaking and enhancing but the results are far from where I want them to be.

Right now the model is all about predicting who get's the most out of the possessions. Basically I'm trying to predict pace in a very simple manner (averaging them, with home/away included) and from that point on, use four factors (with FTA/FGA) to predict the amout of shots and free throws a team will take. I am actually liking some of these early results, but it's clear that the scores are far too dog friendly most of the time and also pointing towards the Under.

Firstly I would like to know if you, certainly more experienced regarding advanced basketball statistics, could tell me if there is any probability that what I'm doing now, even as a start, is a step in the right direction.

Secondly, I would like to know how to adjust the projections to the fatigue factor, the home court factor and also injuries (on key players of course). Right now im simply giving 4 points of advantage to the home team as, for what I've heard and read, that's the most used number. I'm also thinking about adding a 1/1.5 edge against teams on b2bs on the Road, that travel West to East, or that are playing 3rd game on 4 nights.

Any answers and data I could be pointed on would be great. Thanks in advance.

[DSMok1]

For the impact of fatigue, check out the last post in this thread:

http://godismyjudgeok.com/DStats/APBRme ... =2661.html

[Crow]

When you look at the 4 Factor data, you have the average performance of a team on each factor but you could drill down to the average performance against good, average and poor teams and use that in your calculation instead of the overall average, if you are not already doing so. So if one team is good on own eFG% and the other is average on eFG% you might use the average performance of the defense against good own eFG% teams and the average performance of the offense against average eFG% allowed teams and maybe average them (and perhaps look to actual league-wide performance data and see how those interactions have done over time and maybe shade the prediction one way or another a bit since the team level data is small sample). And then you might go further into how a specific factor match-up advantage affects other factors in those match-ups. One level of interaction or two or three. Dean Oliver did something perhaps along this line years ago and called it "roboscout" and never reveal any details in public. Don't know how deep or accurate it was. You could also use last season's data to some degree, especially this early.

The home court advantage could be tweaked from 4 pts to a more accurate average based on other studies out there or to a team specific value that might be refined slightly as the season goes on.

Player injuries / absences could be considered. Starters vs starters lineup matchup might be given some additional attention. You could look at factor level consistency data game to game and compute a probabilistic based match-up projection rather than using an average. You could look at end of game "clutchness", and if you do that you could look at the distorting effects of garbage time too on the average data. Plenty of ways to tweak it, depending on how time you want to put into it.

[v-zero]

Four points HCA is miles too high. The average when team strength is accounted for and rest is accounted for is about 2.8 points. A team playing back-to-back against a team that isn't loses about 1.8 points on the spread.

[DSMok1]

Four points HCA is miles too high. The average when team strength is accounted for and rest is accounted for is about 2.8 points. A team playing back-to-back against a team that isn't loses about 1.8 points on the spread.

When accounting for rest days (see my previous post) HCA is about 3.24.

[v-zero]

When accounting for rest days (see my previous post) HCA is about 3.24.

When constructing teams as they appeared using player models and taking into account rest and only accepting what is testably significant (probably the difference here) I found that it's almost exactly 2.8, YMMV...

[DSMok1]

When accounting for rest days (see my previous post) HCA is about 3.24.

When constructing teams as they appeared using player models and taking into account rest and only accepting what is testably significant (probably the difference here) I found that it's almost exactly 2.8, YMMV...

2.8 Pts/100 Possessions?

What do you mean by testably significant? You mean you assume zero effect of rest unless it's statistically significant? I wouldn't limit it that way (obviously).

I did not include individual players when I did that research, so your system is likely better. The difference may be that coaches pull players when losing games, exaggerating the effect.

[v-zero]

2.8 Pts/100 Possessions?

What do you mean by testably significant? You mean you assume zero effect of rest unless it's statistically significant? I wouldn't limit it that way (obviously).

I did not include individual players when I did that research, so your system is likely better. The difference may be that coaches pull players when losing games, exaggerating the effect.

I work with per-minute data as from what I've found players have their own intrinsic pace and so a per-minute stat makes more sense for my stuff - I looked at whether games with higher pace had a higher HCA but didn't find anything significant. On significance: if it reduces the BIC of the model then I accept it, otherwise I reject it. I wasn't trying to say mine was better, only stating that I was speaking from what I've found rather than what I sort of knew/guessed at/had seen around.

