2023-24 team win projection contest

[Mike G]

Not much has changed, other than the leader's lead has grown again. Top to bottom diff is season high also.

Code: Select all

.  avg err   rmse   r^2      .  avg err   rmse   r^2
medi   5.00   6.0   .79      dtka   7.08   8.3   .60
EExp   6.15   7.3   .69      EBPI   7.24   8.5   .59
ChKl   6.18   7.8   .63      DQin   7.45   9.2   .46
vegas  6.26   7.5   .68      KPel   7.49   9.2   .52
vzro   6.45   8.1   .59      DRKO   7.50   8.9   .54
avgA   6.53   7.9   .66      ncs.   7.74   9.5   .44
IanL   6.54   7.7   .63      LEBR   7.95   9.4   .48
trzu   6.76   7.8   .68      eWin   7.95  10.1   .45
Crow   6.88   8.3   .60      23re   8.10   9.7   .50
emin   6.93   8.4   .55      NuFi   8.80  10.1   .38

The avg avg error peaked on Jan.14 and is now .36 less.
https://www.basketball-reference.com/fr ... _prob.html

[Crow]

Still 4th in contest. Still decent but not to aspirations or past peaks.


Any concerns or change motivations by authors of Darko and LeBron based of weak performance this year?

Any dialogue from them? Or any other participant or listing? That was and is sought with the exercise.


I see again that my subjectively based projection performance could probably be better taking in more awareness of and influence from the crowd and not just past top performers.

[v-zero]

If I had given subjectivity any sway I would have pulled down my estimate for the Knicks, and bumped the Thunder, but otherwise I don't think I would personally have tweaked anything much.

As a Spurs fan I am depressed by their performance, or rather I can see that Wembanyama is already one of the thirty or so best players on the team, but he is surrounded by a core that has not developed, and I am unsure what on earth the front office is doing at this point.

I'm satisfied enough with where I sit, though I am once again defying the god of errors by having a low MAE for my RMSE.

[Mike G]

... having a low MAE for my RMSE.

And if SMRE were a thing -- Square of the Mean of the square Roots of Errors -- you'd be in 2nd place there; just better than the EExp gang.
At exponent .10, you're only .35 off the lead!

APBR entries now avg .20 better than the outsiders.
Mediocre's avg err is lowest since Dec.29

[Mike G]

NOTE: Fantasy/ thought experiment!

Over on the b-r.com playoff projection page, they show along with the projected wins, a maximum and minimum win total for each team. Assume these are the highest and lowest from their 10,000 simulations; thus about .0001 chance of either. All teams have some non-zero chance of ending up 10 to 11 wins more or less than their median.

So I wish the Celtics would win 10-11 fewer than their projected 62.5, and that the Pistons would win that many more than 15 now expected. But these are unrealistic hopes; so I've entered numbers in my spreadsheet that are up to half the max variation, to force them closer to my guesses. I just manually added/subtracted up to 5.5 wins with each team. Perhaps worse than 1 in a million chance that all 30 teams would accommodate me in this way.

The majority of teams are doing either better or worse than any of us guessed, so we all do better here, in absolute terms:

Code: Select all

.  avg err   rmse   r^2      .  avg err   rmse   r^2
avgA   3.397  4.4   .76      DRKO   4.17   5.7   .61
vegas  3.400  4.0   .80      eWin   4.20   6.2   .60
medi   3.52   4.2   .81      IanL   4.33   5.3   .70
trzu   3.53   4.3   .78      KPel   4.40   5.6   .62
EExp   3.60   4.2   .79      DQin   4.52   6.0   .58
EBPI   3.92   4.9   .71      LEBR   4.62   5.9   .57
dtka   4.00   4.8   .72      ncs.   4.66   6.3   .52
ChKl   4.00   5.1   .70      emin   4.67   6.2   .62
vzro   4.02   5.4   .66      23re   4.77   6.0   .60
Crow   4.07   5.1   .68      NuFi   5.54   6.8   .54

Best-best-case scenario for eWins still doesn't get out of the 2nd division; and Mediocre still wins here (barely).

[Crow]

"The majority of teams are doing either better or worse than any of us guessed..."

So we are left to think that performance is often significantly influenced by variables not in the projection approaches (Coaching, analytics, specifically lineup & sublineup management, variable aging, "real" game matchup dynamics not captured in average stats... what else?) or randomness in play, randomness or subliminal or active bias in refereeing, or playoff effort / settle for picks behavior?

Simple regression to mean may have been too dominant a theory / practice, especially in some projections

[v-zero]

Teams sitting within plus or minus six wins of their predictions would fit 95% confidence at this point based purely on randomness alone.

