2022-23 team win projection contest

[v-zero]

Conservative, at the time, was just what occurred to me. They are about as conservative as is warranted, extreme predictions are in general poor predictions; as you say they were not relatively all that conservative, anyway.

[Mike G]

Everyone looks worse over the last 6 days, by an avg of about +.40 -- from MPra +.10 to KPel +.97

Code: Select all

.   avg err   rmse          avg err   rmse
KPel   5.23   6.40      BIPM   6.35   8.04
22Re   5.38   7.42      538E   6.54   8.76
LEBR   5.47   6.51      TmRk   6.60   8.31
ncs.   5.56   6.97     vegas   6.63   8.45
vzro   5.75   7.10      emin   6.82   8.21
dtka   5.75   6.86      4141   6.86   9.02
avg.   5.82   7.06      538R   6.99   7.99
DRKO   5.82   7.07      AnBa   7.04   8.57
EBPI   5.99   7.08      nuFi   7.15   8.72
trzu   6.04   7.41      EExp   7.46   8.77
Crow   6.24   7.47      MPra   7.55   9.22

Celtics gained almost 3 wins in their projection, by demolishing the Suns, who lost as many.
https://www.basketball-reference.com/fr ... _prob.html

[DarkStar48]

@mike Can you add in R^2 as well?

Yeah @Mike G , adding a r^2 column to the summary table would be useful in showing which projections are best at balancing the finite number of total team wins.

[Mike G]

Is this just the square of the correlations?

[DarkStar48]

Yes it is, aka, coefficient of determination.

[Mike G]

Code: Select all

.    avg err   rmse   r^2          avg err   rmse    r^2
KPel   5.08   6.27   .521      BIPM   6.22   7.88   .602
LEBR   5.31   6.31   .570      TmRk   6.45   8.24   .417
22Re   5.33   7.30   .344      538E   6.47   8.60   .304
ncs.   5.40   6.82   .497     vegas   6.49   8.38   .418
vzro   5.61   6.90   .488      emin   6.67   8.09   .404
dtka   5.61   6.71   .550      4141   6.83   9.00   
avg.   5.68   6.93   .517      538R   6.86   7.84   .455
DRKO   5.79   6.99   .476      AnBa   6.97   8.50   .455
EBPI   5.85   6.97   .502      nuFi   6.98   8.59   .394
trzu   5.90   7.32   .520      EExp   7.31   8.66   .406
Crow   6.11   7.38   .501      MPra   7.49   9.13   .411

[Crow]

There's a negative-0.41 correlation between the age of a team and how they're over-achieving relative to predictions.

Code: Select all

tm     age   over       tm    age   over
LAC   29.7  -10.0      NOP   25.9    5.5
Mil   29.5   -0.5      Tor   25.7   -3.4
Brk   28.7   -5.3      Cle   25.6    8.6
LAL   28.4   -8.2      Min   25.5   -7.3
Phl   28.3   -5.8      Por   25.4    9.4
Dal   28.2   -0.5      Sac   25.4    8.3
Mia   28.1   -9.0      Atl   25.1   -2.6
Chi   28.0    2.5      NYK   25.1   -3.0
Phx   27.7    3.3      Mem   24.6   -3.5
GSW   27.5  -10.1      SAS   24.4   -7.3
Bos   27.4    3.7      Ind   24.2   14.0
Uta   27.1   12.6      Det   24.0   -1.1
Den   26.8   -6.9      Orl   23.1    2.4
Was   26.6    4.5      OKC   22.9   12.6
Cha   26.0   -6.3      Hou   22.4    4.0

Relative to the avg of predictions here (not including dummy entries), the 10 oldest teams (by age on the court) are projecting to under-achieve by 43 wins.
The 10 youngest teams appear headed to win 24 more games, and the middle 10 total 20 more than expected.

The contrast is even more dramatic if the old guys were getting their expected minutes: Lakers without LeBron, for example, register as younger than they'd be with him on the floor.

Comparing actual to BRef Pythagorean based "expected" wins, teams 25 and under combined and 28 and over have exactly the same totals.

The issue appears to be a projection issue and possibly an age curve issue or a failure to regress adequately to the mean.

The original correlation should be rechecked and other research done to advance understanding.

Past seasons and parts of seasons should be checked for split trends.

For last season, both 25 and under and 28 and over are -3 combined for full season. For '20-21 it is -1 and -5 respectively. There are some extreme cases of under and over performance. Thunder over and Pistons under last season and Spurs under this season. In '19-20 it was -1 and +5.
There may not be a trend here long-term but the 3 year trend is slight over-performance by teams between 25-28. Slight average under-performance by 25 and under and 28 and over. That would be believable / in line with my expectations. In a race decided by tenths or hundredths of a point of error, it might help to adjust a bit for this. Most wouldn't care. Checking longer would make sense. Eh, the average trends reverse when adding the 4th and 5th years... but the trends within the period get bigger. More tanking and load management? Maybe. Premature to conclude but notes offered to draw attention.

