2023-24 team win projection contest

[v-zero]

They just take the current SRS or efficiency metrics (depends which you use) of each team and give you their expected wins from an 82 game season.

[Mike G]

Right. But we are predicting actual wins here.
b-r.com adds current wins to expected future wins; taking into account SRS and future schedule strength.

Winning close games may or may not be entirely 'luck'. Blowout games skew SRS a bit out of proportion to their significance as a predictor.

Regarding terminology: I think of 'projection' as just extrapolating current performance into the future. Players and teams come into a season somewhat changed from the previous year; and after 12 to 15 games, last year is irrelevant.

'Prediction' is based on best information available, statistical or otherwise. Like a weather forecast -- 20% chance of rain, but it's raining, so their prediction was off.

Depending on who initiates the 'contest' thread, it may be called Prediction or Projection.

[v-zero]

I'm saying it will act as a luck adjustment, as it removes the clustering of luck from game scores. Winning close games is more luck than anything else, though certainly aspects of composure and coaching will apply.

A 20% prediction of rain and then it rains doesn't mean their prediction was 'off', it equally just means the dice fell in the 20%. Accuracy isn't a good metric for categorical prediction, this is why we use logloss/cross-entropy.

Last year's data never becomes irrelevant in the NBA, it has predictive value all the way through the year. It diminishes but it isn't zero.

SRS is overly aggressive, a version using ridge regression would be better. IIRC JE has long wished BBRef would replace their SRS with a ridge version, I agree.

[Mike G]

Last year's data never becomes irrelevant in the NBA, it has predictive value all the way through the year. It diminishes but it isn't zero.

I looked at a few years' standings and SRS -- after 5 games, after 10 etc. -- and checked how much each of these components (last year and this year) would correlate to the final Win totals:
(current W%*G + PythW%*RG)*n + (1-n)*avgA
RG = remaining games; avgA is the avg of APBR predictors.
Pythagorean W% can use SRS or simple MOV, it doesn't much matter.

SRS tends to stabilize within 10-12 games, and highest correlation with ultimate Wins requires n=1.0 at around 15 games.
Current team success takes all the weight from predictions based on last year's performance, with 20% of a season played.

One could maintain that we were well "off" to collectively guess OKC would win around 40 games. Of course there's a dice chance they do much better (or worse).

[v-zero]

At a player level last year's data remains significant all the way to the end of the season, and predictions are made using player level statistics in general.

Also it would be better to run that regression attempting to predict the remaining games than the total, that would provide a more accurate split of the value of past to present data. Otherwise the collinearity between "current W%*G" and the final standings invalidates the regression.

To put it simply: the best possible model to predict the final standings today, is the best model to predict the remaining games, to which you then add the results of games for which you already know the result.

[Mike G]

You are correct. Formula would change from :
(current W%*G + PythW%*RG)*n + (1-n)*avgA

to:
current W%*G + (PythW%*RG)*n + (1-n)*avgA*(RG/82)

Maybe I can find that old exercise, in my files or what I reported here.

[v-zero]

Believe it should be:

current W%*G + (PythW%*RG)*n + (1-n)*avgA*(RG/TG)

Where TG is the total games in the year, so usually 82 but not always.

[Mike G]

Correct yet again haha. Thanks.

[Mike G]

Code: Select all

.   avg err  rmse   r^2      .   avg err  rmse   r^2
medi   4.89   6.2   .78      emin   7.01   8.5   .55
ChKl   5.77   7.7   .64      EBPI   7.22   8.4   .61
EExp   5.87   7.1   .71      KPel   7.49   9.3   .51
vegas  6.05   7.3   .71      DRKO   7.58   9.1   .52
IanL   6.19   7.3   .67      ncs.   7.80   9.5   .44
avgA   6.35   8.0   .66      DQin   7.87   9.4   .45
trzu   6.44   7.9   .67      LEBR   8.04   9.6   .46
Crow   6.89   8.3   .60      23re   8.14  10.0   .46
dtka   6.93   8.3   .61      eWin   8.20  10.1   .47
vzro   6.97   8.4   .56      NuFi   8.49  10.0   .39

[Crow]

I am currently 3rd in "the contest". That is probably the high water mark for me this season.

[Mike G]

We are "all" about 1 MAE worse than last year at this time. I was wondering if there was too much preseason consensus this year, or just what could be going on.

For some teams, our guesses were quite closely packed, and for others not so much. So I wonder if the 'spread' between highest and lowest guesses would be correlated to avg error.
Here I am defining "high" guess as the avg of the top 2. This mitigates extreme outliers without excluding them.
Same deal for "low" guess; and 'spread' is just high minus low.

At this point, relative to the b-r.com projections, 2/3 of teams are completely outside our guess ranges; possibly excepting a single guess. *
We were all * too low on the top 11. The middle 10 fall within our spread. Bottom 9 are worse than any * of us guessed.

