2015-16 Team win projections

[Crow]

Not all the predictions on the bigger listing are strict formula based models. At least 15% are not and who knows if or how often formula driven models adjusted results at the end, perhaps using discretion in rounding off or beyond in cases that didn't "look right". For expected results in future a strict formula model with lower RMSE model may be better but for the contest this season every incremental win miss is same 1 win miss. Perhaps there will be 2 "winners" on these different measures. But pretty good chance one prediction will win both.

[kmedved]

Errors for last year:



Not quite normal, especially on the downside, which I imagine is a Durant effect.

[Mike G]

At about 20% of the season. Absolute avg errors relative to b-r.com's projections:

Code: Select all

KF   6.39     DF    6.8      nr    7.6
tzu  6.42     Crow  7.0      BD    7.9
Cal  6.5      fpli  7.1      taco  8.1
km   6.6      itca  7.2      DrP   8.1
DSM  6.7      yoop  7.2      Dan   8.1
bbs  6.7      snd   7.2      rsm   8.2
AJ   6.8      MG    7.3      EZ    8.2

[kmedved]

Post-Thanksgiving update here as well:

[Mike G]

In just 2 days, almost everyone improved by 0.2 to 0.6

Code: Select all

KF   6.16      DSM  6.6      nr   7.4
Cal  6.24      bbs  6.7      BD   7.6
tzu  6.24      snd  6.8      rsm  7.7
km   6.5      fpli  6.8      DrP  7.8
AJ   6.5      yoop  6.8      taco 7.9
DF   6.5      itca  6.9      Dan  8.0
Crow 6.6       MG   7.0      EZ   8.0

[kmedved]

Update after Friday:

[kmedved]

538's win projections have gone live: http://projects.fivethirtyeight.com/2016-nba-picks/

I've added them to the average used in the standings I'm posting. I've also shifted to averaging the B-Ref projections and the SRS-method I was using, before adding them into the bigger average, because the SRS method and the B-Ref method are really the same thing, so we were ending up with SRS-methods overweighted:

[ampersand5]

can someone explain to me the 538 love for ELO? Specifically, why it would be preferable to something like SRS for their team quality blend?

Thanks

[Nate]

can someone explain to me the 538 love for ELO? Specifically, why it would be preferable to something like SRS for their team quality blend?

Elo calculations are much easier to follow since they happen 'one game at a time', that makes for more intelligible narratives. Elo is also a bit more statistically sophisticated so it's less likely to overfit, has an implicit recency bias that seems to work quite well, and there's a natural way to incorporate prior strength estimates. 538 is also an "Elo shop" - it's a model they're familiar and experienced with, so they will hopefully be more aware of potential pitfalls.

P.S. Elo is a guy's name, so you'd normally only capitalize the E.

[Mike G]

A month ago, AJbaskets was middle of the pack, and he's now surged into the lead.

Meanwhile, here's a summary of our avg predictions, ranked by difference from b-r.com's current projections.
http://www.basketball-reference.com/fri ... s_forecast
From better than expected, to worse than expected:

Code: Select all

err  tm  avg  proj    err  tm   avg  proj     err  tm   avg  pro
13  Orl   31   44      4   SAS   56   59      -2   LAL   25   23
13  Cha   36   48      3   Dal   39   42      -5   Okl   55   51
12  Min   27   39      3   Phx   37   39      -5   Phl   22   18
10  NYK   26   37      2   Den   29   31      -6   Was   44   38
9   Ind   40   50      2   Uta   41   44      -9   Cle   56   47
7   Det   36   43      2   Tor   47   49     -10   Mil   39   30
6   GSW   61   67      1   Por   36   38     -10   LAC   54   44
4   Brk   25   30      0   Atl   49   50     -10   Mem   49   39
4   Bos   45   49      0   Chi   47   46     -15   NOP   44   30
4   Mia   43   47     -1   Sac   36   35     -22   Hou   54   32

In double-digit errors it's East 7, West 2
Did anyone notice Charlotte becoming the 6th best team in the league?

[Dr Positivity]

The Scott Skiles effect has been impressive yet again, although usually it takes until his 2nd season to click in

[Mike G]

Probably just this once: Exponent 1.5, relative to b-r.com projections:

Code: Select all

km   7.0      Crow  7.7       BD   8.7
KF   7.1      yoop  7.8       nr   8.8
AJ   7.2      fpli  7.9      DrP   9.0
Cal  7.2      snd   7.9       EZ   9.1
tzu  7.2       MG   7.9      Dan   9.1
DF   7.3      itca  8.1      taco  9.4
bbs  7.3      rsm   8.3      15py 10.2
DSM  7.4                  

[caliban]

Let me know if I missed someone, there's a number wrong or you have a specific blend you want to be displayed for the 30 games in edition

[Crow]

Thanks for the newest detailed tracking and all the tracking by several folks. Makes it an interest ongoing story.

As far as blends go, I'd ideally like to see how the best blend of the top 5 entrants last season (optimal explanation of last season) who continued this season is doing this season with that same blend. Maybe regress the weights a bit towards more equality for future performance. My entry relies heavily on best 4 from last season (and unfortunately Vegas) but the blend was not systematically optimized and I did add some quick adhoc subjective adjustments.

[kmedved]

Sorry, are you saying your entry "peeked" at the entries of the top finishers last year?