2021-22 team win prediction contest

[Mike G]

"41 for everybody" continues to win the contest. This is aggravating for a number of reasons. One is that absolute avg errors are starting small and will grow all season -- indicating that regression toward the mean is too much at b-r.com
Another issue is that of course the most-regressed predictions will be favored at this point, and the contest will be artificially rearranged as the season progresses.

So I've got a de-regressed version of the team projections. Still using b-r.com's number, we can readily double the difference from 41:
BR*2-41 gives a mostly-performance-based projection; still with some regression toward .500

Code: Select all

west   W      east   W
Uta   62      Mia   71
Por   51      Chi   59
Den   50      NYK   55
GSW   49      Phl   51
Min   44      Was   50
Sac   43      Cha   46
Mem   38      Tor   46
SAS   38      Mil   41
LAC   38      Brk   41
LAL   33      Atl   37
Phx   33      Ind   36
Dal   31      Cle   34
NOP   30      Bos   33
Hou   23      Det   29
OKC   14      Orl   21

Compare to the lukewarm projections currently at b-r.com
https://www.basketball-reference.com/fr ... _prob.html

And relative to these, our contest (with a few dummy entries) would look like this:

Code: Select all

.  avg err   rmse         avg err   rmse
cali   8.7   10.3      dtka   9.4   11.4
21re   8.9   10.7      shad   9.4   11.0
bpmW   9.0   10.9      perW   9.4   11.3
avgA   9.0   10.6      Crow   9.6   11.1
WShr   9.1   10.8      emin   9.6   11.5
5.38   9.1   10.8      4141   9.7   12.2
trzu   9.2   11.2      vegas  9.8   11.3
lisp   9.3   11.1      2021   9.9   12.0
eWin   9.3   11.0            

Dummies:
2021 is just last year's Wins * 82/72
21re is that number regressed halfway to .500
avgA is the avg of 8 actual APBR submissions plus my 3 extras.
4141 is 41 wins for every team

I'm inclined to phase-in this 'de-regressed' projection to get the dummies in their rightful place a bit sooner, and earlier props to those who got things right. Like today I'm going with 75% straight b-r.com and 25% the less mushy version; adding another 5% every day to the radical side?

Code: Select all

.   avg err   rmse    .25   avg err   rmse
21re   5.71   7.00      dtka   7.57   8.90
4141   6.07   7.63      shad   7.57   8.67
eWin   6.14   7.38      lisp   7.58   8.71
perW   6.22   7.49     vegas   7.65   8.77
bpmW   6.51   7.89      5.38   7.71   8.63
avgA   6.77   7.97      Crow   7.72   9.04
cali   6.98   8.05      emin   8.39   9.95
WShr   7.22   8.44      2021   8.48  10.13
trzu   7.42   9.02            

[Crow]

Ok.

Figuring it out by mid-point of season is a decent goal.

I'm not going to react a lot til then or at least 20 games.

[Mike G]

"Last year's W%" is dead last. Averaging that with 41 leads the pack.

Code: Select all

.   avg err   rmse   .30    avg err   rmse
21re   5.83   7.38      trzu   7.26   9.10
eWin   6.30   7.71      shad   7.33   8.60
perW   6.47   7.95      lisp   7.35   8.66
avgA   6.54   8.03      Crow   7.40   9.00
bpmW   6.57   8.03      dtka   7.44   8.98
cali   6.62   7.97      5.38   7.47   8.55
4141   6.78   8.45     vegas   7.48   8.75
WShr   6.84   8.44      emin   8.28   9.77
.                       2021   8.30  10.02

[eminence]

Hot start

Bummed that Zion is out so long, fun player to watch.

