2024-25 team win projection contest

[Mike G]

One year ago, mediocre hit their season low 3.86 MAE. A couple of entries were almost twice that.

Code: Select all

.  avg err   rmse   r^2      .  avg err   rmse   r^2
bpmW   6.44   8.0   .38     vegas   7.40   9.6   .34
WShr   6.54   8.2   .35      eWin   7.59   9.4   .19
avgA   6.68   8.5   .36      emin   7.84   9.9   .32
24pr   6.77   8.6   .25      medi   7.91  10.0   .33
TmTj   6.78   8.7   .41      4141   7.95   9.9   
dtka   7.11   8.8   .38      perW   8.09   9.9   .13
DQin   7.30   9.1   .39      24py   8.79  11.2   .25
Crow   7.36   9.4   .38      bmgm   9.71  11.8   .26

https://www.basketball-reference.com/fr ... _prob.html

A clumsy visual of highest and lowest guesses by entries -- in order of least current error -- and current closest and furthest from b-r.com projections:

Code: Select all

entry    highest      lowest       closest        furthest
BPM        ...         ...         Bos Dal          ...
WS     Brk Chi Tor     Mem         Brk Dal          Mem
TmTj     LAL Phx     Den Uta         Den            Phx
dtka   Cle Mil LAL   NOP Por    Cle Orl Was NOP   Mil Por
DQ       Dal LAC     Was Phx     LAC LAL Phx    Det Was Dal
Crow   Orl GSW LAL  Brk Mia Tor    Mia GSW        Brk Tor
"       Mem Sac      SAS Uta                      Sac SAS
vegas     Mil        Det Por   Dal LAL Mem OKC  Det Mil Por
eWin    Det Was    Bos Ind Mil   Det Ind Mil    
"       Hou Por    LAC LAL OKC     Hou Por      LAC LAL OKC
medi      Phl        Cha Chi         NYK        Cha Chi Det
"       NOP OKC      Det Sac         Sac          Phl NOP
PER   Atl Cha Ind  Cle NYK Orl   Cha Chi Min    Atl Cle Ind
"       Tor Uta      GSW Min                      Orl GSW

Ties (and near ties) become duplicates. Omissions stand to be corrected.
Avoiding the extremes, you never get the worst (furthest) guess, but you may get the closest.
Right now, 16 of 30 teams are projecting either above or below all our guesses.

UPDATE Nov.23
Not much change in the standings; just to note that except for the bottom 2, all are showing biggest errors of the season.

Code: Select all

.  avg err   rmse   r^2      .  avg err   rmse   r^2
bpmW   6.47   8.1   .38      vegas  7.45   9.6   .33
WShr   6.60   8.2   .35      eWin   7.79   9.5   .18
avgA   6.83   8.5   .37      emin   7.96   9.9   .32
TmTj   6.94   8.7   .42      4141   8.22   9.9   
24pr   6.95   8.7   .26      medi   8.06  10.1   .32
dtka   7.32   8.8   .38      perW   8.13  10.0   .13
Crow   7.44   9.4   .39      24py   8.81  11.2   .25
DQin   7.45   9.1   .39      bmgm   9.64  11.8   .26

The overachieving teams keep overing and the unders keep undering.

UPDATE Nov.24

Code: Select all

.  avg err   rmse   r^2      .  avg err   rmse   r^2
bpmW   6.22   7.9   .38      Crow   7.52   9.4   .38
WShr   6.35   8.1   .35      vegas  7.55   9.4   .35
avgA   6.53   8.4   .37      perW   7.75   9.8   .14
24pr   6.68   8.6   .25      medi   7.81   9.9   .34
TmTj   7.11   8.7   .41      emin   7.90   9.9   .32
DQin   7.20   8.9   .41      4141   8.13   9.8   
dtka   7.33   8.7   .38      24py   8.80  11.2   .25
eWin   7.33   9.3   .20      bmgm   9.62  11.7   .27

Avg overnight change was improvement of -.17 with wide variation -- from +.17 (TmTj) to -.46 (eWin).

UPDATE Nov.27

Code: Select all

.  avg err   rmse   r^2      .  avg err   rmse   r^2
WShr   6.60   8.4   .35      vegas  7.49   9.7   .34
bpmW   6.62   8.3   .36      medi   7.71  10.1   .33
avgA   6.87   8.7   .36      eWin   7.80   9.7   .18
TmTj   6.97   9.0   .40      emin   7.86  10.0   .32
24pr   7.09   8.9   .25      perW   8.28  10.2   .13
dtka   7.33   9.0   .37      4141   8.42  10.3   
Crow   7.39   9.7   .36      24py   8.81  11.3   .25
DQin   7.41   9.3   .39      bmgm   9.65  11.8   .26

Worst MAE of the season for WS, bpm, avgA, 24pr, DQ, eW, 41 -- and for all together.

