2025-26 team win projection contest

[nbacouchside]

Not sure if this would be helpful as a thing to track against, but figured I'd offer it up.

Basically built end of season projections that use Vegas adjusted win totals as the prior and then use opponent, HCA adjusted performance, with FT and 3PA luck adjustments (regressing outliers towards league average). The weights change after every game played for a team, so basically a time-series analysis.

I have been posting the updates after every night over on X and Bluesky.

[Crow]

7 over 50 wins, 5 under 30 with this.

BRef up to 4 and 4.

Your distribution is more realistic based on history.

I'd like to see it used. Might manually apply to my projections.

The two will move closer over the weeks and ultimately are replaced by the actual. But seeing each story would be more interesting than just BRef, especially early. A blend is also possible.


Using this, playoff seeds can be anticipated. That should draw attention. And possible reaction / manipulation.

Thunder headed, at this point, for 1st round matchup with Blazers or Clippers. BRef is projecting Thunder vs. Blazers as well, but as a 3 - 5 matchup. BRef method "likes" Blazers far more than Clippers (projected 11 seed).

[Crow]

Using Kevin's target data, my projections are an average of 4.76 different at this time.

[Mike G]

This seems to assume that the 5-1 Bulls will henceforth turn into their predicted 33-win selves, and the 2-6 Mavs will in fact play like a 42-win team the rest of the way; with minor "luck" adjustments.
The season is already about halfway to that 15-game mark where preseason expectations have overall zero bearing on the 82-game outcome. Around that time, only to-date performance is correlated to final win%.

[Crow]

BRef method not revealed to my knowledge. Highly regressed current net margin is assumed to be a prominent feature in simulation but details not known.

[Crow]

Darko by Kevin's target data is 4.37 different.

By Kevin's target data I am more than 60% closer to Darko, with whatever else happens to others.

[Mike G]

Side by side with errors from b-r.com projections.
Unsurprisingly, all errors are smaller; and vegas with the early large lead, with "error" less than half of previously believed.

Code: Select all

.   ncs.          b-r.com
vegas  3.44      DRKO   6.25
EExp   4.11      EExp   6.61
medi   4.34      Walr   6.66
DRKO   4.36      Mgoo   6.67
avgA   4.42      eWin   6.67
bpmW   4.46      knar   6.68
Walr   4.50      avgA   6.71
Mgoo   4.56      bpmW   6.73
Crow   4.76      perW   6.73
dtka   4.89      dtka   6.86
KPel   4.90     vegas   7.04
perW   5.12      WShr   7.13
eWin   5.18      Crow   7.28
knar   5.19      KPel   7.42
WShr   5.36      medi   7.99
DQin   6.74      DQin   9.16

[Crow]

Thanks for the comparison.

Vegas' lead is not really surprising since they serve as the base of this method.

Generally entries shift by a few spots but there are some big differences.

In a 50/50 blend, a few average better than 5.5: Vegas, Darko and EEXP.

[nbacouchside]

This seems to assume that the 5-2 Bulls will henceforth turn into their predicted 33-win selves, and the 2-6 Mavs will in fact play like a 42-win team the rest of the way; with minor "luck" adjustments.
The season is already about halfway to that 15-game mark where preseason expectations have overall zero bearing on the 82-game outcome. Around that time, only to-date performance is correlated to final win%.

For the record, the analysis I did suggests that at 13 games, Vegas expectations are still worth about 50% of end of season win projection. It drops below 25% after 31 games. It doesn't drop below 5% until after game 51.

[Mike G]

Interesting. Was this for last season, or for more than one season?
I've done it several times but maybe not last year. At ~13 games, more teams end up closer to their 13-game W% than to pre-season guesses. And by December, those guesses tend to be useless.

[Mike G]

Here's a post from Nov. 17 last year:
https://www.apbr.org/metrics/viewtopic. ... t&start=71

As of Nov. 16, teams had played 11 to 14 games. The Cavs were 14-0, and b-r.com was projecting them to win 60 games.
Our APBR-submitted average, known as avgA, was 48 for Cle; they ended up winning 64; so the b-r.com projection would end up being better by 12.
https://www.basketball-reference.com/fr ... &lg_id=NBA

A retro-contest between pre-season avgA and Nov.16 b-r.com produces a clear winner. We can multiply one contestant's guesses by X and the other's by (1-X), and sum them, to get a lowest error at season's end. It turned out like this:

Code: Select all

avgA  b-r   MAEr
1.0   0.0   7.82
0.8   0.2   7.34
0.6   0.4   7.13
0.4   0.6   6.92
0.2   0.8   6.71
0.0   1.0   6.52
-.2   1.2   6.47

The lowest error is achieved at 1.22 to -.22
At less than 15 games, there was no overall value in the pre-season prediction. That 7.82 was in fact avgA final MAE, which finished 3rd among 17 entries. Vegas surged from 12 to #8 at 8.40 on last-day tanking.