[Mike G]

Code: Select all

.   avg err   rmse   .35    avg err   rmse
21re   6.04   7.78      Crow   7.25   9.16
eWin   6.39   8.01      lisp   7.37   9.11
cali   6.52   8.29      trzu   7.40   9.33
avgA   6.61   8.36      vegas  7.50   8.97
perW   6.66   8.33      dtka   7.57   9.39
bpmW   6.95   8.47      5.38   7.57   9.00
4141   7.02   8.83      emin   8.37  10.01
WShr   7.19   8.70      2021   8.49  10.30
shad   7.21   8.88            

Nov. 3 --
Errors get bigger at the top and less at the bottom. As "predicted", caliban surges to the top.

Code: Select all

.   avg err   rmse   .40    avg err   rmse
21re   6.32   8.06      5.38   7.39   8.92
cali   6.75   8.24      lisp   7.47   9.08
avgA   6.81   8.42      trzu   7.49   9.22
eWin   6.82   8.67     vegas   7.51   8.96
Crow   6.98   8.93      dtka   7.53   9.44
perW   7.04   8.91      4141   7.54   9.67
shad   7.13   8.79      emin   7.92   9.55
bpmW   7.26   8.78      2021   8.15   9.97
WShr   7.30   8.80            

[sndesai1]

i've updated this sheet as well: https://docs.google.com/spreadsheets/d/ ... 1z2U9/edit

using the average of TeamRankings, BPI, Bk-Ref, and 538 current projections to estimate

aside from the late BPI entry, caliban and shadow (along with the "Average") as usual are at the top of the list. 538 RAPTOR is not looking great - we'll see how that goes


also, a couple of charts that don't seem to translate well from excel into google docs


min, max, and average of the entries vs. the current projection (gray box). white rectangle shows teams currently outperforming and black rectangle the opposite


range of current in-season projections along with average (gray box)

[Crow]

4 top scored predictive metrics are all doing poorly. BPI #1. Strange early results.

Let's see in a couple of weeks.

Similarity of entry to measuring stick on general degree of regression to mean right now from last season or "talent" estimates may matter a lot now but firm specific team performance numbers later might not match as closely.

The database indicates that my variance from average was middling. I am ok with that. Some years I've been more extreme, sometimes very successfully, sometimes bitten by being too extreme in a few cases. In the pocket, mildly different, way different, any style COULD work.

If in the pocket is the place to be, then expect Caliban and Shadow to be in top tier. I have to have an edge coming from accuracy in variance from entry average and the most in the pocket entries.

Some substantial differences on the 2 scorecards because of different measuring sticks and number of entries tracked. Leaders will tend to converge with time, though the final results will be independent of these measuring choices.

[sndesai1]

since we're still very early in the season, the fact that BPI knew the outcome of those first two games is very impactful

by the end of the season, i would guess the value of 2 games worth of information to the final RMSE will be an improvement of about 0.1-0.2 - still valuable, but not quite as much as it seems to be at the moment

[Mike G]

Big changes from yesterday.

Code: Select all

.   avg err   rmse   .45    avg err   rmse
21re   6.97   8.58      eWin   7.61   9.50
Crow   7.01   9.01      bpmW   7.69   9.39
avgA   7.06   8.87      lisp   7.73   9.43
cali   7.19   8.72      dtka   7.77  10.09
shad   7.44   8.99      perW   7.78   9.80
vegas  7.46   9.00      trzu   7.79   9.48
5.38   7.53   9.49      4141   8.11  10.52
WShr   7.57   9.33      2021   8.12   9.97
emin   7.57   9.31            

Nov. 5 update:

Code: Select all

.   avg err   rmse   .50    avg err   rmse
Crow   7.09   9.11      eWin   7.84   9.78
21re   7.36   8.86      bpmW   7.91   9.61
avgA   7.38   9.03      dtka   7.92  10.08
cali   7.51   8.97      trzu   7.95   9.51
5.38   7.70   9.51      lisp   8.08   9.46
emin   7.72   9.48      perW   8.08  10.09
WShr   7.73   9.50      2021   8.17  10.10
vegas  7.75   9.28      4141   8.41  10.85
shad   7.76   9.17            

Nov. 6
Teams suddenly resemble their last year W% (2021) relative to our guesses.