UPDATE Dec.01
Best MAE of the season* for Teemo, DQ, Crow, dtka, vegas, medi, emin, 24py, and bmgm

Code: Select all

.  avg err   rmse   r^2      .  avg err   rmse   r^2
bpmW   6.25   8.0   .41     vegas   7.13   9.3   .38
WShr   6.35   8.1   .39      medi   7.27   9.8   .36
avgA   6.41   8.4   .40      eWin   7.42   9.5   .21
TmTj   6.42   8.6   .44      emin   7.55   9.9   .33
24pr   6.72   8.8   .27      perW   8.04   9.9   .16
DQin   6.75   8.8   .44      4141   8.38  10.3   
Crow   6.86   9.5   .39      24py   8.55  11.1   .27
dtka   7.03   8.6   .42      bmgm   9.22  11.5   .29

* - "of the season" = "since mid November"
TmTj now is much closer to the top than to the rest of the pack.
Atl beating Cle -- twice! -- brought everyone's MAE down about .50

Leaders at various exponents:

Code: Select all

e = .05           e = .25          e = .50   
24pr   3.71      Crow   4.46      TmTj   5.18
Crow   3.77      TmTj   4.52      bpmW   5.19
TmTj   3.98      bpmW   4.58      Crow   5.30
bpmW   4.04      24pr   4.66      WShr   5.37
WShr   4.45      WShr   4.86      24pr   5.46
                     
e = 1.0           e=3.0            e = 6.0   
bpmW   6.25      bpmW   9.60      dtka   13.0
WShr   6.35      WShr   9.74      WShr   13.4
TmTj   6.42      dtka   9.95      bpmW   13.5
24pr   6.72      TmTj  10.40      TmTj   14.3
DQin   6.75      24pr  10.57      DQin   14.4

[Crow]

Pretty tight.

How are exponents applied and what is the layman's interpretation of that data?

[TeemoTeejay]

Pretty tight.

How are exponents applied and what is the layman's interpretation of that data?

I for sure got cooked today lol. I wonder how it would have gone if I had gone with MAE or RMSE as a target metric instead lol insane oversight on my part

^ i think it’s like, e3 is RMSE but cubed instead of squared, etc, 4 would be to the power and root of 4, etc

Practically higher exponents will penalize larger errors more

[Mike G]

How are exponents applied and what is the layman's interpretation of that data?

^ i think it’s like, e3 is RMSE but cubed instead of squared, etc, 4 would be to the power and root of 4, etc
Practically higher exponents will penalize larger errors more

Exactly that. And tiny exponents reward exact hits, dismissing the big misses.

Depending on the layman -- for most it's difficult to express what RMSE is or why it is a thing. But it's popular and mainstream among statisticians, I think because it's historically been possible to calculate.
Right now, the 4141 prediction becomes the 'leader' at e > 20. Their worst guess is off by 22 (Was), and everyone else has at least one which is worse.

Anyway, we now see a 3-way race for the lead:

Code: Select all

.  avg err   rmse   r^2      .  avg err   rmse   r^2
bpmW   6.48   8.3   .38     vegas   7.35   9.6   .36
WShr   6.58   8.5   .36      medi   7.60  10.1   .34
TmTj   6.59   8.8   .42      emin   7.70  10.2   .32
avgA   6.72   8.7   .37      eWin   7.71   9.8   .20
24pr   6.94   9.1   .25      perW   8.19  10.2   .14
Crow   7.01   9.7   .37      4141   8.55  10.5   
DQin   7.09   9.1   .41      24py   8.73  11.4   .25
dtka   7.25   8.8   .40      bmgm   9.40  11.8   .27

UPDATE Dec. 4

Code: Select all

.  avg err   rmse   r^2      .  avg err   rmse   r^2
bpmW   6.29   8.2   .43      vegas  7.05   9.3   .40
TmTj   6.33   8.6   .46      medi   7.25   9.9   .37
WShr   6.34   8.3   .42      emin   7.49   9.9   .35
avgA   6.54   8.5   .42      eWin   7.55   9.7   .23
Crow   6.70   9.5   .41      perW   7.92  10.2   .18
DQin   6.72   8.8   .46      24py   8.24  11.1   .30
24pr   6.77   9.1   .30      bmgm   8.92  11.4   .32
dtka   7.01   8.5   .45      4141   8.96  10.8 

Lowest errors so far for all but my gang of 4. And 4141 securing the basement.