Here are the teams arranged by avgA- minus b-r.com errors at season's end:

Code: Select all

A-b   tm   avgA  b-r   W       A-b   tm    W   avgA  br
20.3  Phl   49   29   24     -12.7   Brk   26   27   40
17.1  NOP   45   28   21      -8.5   Mil   48   46   37
12.0  Cle   48   60   64      -7.4   Ind   50   44   37
9.9   Hou   42   52   52      -7.0   GSW   48   46   57
8.4   Det   30   38   44      -6.6   Min   49   51   41
5.6   Phx   47   41   36      -5.6   SAS   34   34   40
5.4   Por   26   31   36      -5.5   Cha   19   30   35
5.4   Sac   47   42   40      -4.8   Den   50   49   44
5.3   Dal   50   45   39      -4.3   Orl   41   43   47
5.1   Chi   31   36   39      -4.2   Was   18   23   27
4.8   LAC   37   42   50      -4.1   Atl   40   39   35
4.6   Mem   43   48   48      -3.2   Bos   61   58   55
4.1   OKC   56   60   68      -1.7   Tor   30   30   28
2.9   Uta   30   27   17      -1.2   NYK   51   47   46
2.6   Mia   43   40   37                  
2.5   LAL   41   43   50                  
116.0   total               -76.8   total         

On the left, b-r.com at 13 games would end up closer; on the right, pre-season apbr.
b-r.com does better by 1.3 per team

[Mike G]

Improvements from 0 to 60, avg .30

Code: Select all

.  avg err    rmse   r^2      .  avg err    rmse   r^2
DRKO   6.14   7.67   .40      knar   6.52   7.81   .36
EExp   6.18   7.42   .44      dtka   6.53   7.95   .38
Mgoo   6.23   7.81   .41      25pr   6.84   8.37   .33
Walr   6.29   7.49   .43      Crow   6.85   8.74   .36
perW   6.31   8.46   .32      WShr   7.11   8.27   .35
bpmW   6.39   7.50   .44      KPel   7.19   8.39   .40
avgA   6.39   7.72   .41      25re   7.30   9.03   .31
eWin   6.44   8.25   .34      medi   7.39   8.61   .42
vegas  6.44   7.79   .50      DQin   9.01  10.52   .24

[Crow]

2 different accountings of the value of pre-season Vegas lines and APBR average projections. Further discussion or possible reconciliation would be worthwhile.

Separately there is the question of the relative value of BRef projections now or after say 15 games vs. Kevin's approach. Let's capture the difference after 15 games and see how that comparison ends up.

Then we get to the possibility of others trying to predict end results by the 15th game. I might make such a projection set. Anybody else interested? If so, put it here by Nov. 18. Pre-season has a lot of unknowns. Still a lot now but there is more information to process and try to get a closer understanding and projection. It is a second contest. Which is more interesting or valuable is up for discussion. With more information to process and likelihood of greater accuracy, Nov. 18 projections appear more lasting interesting (and unique) to me than Oct. 21 projections.

[DQuinn1575]

This seems to assume that the 5-2 Bulls will henceforth turn into their predicted 33-win selves, and the 2-6 Mavs will in fact play like a 42-win team the rest of the way; with minor "luck" adjustments.
The season is already about halfway to that 15-game mark where preseason expectations have overall zero bearing on the 82-game outcome. Around that time, only to-date performance is correlated to final win%.

For the record, the analysis I did suggests that at 13 games, Vegas expectations are still worth about 50% of end of season win projection. It drops below 25% after 31 games. It doesn't drop below 5% until after game 51.

I must be missing something, if I do a simple Excel correlation, I get about 61-62% for FanDuel and SportsOdds for the whole season last year, while our entries were in the 64-66% range.

[nbacouchside]

Interesting. Was this for last season, or for more than one season?
I've done it several times but maybe not last year. At ~13 games, more teams end up closer to their 13-game W% than to pre-season guesses. And by December, those guesses tend to be useless.

I used every 82 game season (so removed all the weird shortened seasons for COVID and lockout) from 2008 on.