Code: Select all

.   avg err   rmse   .55    avg err   rmse
Crow   8.09   10.2     vegas   8.84   10.6
WShr   8.30   10.7      cali   8.85   10.4
21re   8.39   10.1      bpmW   8.87   10.9
avgA   8.49   10.4      trzu   8.88   10.8
2021   8.68   10.9      eWin   9.02   11.2
emin   8.70   10.4      lisp   9.04   10.8
shad   8.74   10.5      dtka   9.11   11.6
5.38   8.79   11.0      perW   9.13   11.5
.                       4141   9.42   12.2

[Mike G]

The worst team in the league by most measures has been the Pistons. They are headed to about 10 fewer wins than our average predicted.
I got minutes estimates from a couple places and averaged them. They are included in this table, along with last season's PER, BPM, and WS/48; and this year's numbers thru 9 games. Expected vs Actual

Code: Select all

Detroit    expected /  last year            actual /  this year
. Pistons      mpg   PER   BPM   WS/48      mpg   PER   BPM   WS/48
Saddiq Bey     29   12.6   -0.8   .082      33   13.3   -1.7   .034
Jerami Grant   32   16.9    0.9   .083      29   14.0   -1.1   .021
Isaiah Stewart 28   16.4   -1.2   .132      24   15.3   -2.4   .096
Kelly Olynyk   21   16.4    1.1   .121      23   20.0    2.9   .125
Josh Jackson   18   12.0   -3.3   .004      22   10.3   -4.5   .009

Killian Hayes  26    5.3   -7.3  -.081      22    6.9   -6.3  -.036
Cory Joseph    19   12.5   -2.1   .069      21   12.7   -2.3   .085
Frank Jackson       13.3   -2.1   .076      19    5.4   -7.8  -.064
Trey Lyles     13   11.4   -1.9   .085      16   18.2    0.4   .113
C Cunningham   31   17.0    1.8   .135      12    2.0  -11.8  -.211

. Pistons      mpg   PER   BPM   WS/48      mpg   PER   BPM   WS/48
Hamidou Diallo 19   14.3   -2.3   .063       9    9.1   -6.0   .004
Luke Garza      1    8.0   -6.4  -.022       4   14.4   -3.1   .125
Saben Lee           13.8   -1.9   .091       3    3.3   -9.0  -.135
R McGruder          16.1    0.7   .121       3    4.2   -6.4  -.024
Isaiah Livers   6    8.8   -5.7  -.008       0         

. totals      240   13.4   -1.6   .068     239   12.1   -3.2   .026
w exp. min.   240   13.6   -1.4   .070               

Minutes allotment is not the problem; hardly anyone is producing as well as expected.
Cunningham may just take Frank Jackson minutes, eventually.

[Mike G]

mayhem continues. "Last year" now in contention, up from the bottom.

Code: Select all

.   avg err   rmse          avg err   rmse
Crow   8.17   10.1      shad   8.81   10.4
21re   8.34    9.8     vegas   8.90   10.4
WShr   8.37   10.5      cali   8.91   10.3
2021   8.44   10.6      eWin   8.96   11.1
avgA   8.56   10.3      lisp   9.05   10.7
emin   8.66   10.3      dtka   9.07   11.4
bpmW   8.71   10.8      perW   9.13   11.4
5.38   8.76   10.9      4141   9.32   12.0
trzu   8.77   10.6   

Nov. 8 update

Code: Select all

.   avg err   rmse          avg err   rmse
21re   7.73    9.3      emin   8.48   10.3
Crow   8.09   10.1      lisp   8.49   10.5
5.38   8.16   10.7      bpmW   8.59   10.5
2021   8.19   10.4      eWin   8.63   10.7
avgA   8.21   10.1      perW   8.65   11.0
trzu   8.29   10.4      cali   8.80   10.2
WShr   8.32   10.4      dtka   8.87   11.3
vegas  8.38   10.1      4141   9.04   11.3
shad   8.42   10.1   