[DarkStar48]

The r^2 for the different predictions are unexpectedly all over the place. Some high error ones have very substantial correlation (see: BIPM, trzu), while others with smaller errors are not meaningfully correlated (e.g., 22Re, DRKO).

It is still interesting to see which “models” are best at replicating the distribution of actual (or, expected as of now) team wins.

We have to wait till the end of the regular season, but I would be curious to see whether this is an effect of the nature of the BBallRef projections, or if this is a systematic property of the year’s entries.

For comparison, observe below the 2021-22 trends for a selection of different estimates.

[Mike G]

The r^2 for the different predictions are unexpectedly all over the place. Some high error ones have very substantial correlation (see: BIPM, trzu), while others with smaller errors are not meaningfully correlated (e.g., 22Re, DRKO).

"Correlation" seems to be high when the ordering is similar, regardless of the magnitude of error.
BIPM has a lot of highest and lowest guesses. Several are what I'd call outliers:
Bos 62, Brk 58, Mil 63, NOP 55 are at least 2 wins higher than anyone else, while Det 19 and Hou 13 are much lower. Also high on Sac and low on MIa.

If I manually switch his Mil prediction from 63 to 53 -- 53 being exactly where they're headed -- his r^2 drops from .617 to .608. Go figure.
Same if they'd guessed 27 for Hou, rather than 13. When I 'correct' both of these, it's down to .598.
Is something wack here?

Now if I manipulate the 41-41 column by adding merely 1 win to the teams expected to win 51 or more, and take one away from consensus 'bad' teams (31 or less), the r^2 is .518 -- better than Pelton or 3/4 of the field.

[DarkStar48]

Out of curiosity, what is the r^2 for the 4141 “projection” without the manual alteration mentioned above? I’m trying to observe something.

[Mike G]

#DIV/0!

Code: Select all

.   avg err   rmse   r^2          avg err   rmse   r^2
KPel   5.03   6.23   .51      emin   6.32   7.88   .42
LEBR   5.26   6.42   .55      BIPM   6.55   7.87   .61
ncs.   5.37   6.82   .49      TmRk   6.57   8.18   .42
22Re   5.48   7.36   .30     vegas   6.61   8.29   .42
DRKO   5.59   6.73   .50      538E   6.68   8.68   .28
dtka   5.68   6.68   .55      4141   6.83   8.76  #/0!
avg.   5.76   6.89   .51      538R   6.89   8.00   .43
vzro   5.78   7.10   .45      nuFi   6.98   8.58   .39
trzu   5.81   7.17   .53      AnBa   7.30   8.58   .44
Crow   5.82   7.11   .53      MPra   7.31   8.84   .44
EBPI   5.95   7.11   .47      EExp   7.51   8.71   .39

Poor ESPN Experts! Last time they were in this contest, didn't they run away with it?

Well kinda -- Final standings from year before last:

Code: Select all

.   avg err   rmse       .  avg err   rmse
EExp   5.47   7.13      dtka   6.13   8.12
vegas  5.55   7.37      DSM1   6.17   8.38
shad   5.73   7.72      trzu   6.23   8.24
TeRa   5.74   7.68      KPel   6.40   8.44
bbst   5.82   7.70      lisp   6.50   8.69
cali   5.88   7.68      538R   6.60   8.89
Crow   5.93   7.97      emin   6.73   8.67
avgA   5.95   7.90      538E   8.33  10.24
.                       EBPI   9.13  11.52

[DarkStar48]

#DIV/0!

Oh! For some reason I thought “4141” was Pythagorean expectation plus regression.

[Crow]

0.42 average error improvement for me in 4 days.

[Mike G]

Code: Select all

.   avg err   rmse    r^2          avg err   rmse    r^2
KPel   4.96   6.13   .507      emin   6.15   7.74   .431
ncs.   5.24   6.74   .488      TmRk   6.48   8.00   .432
LEBR   5.38   6.54   .526      538E   6.49   8.54   .289
DRKO   5.39   6.61   .503      BIPM   6.54   7.79   .617
22Re   5.42   7.14   .315     vegas   6.55   8.12   .437
dtka   5.55   6.65   .544      538R   6.69   7.86   .439
trzu   5.62   7.10   .537      4141   6.83   8.57   
Crow   5.63   6.96   .540      nuFi   6.83   8.30   .416
avg.   5.64   6.77   .521      AnBa   7.21   8.45   .453
vzro   5.75   7.05   .447      MPra   7.21   8.68   .457
EBPI   5.82   7.11   .466      EExp   7.33   8.60   .402

Including dummy entries, errors have improved by an avg of 0.20 in 5 days.

[Crow]

0.61 improvement in 5 days for me. 0.25 behind 3rd.