[Crow]

Now I'm up to 5th. 1st place is only half as far away as it was 2 weeks ago.

[Crow]

Straight record projection would essentially yield 8 50+ teams. BRef projects 6 and 2 very close. I projected 8.

Vegas 6, APBRmetrics average 3.

[Mike G]

A good night for those who don't like the Lakers, Clipps, Bucks' chances.

Code: Select all

.  avg err   rmse   r^2      .  avg err   rmse   r^2
WShr   6.11   8.2   .43     vegas   7.05   9.3   .40
bpmW   6.12   8.1   .45      medi   7.07   9.8   .38
TmTj   6.34   8.6   .45      eWin   7.18   9.5   .26
avgA   6.45   8.4   .43      emin   7.36   9.8   .37
24pr   6.53   9.1   .30      perW   7.82  10.1   .19
DQin   6.69   8.7   .46      24py   8.23  11.0   .30
dtka   6.82   8.5   .45      bmgm   8.83  11.4   .32
Crow   6.91   9.5   .40      4141   8.92  10.8   

Season low errors for WS, DQ, kavana, vegas, medi, emin, bmgm
There was a major upheaval or recalibration since yesterday at b-r.com. And it seems the "NBA Cup" is resolving, such that teams are now totaling closer to 1230 wins.

[Crow]

Further regression to 41 is probably the wrong move but whatever. It will eventually revert to the actual.


To be clear, are you using BRef's projections exactly or adjusting?

[Mike G]

Sorry, I was unclear and now edited. Until a few days ago, b-r.com was giving W-L projections for 80 game seasons, and I was adding a win and a loss to all. Now it's just +.11 win adjustment.
I have not "de-regressed" the b-r.com numbers, as this makes all errors worse.

DQuinn the big overnight riser, and a virtual 3-way tie at the top.

Code: Select all

.  avg err   rmse   r^2      .  avg err   rmse   r^2
WShr   6.33   8.3   .44      medi   7.12   9.8   .39
bpmW   6.34   8.1   .46      vegas  7.17   9.3   .41
TmTj   6.36   8.6   .47      eWin   7.27   9.6   .26
DQin   6.59   8.6   .48      emin   7.62   9.7   .38
avgA   6.61   8.4   .44      perW   8.04  10.2   .19
24pr   6.68   9.1   .32      24py   8.32  10.9   .32
Crow   6.90   9.4   .42      bmgm   8.85  11.2   .34
dtka   7.00   8.5   .46      4141   9.06  11.1   

WS leads at exponent from .95 to 1.02

UPDATE Dec. 7
Wins by Phl, Atl, Ind, Sac, Min, Sac; losses by Orl, Chi, SAS, GSW, Por -- that's 11 of 14 teams that played last night -- moved errors down by an avg of .50. The range of improvement was from .25 (Crow) to .60 (bpm, emin). Everyone with best marks of the season

Code: Select all

.  avg err   rmse   r^2      .  avg err   rmse   r^2
bpmW   5.74   7.8   .49      vegas  6.64   8.9   .45
WShr   5.77   7.9   .48      Crow   6.66   9.1   .45
TmTj   5.84   8.2   .50      eWin   6.92   9.3   .30
DQin   6.04   8.2   .52      emin   7.02   9.4   .41
avgA   6.08   8.0   .48      perW   7.51   9.8   .23
24pr   6.38   8.9   .35      24py   7.78  10.6   .35
dtka   6.44   8.0   .50      bmgm   8.38  10.8   .37
medi   6.59   9.4   .43      4141   8.98  10.9   

[TeemoTeejay]

We can just guess Wins; losses naturally follow.
I have 4 submissions:

Code: Select all

west  eWin   perW    WS    bpmW      east  eWin   perW    WS    bpmW
Dal   51.3   47.0   48.4   48.0      Atl   39.6   43.5   43.6   41.1
Den   55.6   48.2   51.4   50.3      Bos   55.1   48.2   60.3   55.5
GSW   41.7   41.0   46.0   46.3      Brk   32.0   36.2   33.0   31.9
Hou   47.6   42.5   44.7   45.3      Cha   22.9   32.1   22.0   24.3
LAC   29.1   36.3   42.9   39.5      Chi   30.8   35.1   30.8   33.5
LAL   38.5   40.1   38.1   39.3      Cle   42.5   40.6   45.2   44.0
Mem   39.9   38.7   30.9   38.9      Det   34.5   37.7   30.7   33.4
Min   51.9   43.2   51.8   50.1      Ind   43.4   46.1   44.0   43.6
NOP   45.5   43.9   45.5   45.9      Mia   47.5   42.4   45.4   44.4
OKC   51.6   48.1   58.4   54.8      Mil   41.2   41.7   42.1   39.6
Phx   45.6   45.0   48.3   47.6      NYK   40.8   39.7   44.0   42.7
Por   30.1   33.2   23.1   27.7      Orl   40.6   38.1   46.0   42.3
Sac   48.9   46.0   47.9   49.1      Phl   49.3   44.6   45.0   43.0
SAS   33.9   37.7   29.8   34.6      Tor   28.5   36.0   30.8   32.0
Uta   35.2   39.3   32.3   31.4      Was   35.0   37.7   27.6   30.0

Assumed same player-minutes across each of these stats.
Nobody gets more than 2500 min. Players who missed a lot of games are assumed to play more this year.

Just wondering, when you did these projections was it just the stat * by players minutes with wins as a target variable, or was it a main component in a larger prediction

[Mike G]

For PER and BPM, I had formulae converting to 'wins'; multiply by minutes and minimizing differences w actual league wins.
perW = (PER - 5.75)*min/4359
bpmW = (BPM + 5.25)*min/2621
WS = (WS/48)*min/49.08

Zero 'wins' are created by PER<5.75 or BPM< -5.25

These all yield 1230 wins from last year's players.
The PER win distribution was especially flat, and WS was very broad. It seems most of the better projections here are regressed toward 41-41, so I regressed WS and de-regressed eWins and PER to match the avg deviation of BPM. It turns out that these were all less than the AveDev of everyone else here; with mixed results.

Note: player eWins is by design pre-regressed when totaled for a team. A team whose players total 51 eW should win 61 games. So my eWin guesses are roughly in the middle.

Note2: I don't know how b-r.com does Win Shares. During playoff series, player WS (as WS = ws/48*min/48) never add up to 1.00 per game, and this is true in regular season as well. Overtime minutes don't explain it.

I used minutes from Kevin Pelton of espn.
Is anyone going to find the espn Experts' prediction? or Pelton's? I can toss out the "41 for everyone" dummy.

[Crow]

ESPN panel

https://www.espn.com/nba/story/_/id/418 ... -odds-more

Pelton rounded

https://www.mavsmoneyball.com/2024/9/27 ... rojections

See pm for more.

[Crow]

Heavily regressed to 41Ws may be an effective strategy on average but it doesn't impress me much as insight.

ESPN Expert projections have 2 over 50, 1 less than 30. Currently BRef is projecting 6 and 4 respectively.

Of my 8 projections of 50+, 6 are projected that way or 1 win off by BRef. I picked 3 of the 4 projected by BRef to end up below 30 as such. A few more I picked for under 30 are currently projected in 30s but could fade for injuries and draft maneuvering.

If I lose because I wasn't cautious / indifferent / regressed enough in the middle, so be it, for me.

[TeemoTeejay]

Even though I think bbref prediction on the lakers win total is pretty harsh, very happy that I didn’t arbitrarily give the lakers 10 extra wins haha

45-48 still seems about right to me

^ it’d be interesting to see different predictions over Ann std dev in wins.

[Mike G]

2 new entries thanks to data leaks; and they are frontrunners.
Crow and emin again with lowest err of the season.

Code: Select all

.  avg err   rmse   r^2      .  avg err   rmse   r^2
KPel   5.84   7.4   .61      vegas  6.65   8.8   .46
WShr   5.87   8.0   .47      24pr   6.65   9.0   .33
bpmW   5.87   7.8   .50      medi   6.85   9.4   .42
TmTj   5.90   8.2   .51      emin   6.94   9.4   .42
eExp   6.14   8.1   .55      eWin   6.98   9.2   .32
avgA   6.19   8.0   .48      perW   7.64   9.8   .23
DQin   6.30   8.3   .52      24py   7.91  10.8   .33
dtka   6.60   8.1   .50      bmgm   8.45  11.1   .35
Crow   6.60   9.0   .46   

Pelton's predictions sum to 1230.9
He's correctly outlier-high on Brk, Mem; also highest on Tor, Por; low to very low on Atl, Bos, Min, NOP. A few best guesses in there and only one worst (Bos, tie).