2015-16 Team win projections

[kmedved]

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

[Mike G]

Caliban, what are these numbers? I see the b-r.com projected wins at top. And for each entrant, our predicted wins are in green. But what is the row of numbers below that?

[caliban]

Caliban, what are these numbers? I see the b-r.com projected wins at top. And for each entrant, our predicted wins are in green. But what is the row of numbers below that?

Thank you. Mixed up a row when switching from a horizontal view to fit the forum better. Updated. The row below is the variance.

[Mike G]

OK, thanks. But the average errors look the same, and fpliii should not be on top, according to my chart.
You have his Cha guess at 44, but he actually guessed 35; Chi at 44, but he said 49; LAC 56 vs 54; Dal 40 vs 39; Min 30 vs 22 ...

I do love the chart, though. All the info is there, and it fits on the screen!

[caliban]

Hopefully we're closing in on something, thanks again sir.

esthetic overfitting?

[BasketDork]

Wow, everyone off on Houston by an average of 21 wins. Big ol' fly in the ointment so far...

[fpliii]

I feel like Houston should figure things out, New Orleans on the other hand, I'm more nervous about. Last I checked, all their lineups with Gordon were problematic.

Minnesota looks to be better than anticipated. It's a huge credit to KG that he's (presumably) doing such a great job of mentoring the younger players there.

[kmedved]

Added average misses here:

[Crow]

If there was no interest in optimizing a blend based on last season performance, a simple even blend of top 4-6 performers this season might be good to see in comparison with the components by themselves and the average of all projections.

[Mike G]

Some of the projections are getting pretty extreme, i.e., close to what we get by just adding current wins + pythagorean W% to remainders of season.

Code: Select all

GSW   73      LAC   46      Hou   33
SAS   68      Dal   44      Sac   32
Okl   57      Uta   44      Was   32
Cle   55      Det   44      Mem   32
Cha   54      Chi   44      Den   28
Ind   53      Orl   43      Mil   26
Bos   52      Por   39      Brk   26
Tor   52      Phx   39      NOP   24
Mia   46      Min   37      LAL   16
Atl   46      NYK   35      Phl   10

Relative to these, our avg errors are all larger:

Code: Select all

KF   7.74      rsm   8.61      yoop   9.40
km   7.74      fpli  8.67      DrP    9.45
tzu  7.76      Crow  8.69       BD    9.58
bbs  7.85      snd   8.88       nr    9.61
DSM  7.97       MG   8.93       EZ   10.06
AJ   8.02      itca  9.00      taco  10.47
Cal  8.03      Dan   9.07      15py  11.51
DF   8.08                  

We are also seen to be breaking into upper, middle, and lower tiers.

[Mike G]

Yesterday, the Bucks were projected by b-r.com to finish with 28.7 wins.
Today, they're at 31.2 -- an uptick of 2.5 -- just for beating the Warriors.

They're still the 3rd worst team in the East, and a 17-point underdog vs GSW.
But they managed to be >30 points better than expected last night.
In the course of a season, every team can expect to meet one other team that's having its best game of the season.

[caliban]

added your idea to the spreadsheet Crow. I did test with a higher then 15% consideration to the current top5 but it looked too good pretty quickly in comparison to the angles kmedved so excellently already covers.

with a working template I'll try to make this a weekly update before the Sunday games. The last couple of days has been particularly kind to rsmth, bbstats (who is now in the front seat) & nbacouchside.

[kmedved]

I've also added rankings for the 2015 Top 5 (based on RMSE here: viewtopic.php?f=2&t=8633&start=255) as well as the current average of the top 5 (also based on RMSE).

[Wrong chart included : see 2 posts down]

[Crow]

Thanks to both. Hot performance by 2016 top5 was kinda expected. Surprised a bit that 2015 top5 just barely beats 2015 average of all but regression to the mean is a powerful friend on this it seems. Elite performance on one set of data and volume of predictors creating a regressed average both matter, maybe the later more when the number is healthy and the gap in quality isn't huge or proven reliably over long time as huge.

[kmedved]

Your question got me wondering, and I found a coding error in my prior set above (I had AcrosstheCourt and AJBaskets confused for inclusion into the 2015_Top5 Set). This is corrected:



One other process note: I am using the raw win projections as compiled here and linked in the Google Docs (plus a few stragglers posted later in the thread): viewtopic.php?p=25444#p25444

I am not using the scaled projections. The difference is for people who didn't have their wins sum to up 1230, which includes the Westgate line, but also curiously the 538 numbers? I have no sense as to whether all the numbers should be scaled for contest purposes.