Nov. 9 update:

Code: Select all

.   avg err   rmse          avg err   rmse
21re   7.30    9.0      emin   8.18   10.4
WShr   7.82   10.3      bpmW   8.22   10.3
Crow   7.85   10.1      cali   8.24   10.0
avgA   7.88   10.0      lisp   8.31   10.5
5.38   7.94   10.6      eWin   8.56   10.6
2021   7.95   10.2      perW   8.59   10.8
shad   8.08   10.1      dtka   8.69   11.3
trzu   8.12   10.3      4141   8.87   11.0
vegas  8.13   10.0            

[sndesai1]

another update: https://docs.google.com/spreadsheets/d/ ... OH861z2U9/



charts: https://imgur.com/a/AHkBy4W

[Crow]

I am 4th overall here on mean average error. That is what matters to me. 1st on self-entered "contest" entries, looking beyond a few metrics.

3rd on last update from Mike G, after a few days in lead there.

Long way to finish line.

[Mike G]

Big improvements by everyone.

Code: Select all

.   avg err   rmse          avg err   rmse
21re   6.51    8.8     vegas   7.58    9.6
Crow   7.11    9.6      lisp   7.63   10.0
WShr   7.31   10.0      bpmW   7.70    9.9
avgA   7.33    9.6      eWin   7.81   10.4
2021   7.40    9.6      trzu   7.81    9.9
emin   7.48    9.7      perW   7.90   10.8
cali   7.49    9.5      dtka   8.32   11.0
5.38   7.53   10.3      4141   8.54   11.2
shad   7.54    9.6            

Nov. 13 update: a brutal 2 days

Code: Select all

.   avg err   rmse          avg err   rmse
21re   6.83    9.0      cali   7.82   10.0
Crow   7.42   10.0      WShr   7.84   10.5
avgA   7.57    9.9      lisp   7.86   10.2
5.38   7.59   10.4      trzu   8.03   10.1
emin   7.71   10.1      eWin   8.11   10.8
shad   7.72    9.9      perW   8.18   11.1
vegas  7.74    9.9      2021   8.23   10.0
bpmW   7.77   10.3      dtka   8.41   11.2
.                       4141   8.71   11.1

update Nov. 14

Code: Select all

.   avg err   rmse         avg err   rmse
21re   6.66   8.6      emin   7.87   10.0
Crow   7.47   9.8      lisp   7.91   10.1
avgA   7.55   9.7      trzu   8.00   10.0
shad   7.69   9.7      eWin   8.03   10.4
bpmW   7.71   9.9      perW   8.04   10.7
vegas  7.71   9.6      2021   8.06    9.8
5.38   7.74  10.3      dtka   8.42   11.1
WShr   7.75  10.1      4141   8.44   10.7
cali   7.83   9.9            

update 11-14: The separation between 1 and 2 dropped almost 80% overnight.

Code: Select all

.   avg err   rmse         avg err   rmse
21re   7.07   8.7      emin   7.67    9.7
Crow   7.23   9.6      2021   7.73    9.6
vegas  7.50   9.4      trzu   7.94    9.8
avgA   7.54   9.5      lisp   7.96   10.0
shad   7.57   9.5      bpmW   7.96    9.8
5.38   7.58  10.1      eWin   8.12   10.3
WShr   7.61   9.8      perW   8.33   10.6
cali   7.63   9.5      dtka   8.66   11.0
.                      4141   8.85   11.0

[Crow]

21re is a pretty smart but simple strategy.

I try to stay fairly close it (especially in less "understood" cases), but I hope to add more than I lose in specific evaluations. The margin might be thin but sometimes it is more than thin.