Predictions 2014-2015

[AcrossTheCourt]

Okay but what's wrong with using a multi-year RAPM? All in all, it's just a player estimate, and including multiple years means you have more samples to work with and there are more instances of guys switching teams. If you want to call out that method, then show how single year NPI RAPM is better over a larger number of seasons.

[sideshowbob]

Is single-year RAPM informed by a prior considered multi-year? I don't see how that's the same as an actual multi-year set.

[permaximum]

Okay but what's wrong with using a multi-year RAPM? All in all, it's just a player estimate, and including multiple years means you have more samples to work with and there are more instances of guys switching teams. If you want to call out that method, then show how single year NPI RAPM is better over a larger number of seasons.

What you are saying is correct but it's not related to the problem here.

The problem is simply a mismatch. You can't use 1 year of player-tracking data as X and multiple years of RAPM values as Y. In short, he tried to estimate multiple years (2014-2013-2012...not sure how many years) of RAPM values with one year (2014) of player tracking data.

Is single-year RAPM informed by a prior considered multi-year? I don't see how that's the same as an actual multi-year set.

Although it has the memo of previous years it can't be considered multi-year. However, RAPM informed by previous years wouldn't be the ideal data to develop a PTPM. He simply needed a single-year NPI-RAPM. (When this season ends he can use 2 years of PT data and 2-year RAPM values found by one big ridge regression.)

Anyways, I don't think that's the case here. The data he used is simply multi-year RAPM where the recent years weighted more. J.E. can confirm this by spending 5 seconds if he wanted. He doesn't care about it I think.

[willguo]

one thing i don't get is how you get 15-15 for the espn summer forecast? if espn had 56 for the clippers and the closing line is also 56, shouldn't you consider it as no bet?

You're totally right - should be 14-12 on all plays vs the close, 11-1 with 1 game diff, 9-4 with 2 game diff, 4-3 with 3 game diff.

[sideshowbob]

The data he used is simply multi-year RAPM where the recent years weighted more. J.E. can confirm this by spending 5 seconds if he wanted. He doesn't care about it I think.

I was under the impression that JE refers to single-year prior informed as "Pure RAPM", as this is what I've seen him call it before. I suppose there's a degree of labeling confusion with all the sets that have been made pubilc.

EDIT: and yet the set on JE's site is labeled "pure RAPM" on the home page but listed as 2-year on it's actual page. Again, confusion.

[permaximum]

The data he used is simply multi-year RAPM where the recent years weighted more. J.E. can confirm this by spending 5 seconds if he wanted. He doesn't care about it I think.

I was under the impression that JE refers to single-year prior informed as "Pure RAPM", as this is what I've seen him call it before. I suppose there's a degree of labeling confusion with all the sets that have been made pubilc.

EDIT: and yet the set on JE's site is labeled "pure RAPM" on the home page but listed as 2-year on it's actual page. Again, confusion.

To put an end to this discussion since J.E. doesn't want to confirm or deny anything about it, here's what he wrote about the exact data the PTPM developer used. Simply it's multi-year, with coach, age, leading up adjustments etc. It just doesn't use box-score data. It's completely wrong to use these RAPM values to develop PTPM from 1 year of player tracking data.

JE's post about the Pure RAPM we're talking about: http://www.apbr.org/metrics/viewtopic.p ... apm#p18810

No priors, but more years than before, which is essentially almost the same as using priors

Aging Curve effects are incorporated.

Effort Curve (Line?) effects (the topic here) are incorporated.

I derive the coach ratings from one large regression over all data. When I compute the ratings for this season the coaches aren't a variable anymore, their rating just gets added in.

[DSMok1]

Using Multi-year RAPM would almost certainly be better than using single year RAPM, even though the PT data is for a single year. Single year RAPM has so much noise and bias, the addition of multiple years will be helpful to the overall accuracy of the PTPM. Single year RAPM should probably not really be used for anything--collinearity and lack of interconnectedness of the data is just too dominant.

[permaximum]

Using Multi-year RAPM would almost certainly be better than using single year RAPM, even though the PT data is for a single year. Single year RAPM has so much noise and bias, the addition of multiple years will be helpful to the overall accuracy of the PTPM. Single year RAPM should probably not really be used for anything--collinearity and lack of interconnectedness of the data is just too dominant.

1. PTPM developer thought it was 1-year RAPM, then he used it. So, he didn't need multi-year RAPM one way or another.
2. How does single-year RAPM has more bias than JE's pure RAPM? It's completely the opposite.
3. What you are saying is completely invalid from a statistical standpoint. You can't regress 3, 4 or 5 years of RAPM values as response on 1 year of PT predictors.
4. I hope you didn't mismatch box-score data with 14-year RAPM developing BPM. I have a feeling you did the right thing and used 2001-2014 box-score data

[ampersand5]

it would be interesting if the users who submitted projections could adjust their calculations for the actual minute distribution of the 2015 NBA season.