Code: Select all

tm    high    low   proj    spr   over
OKC   45.7   34.1   57.4   11.6   11.7
LAC   44.6   40.4   54.2    4.2    9.6
Min   50.6   40.0   57.5   10.6    6.9
Ind   38.5   32.2   44.9    6.3    6.4
Orl   40.6   32.8   46.9    7.8    6.3
Bos   59.3   53.3   64.3    6.0    5.0
Dal   43.5   39.0   46.8    4.5    3.3
Hou   37.5   26.9   38.7   10.6    1.2*
NOP   48.1   43.1   49.0    5.0    0.9
Cle   52.1   46.3   52.8    5.8    0.7*
Mia   44.2   40.8   44.7    3.4    0.5
               
tm    high    low   proj    spr    in
Sac   48.0   37.9   45.8   10.1    2.8
NYK   53.2   42.8   49.3   10.4    1.3
Den   57.2   46.5   53.0   10.7    1.2
Phl   52.9   41.0   47.9   11.9    0.9
Chi   40.8   34.0   37.8    6.8    0.4
GSW   50.6   39.0   44.7   11.6   -0.1
Uta   40.6   32.7   36.0    7.9   -0.7
Mil   54.5   47.1   49.5    7.4   -1.3
LAL   48.1   39.6   42.4    8.5   -1.5
Phx   48.6   45.0   45.0    3.6   -1.8
               
tm    high    low   proj    spr   under
Por   36.1   24.6   22.6   11.5    2.0*
Brk   42.9   36.6   33.2    6.3    3.4
Atl   44.8   39.3   35.3    5.5    4.0
SAS   35.6   23.7   17.6   11.9    6.1
Det   31.2   21.3   15.1    9.9    6.2
Tor   42.0   38.0   31.5    4.0    6.5
Was   33.8   23.7   16.0   10.1    7.7
Cha   35.9   28.6   19.8    7.3    8.8
Mem   46.7   43.1   30.2    3.6   12.9

* Mediocre is the extremist responsible for these, and I suspect he has a time machine. Should be outlawed.

The wider spreads are more likely to have the b-r projection falling within them. The Dubs are performing smack in the middle of their broad range, while the Suns are at the very bottom of their very narrow spread -- 2 of us guessed 45.

[Crow]

Can't go back and be personally blind to season to date; but looking at list I feel the list divides into: "really trying", "peaked last year" and "not equipped to do much / not effectively trying hard". How many of these descriptors could have been appropriately given last fall? Probably a lot. I am probably over retrofitting the descriptors, but I feel like about 25 could have been given, though not exact win totals.

Regression is going to miss low on the try hards and over-estimate the not try effectivelys. Probably do pretty well with the peaked last years.

I guess some teams might not be clearly in one of the 3 tiers. But it may help to try to use the framework for adjusting predictions, instead of going all toward middle. I kinda do this some but maybe not as firmly and comprehensively as suggested by this season's results. Which may or may not be "normal".

[Mike G]

Code: Select all

.  avg err   rmse   r^2      .  avg err   rmse   r^2
medi   4.95   6.3   .78      EBPI   7.27   8.6   .61
. . . . . . . . .            . . .         
ChKl   5.85   7.7   .66      KPel   7.63   9.4   .52
EExp   5.86   7.1   .72      DRKO   7.70   9.3   .52
. .                          . .         
IanL   6.13   7.3   .69      ncs.   7.95   9.7   .44
vegas  6.15   7.4   .72               
. . . .                      DQin   8.05   9.6   .44
avgA   6.58   8.1   .66               
trzu   6.67   8.0   .68      LEBR   8.15   9.8   .46
. .                         
Crow   6.89   8.4   .61      23re   8.26  10.1   .46
emin   6.97   8.4   .57      eWin   8.33  10.2   .47
dtka   7.05   8.5   .61      . .         
vzro   7.07   8.5   .56      NuFi   8.54  10.2   .39
. .                        

Double-space represents separation of .10 or more, and dots represent multiples of .10 separation.

[Crow]

Thanks again Mike G. for scorekeeping and analysis.


2nd place "in the contest", currently held by trzu, remains possible for me and several others.

Somewhat disappointed to be behind Vegas and ESPN experts average and a few others but the universe of projections can have good sets. The contest still has the best known one.

Not as good a season for several top advanced metrics as usual. Will any of their authors use these results as impetus to tweak formula / model?

[Mike G]

Tangential perhaps, here are all entries ranked by avg deviation from last year's Wins:

Code: Select all

delta radical   delta conserv.
5.67   IanL      3.31   vzro
5.52   eWin      3.33   Crow
4.80   emin      3.47   ncs.
4.56   LEBR      3.64   EBPI
4.43   DRKO      3.69   trzu
4.37   medi      3.80   EExp
4.27   ChKl      3.96   KPel
4.24   vegas     3.99   DQin
4.12   NuFi      4.07   dtka

"We" tended more closely to last season than 'they' did.
Exactly zero correlation with avg errors.