[Mike G]

Teams sitting within plus or minus six wins of their predictions would fit 95% confidence at this point based purely on randomness alone.

So that means 2-3% likely to be more than 6 wins better, and 2-3% would be worse?
Perhaps one team goes beyond in each direction?

[v-zero]

So that means 2-3% likely to be more than 6 wins better, and 2-3% would be worse?
Perhaps one team goes beyond in each direction?

So yeah, we'd expect around two teams to be outside that interval, under the assumption that our preseason predictions were exactly correct in teams of the distribution mean for each team. They could both fall in one tail, or not.

So if you inspect the average prediction of the contributors, prorate it to the number of games played so far, and run a comparison I would be interested to see those numbers.

[Mike G]

...
So if you inspect the average prediction of the contributors, prorate it to the number of games played so far, and run a comparison I would be interested to see those numbers.

The avgA entry is the avg of our predictions. It's doing better than any one of us, other than medi.
I don't recall who is this new #2. "Chuck Klein" pops into my head, but he was a 1940's baseball player.

Code: Select all

.  avg err   rmse   r^2      .  avg err   rmse   r^2
medi   4.64   6.0   .79      dtka   6.98   8.3   .59
ChKl   5.89   7.6   .64      EBPI   7.20   8.4   .60
EExp   5.96   7.2   .69      KPel   7.34   9.2   .50
vegas  6.19   7.4   .68      DRKO   7.49   9.0   .52
IanL   6.20   7.3   .66      DQin   7.67   9.4   .44
avgA   6.45   7.9   .65      ncs.   7.76   9.5   .43
vzro   6.50   8.2   .57      LEBR   7.93   9.5   .46
emin   6.62   8.4   .54      eWin   8.01  10.0   .45
trzu   6.65   7.9   .66      23re   8.03   9.7   .49
Crow   6.79   8.2   .60      NuFi   8.81  10.1   .37

Mediocre's lead over #2 peaked a couple days ago and could yet be bettered.
Top to bottom spread beats the record almost daily.

Here are the over-unders relative to our avgA :

Code: Select all

over   tm      over    tm
16.9  OKC     -14.8   Cha
12.2  LAC     -13.8   Mem
9.8   Min     -12.1   Was
9.7   Ind     -11.2   Det
9.6   Hou     -10.2   SAS
8.9   Orl      -8.3   Tor
5.5   Bos      -5.5   Atl
4.8   NYK      -5.5   Por
4.0   Sac      -5.3   GSW
3.8   Den      -3.3   Brk
3.2   Cle      -2.7   Phx
3.1   Phl      -2.0   LAL
2.7   NOP      -1.1   Chi
1.3   Mil      -1.1   Mia
1.0   Dal         
0.3   Uta

[mediocre]

Mike, what would the standings look like if some luck was removed and instead of projected wins as the basis for scoring, it was something like projected wins + (predicted wins - actual wins)?

[Mike G]

We predicted wins/82 games, so I'd have to pro-rate that number to team games played.
So for example, the Hawks: They are 22-28 thru 50 games, and projected to go 14-18 the rest of the way, for 36.1 wins.
So your prediction of 39.4 wins would pro-rate to 24.0 now; and your formula would then be 36.1 + (24-22) or 38.1 W, leaving an error of 1.3 vs your current 3.3 ?

Your guess was lowest for Atl; DQ was high at 45.6 -- times 50/82 = 27.8 at this time.
Their adjusted guess then is 36.1+(27.8-22) = 41.9, for an adjusted error of 3.7 rather than 9.5

So DQ Atl error is currently 6.2 worse than yours, and with your proposal (as I read it), it would be only 2.4 worse.
The ratio of the 'errors' is unchanged, as both are reduced to ~.39 of their current value.

I'm going out to pull weeds and think about how this is reducing the luck element.

[mediocre]

Sorry, horribly worded on my part. By "predicted wins" I was referring to bball-ref predicted wins.

In hindsight the formula for replacing the current model (projected wins) is even simpler: remaining wins + predicted wins (with predicted wins as bball-ref predicted wins).

This would mitigate some of the luck that has already occurred to date, which has impacted actual wins, but not remaining wins.

tl;dr: instead of our predictions vs. bball ref projected wins, what would the results look like if it were our predictions vs. (bball-ref predicted wins (to date) + bball-ref remaining wins)?

[v-zero]

I think the simplest luck adjustment would be to take 41 + SRS*82/30 and compare that to the predictions, or take Pythagorean wins and compare them to the predictions.

Note that the linear model will exaggerate the extremes more than Pythagorean.

[Mike G]

Do these transmogrifications lead to a season finale of actual wins to predicted wins?