[DSMok1]

Using Multi-year RAPM would almost certainly be better than using single year RAPM, even though the PT data is for a single year. Single year RAPM has so much noise and bias, the addition of multiple years will be helpful to the overall accuracy of the PTPM. Single year RAPM should probably not really be used for anything--collinearity and lack of interconnectedness of the data is just too dominant.

1. PTPM developer thought it was 1-year RAPM, then he used it. So, he didn't need multi-year RAPM one way or another.
2. How does single-year RAPM has more bias than JE's pure RAPM? It's completely the opposite.
3. What you are saying is completely invalid from a statistical standpoint. You can't regress 3, 4 or 5 years of RAPM values as response on 1 year of PT predictors.
4. I hope you didn't mismatch box-score data with 14-year RAPM developing BPM. I have a feeling you did the right thing and used 2001-2014 box-score data

1. I'm pretty sure he knew that it wasn't true 1 year RAPM.
2. Single year RAPM is basically noise. The signal is just too small to be of any use.
3. It's better than regressing 1 year of PT predictors onto noise, in my opinion.
4. Yes, I did the right thing.

[permaximum]

1. I'm pretty sure he knew that it wasn't true 1 year RAPM.
2. Single year RAPM is basically noise. The signal is just too small to be of any use.
3. It's better than regressing 1 year of PT predictors onto noise, in my opinion.
4. Yes, I did the right thing.

1. I read his explenation for the metric in this link and it sounded like he thought it was single-year 2013-14 RAPM. If he confirmed to you that he used multi-year RAPM on purpose, OK. Still it doesn't make sense to use multi-year RAPM since there was no multi-year PT data.
2. I pointed out "bias". It's bias is nowhere close to JE's pure RAPM. As for noise, yes there's considerable noise in single-year RAPM but not as much as you think.
3. I don't think so. It has to be tested. Still, it's statistically invalid. There's no PT data to capture previous years' effect on those RAPM values. How can you regress those values on 1 year of PT? If noise is that much of a problem he should have used RiRAPM instead of multi-year.
4. I know BPM's predictive power is better compared to other box-score metrics I have seen so far. Still I wish you didn't use ast*reb. How much did it help anyway? When you cross that line, it feels like there's endless possibilities to check. Ast*reb, blk*stl, sqrt(to), ftm/fgm, reb*stl, fgx/orb...

[DSMok1]

1. I read his explenation for the metric in this link and it sounded like he thought it was single-year 2013-14 RAPM. If he confirmed to you that he used multi-year RAPM on purpose, OK. Still it doesn't make sense to use multi-year RAPM since there was no multi-year PT data.
2. I pointed out "bias". It's bias is nowhere close to JE's pure RAPM. As for noise, yes there's considerable noise in single-year RAPM but not as much as you think.
3. I don't think so. It has to be tested. Still, it's statistically invalid. There's no PT data to capture previous years' effect on those RAPM values. How can you regress those values on 1 year of PT? If noise is that much of a problem he should have used RiRAPM instead of multi-year.
4. I know BPM's predictive power is better compared to other box-score metrics I have seen so far. Still I wish you didn't use ast*reb. How much did it help anyway? When you cross that line, it feels like there's endless possibilities to check. Ast*reb, blk*stl, sqrt(to), ftm/fgm, reb*stl, fgx/orb...

RE: 4: I checked the most obvious of the endless possibilities. None of them added much except the ones I included. Ast*trb helped... several percent on the R^2. Pretty big improvement at the margins.

[permaximum]

So what's the end result here? I guess PT-PM won and Crow came second...

I bet on Crow before since I like his blending approach and couldn't see someone surpassing him but I didn't think of any player tracking metric and it's potential prediction power then. It seems this is where the focus will be on from now on.

[ampersand5]

So what's the end result here? I guess PT-PM won and Crow came second...

I bet on Crow before since I like his blending approach and couldn't see someone surpassing him but I didn't think of any player tracking metric and it's potential prediction power then. It seems this is where the focus will be on from now on.

I think the end result can only be truly discerned if we insert the 2014-2015 season minute distribution into the metrics.

I don't think that people on here put too much thought into their projected minutes and even if they did, its too arbitrary (with things like injuries).

[permaximum]

So what's the end result here? I guess PT-PM won and Crow came second...

I bet on Crow before since I like his blending approach and couldn't see someone surpassing him but I didn't think of any player tracking metric and it's potential prediction power then. It seems this is where the focus will be on from now on.

I think the end result can only be truly discerned if we insert the 2014-2015 season minute distribution into the metrics.

I don't think that people on here put too much thought into their projected minutes and even if they did, its too arbitrary (with things like injuries).

Agree with you but I don't see many people will share their pre-season minute distribution and we can't confirm if the shared ones will be the actual they used or not.