[bbstats]

EDIT: Ignore me, I'm stupid.

[patron]

First of all let me thank you all for the answers, even you bbstats

When you look at the 4 Factor data, you have the average performance of a team on each factor but you could drill down to the average performance against good, average and poor teams and use that in your calculation instead of the overall average, if you are not already doing so. So if one team is good on own eFG% and the other is average on eFG% you might use the average performance of the defense against good own eFG% teams and the average performance of the offense against average eFG% allowed teams and maybe average them (and perhaps look to actual league-wide performance data and see how those interactions have done over time and maybe shade the prediction one way or another a bit since the team level data is small sample). And then you might go further into how a specific factor match-up advantage affects other factors in those match-ups. One level of interaction or two or three. Dean Oliver did something perhaps along this line years ago and called it "roboscout" and never reveal any details in public. Don't know how deep or accurate it was. You could also use last season's data to some degree, especially this early.

The home court advantage could be tweaked from 4 pts to a more accurate average based on other studies out there or to a team specific value that might be refined slightly as the season goes on.

Player injuries / absences could be considered. Starters vs starters lineup matchup might be given some additional attention. You could look at factor level consistency data game to game and compute a probabilistic based match-up projection rather than using an average. You could look at end of game "clutchness", and if you do that you could look at the distorting effects of garbage time too on the average data. Plenty of ways to tweak it, depending on how time you want to put into it.

I think I understand what you mean, instead of averaging the regular efg% own and def, average the efg% own against bad efg% teams and so on. My question is, where can I find this data? Is it available? Also I didn't get the last part. Keep in mind I'm a newbie in this!

For the impact of fatigue, check out the last post in this thread:

http://godismyjudgeok.com/DStats/APBRme ... =2661.html

I found those stats to be really helpful even if its only three season's data that's already a significant amout. And it's also really easy to use I believe. My doubt is regarding your discussions as to HCA? I dind't quite understand, you say its 3.24, but does that mean 3.24 everytime, and then add to it the fatigue impact? Also is it 3.24 points or 3.24 RTG? Should I add to the home team and subtract from the away team?

Once again thanks in advance. Also, where can I find a way to test my system against the past years? those regression tools I've heard?

[DSMok1]

For the impact of fatigue, check out the last post in this thread:

http://godismyjudgeok.com/DStats/APBRme ... =2661.html

I found those stats to be really helpful even if its only three season's data that's already a significant amout. And it's also really easy to use I believe. My doubt is regarding your discussions as to HCA? I dind't quite understand, you say its 3.24, but does that mean 3.24 everytime, and then add to it the fatigue impact? Also is it 3.24 points or 3.24 RTG? Should I add to the home team and subtract from the away team?

Once again thanks in advance. Also, where can I find a way to test my system against the past years? those regression tools I've heard?

The way the regression was set up, those are all in terms of rating (pts/100 possessions). In addition, all of the effects are additive. Add the effect of rest for one team, the effect of rest for the other team, and the effect of home court. In terms of offensive and defensive efficiency for home court advantage, calculate first the expected offensive ratings for each team and then split the HCA effect between the two (+3.24/2 to the home team, -3.24/2 to the road team).

I set this all up, using ASPM and projected minutes + rest and home court to generate an expected score, some time ago, but I don't have it posted anywhere right now.

[patron]

All right so I did read that correctly at first.

Perhaps the reason my model keeps giving me under value its because I'm not using the OFF/DEF Ratings of a team? Can there possibly be something here that I'm missing?

thanks again dsmok1

[DSMok1]

I think you should read this primer first: http://www.rockytoptalk.com/2008/2/27/165520/578

[Crow]

"I think I understand what you mean, instead of averaging the regular efg% own and def, average the efg% own against bad efg% teams and so on. My question is, where can I find this data? Is it available? Also I didn't get the last part. Keep in mind I'm a newbie in this!"


Just divide the teams into blocks of 10: good= top 10 teams on that stat, average= middle 10, poor= bottom 10.

82games w-l profile pages are an example of this kind of sorting / thinking.
Example: http://www.82games.com/1112/1112MIA4.HTM