2017-18 team win projection contest

[shadow]

If you legitimately had a model that could pick winners at even a 56% clip long term in any major sport, I'm pretty sure you wouldn't sell access to it for $14 per month, considering that would be worth millions of dollars. Right Angle Sports is the only legit pick service I'm aware of and they charge substantially more money for their college football and basketball picks.

[sndesai1]

yep, only a stupid person would sell a model that was a license to print money for $14/month
stupid people generally aren't capable of creating that model in the first place

i'm not going to claim they're lying but from the descriptions posted, it looks like they're using cherry-picked subcategories and dates to create a story...and also ignoring vig


i don't think that r^2 claim from nba.com will tell you much about how predictive that pie stat is. here's a quick example for last season: https://docs.google.com/spreadsheets/d/ ... sp=sharing

[Mike G]

New frontrunner!

Code: Select all

avg abs err   RMSE      avg abs err   RMSE
sbs.   4.12   5.18      GK5.   4.66   6.04
RyRi   4.14   5.21      sndi   4.71   5.62
shad   4.33   5.23      Nath   4.93   5.92
vegas  4.45   5.48      kmed   5.13   6.00
AJBk   4.47   5.77      Mike   5.27   6.23
lnqi   4.48   5.25      emin   5.32   6.06
Josh   4.55   5.81      ATCt   5.40   6.55
gold   4.58   5.57      ncs.   5.42   6.81
cali   4.62   5.60      Crow   5.43   6.46
538_   4.65   5.95      knar   5.71   6.74

[Mike G]

In the last week, we've virtually gotten worse every day; by an avg of .38 in that span.

Code: Select all

worse from Jan 5  to  12
.24   sbs.   3.93   4.16
.26   GK5.   4.48   4.75
.27   ATCt   5.24   5.51
.28   vegas  4.27   4.56
.29   Nath   4.80   5.09
.30   cali   4.31   4.60
.30   lnqi   4.21   4.51
.31   sndi   4.56   4.87
.31   AJBk   4.22   4.53
.32   shad   4.18   4.50
.32   538_   4.43   4.75
.38   Crow   5.22   5.60
.38   knar   5.33   5.71
.43   gold   4.31   4.74
.44   Josh   4.22   4.67
.46   emin   5.01   5.47
.49   RyRi   3.79   4.28
.49   Mike   4.99   5.48
.54   kmed   4.71   5.24
.66   ncs.   4.91   5.57

In the last week, you have gained on everyone below you in this list, and lost ground to those above you.
On Jan. 5, Ryan had a .14 lead over sbs; with twice as much slippage, RyRi now .12 out of the lead.

[Mike G]

I was stuck in a Detroit-area auto shop and talked with the desk guy about the Pistons. After going over historic players and teams, I realized I was stumped as to why the current squad is so much better than last season. Andre Drummond and Tobias Harris were obvious examples, but then I ran out of memory. No great rookies sprang to mind.

Here they are, this year and last year, in some summary stats:

Code: Select all

Pistons   Min   e484   PER   WS/48  BPM     e484   PER  WS/48   BPM
Harris   1310   1.38   18.2  .158   1.5     1.15   17.0  .128   1.2
Drummond 1242   2.08   23.1  .196   6.6     1.92   20.9  .133    .8
Bradley  1037    .66   11.0  .019  -3.1 Bos 1.07   14.5  .080   -.5
Jackson   894   1.44   16.8  .102   -.2     1.23   14.9  .065  -1.5
Johnson   881    .31    8.0  .053   -.2      .08    7.2  .012  -2.0

I Smith   867   1.11   15.1  .090  -1.9     1.07   14.8  .076   -.8
Tolliver  731    .37   10.5  .095    .0 Sac  .37   11.2  .075   -.6
Bullock   588    .47   12.8  .131   1.3      .40   11.7  .101    .1
Galloway  497    .67   13.2  .103   -.9 Sac  .36   10.4  .041  -2.6
Moreland  429    .47   11.2  .120   2.5 Sac  .00    7.4  .012  -4.5

Leuer     136    .40    8.8  .048  -5.6      .84   14.2  .112    .2
Boban     133   1.83   21.1  .142  -1.9     2.22   29.7  .280    .1
Ellenson  119    .85    9.4 -.024  -7.9      .56    7.5 -.071  -8.3

Among their top 10 players, only newcomer Avery Bradley is down in every stat, and Ish Smith's BPM is a bit off.
How does this convert to wins differential -- expected vs achieved --? It's easy with eWins; and there are some conversions/equivalencies to apply to these other numbers.
Per 484 minutes (10 games), teams create 5 wins on average; players thus get 1.0 per 484 on avg.

Code: Select all

avg   2017 to '18 Wins/484 impr.        wins added
xyz   Pistons   e484  WS/48  BPM      eW+   WS+  bpm+
+.73   Harris    .23   .30   .06      .61   .82   .15
1.77   Drummond  .16   .64  1.10      .41  1.63  2.83
-1.4  Bradley   -.41  -.62  -.50     -.87 -1.32 -1.06
+.65   Jackson   .22   .37   .25      .40   .69   .46
+.74   Johnson   .23   .41   .34      .42   .75   .62

+.01   I Smith   .04   .14  -.21      .07   .25  -.38
+.16   Tolliver  .00   .20   .11      .00   .31   .17
+.27   Bullock   .06   .30   .23      .08   .37   .28
+.53   Galloway  .30   .63   .32      .31   .64   .33
+.99   Moreland  .47  1.09  1.33      .42   .9   1.18

-.25   Leuer    -.44  -.65 -1.10     -.12  -.18  -.31
-.23   Boban    -.39 -1.39  -.38     -.11  -.38  -.10
+.09   Ellenson  .29   .47   .08      .07   .12   .02

4.08   totals                        1.69  4.67  4.20

xyz is the average of the last 3 columns, after doubling eW+
Positive numbers indicate extra wins in this year's minutes, over last year's rates. Stratospheric bonus from bench players!
Almost halfway thru the season, Det is projecting to win 9 games more than our avg guess; I'm off by 12.

[jgoldstein34]

Updated mid-way through today (Results of Raptors-76ers and Pistons-Hornets included). All tracking a little bit worse.



4 teams where no one's projection was within <3 compared to their currently projection:
IND (Closest 3.1 away)
MEM (Closest 3.8)
CHO (Closest 5.5)
BOS (Closest 6.8)

[jgoldstein34]

Update: we've all gotten a tiny bit better over the past few days.

[Mike G]

Well I may as well post this anyway:

Code: Select all

avg abs err   RMSE      avg abs err   RMSE
sbs.   4.38   5.19      GK5.   4.88   6.06
RyRi   4.42   5.18      sndi   4.96   5.76
shad   4.60   5.36      Nath   5.15   6.22
Josh   4.62   5.69      kmed   5.31   6.05
538_   4.62   5.87      ATCt   5.34   6.50
lnqi   4.63   5.28      Crow   5.35   6.42
AJBk   4.63   5.52      emin   5.51   6.41
gold   4.68   5.50      Mike   5.59   6.44
cali   4.70   5.50      ncs.   5.59   6.82
vegas  4.71   5.55      knar   5.61   6.67

Not sure how to explain the differences -- which should be shrinking later in the season.

[eminence]

Anybody ever look into how many players they need to include in their predictions to get something stable in terms of team Win projections? Obviously the top 1 or 2 players from a team isn't enough, but 10-15 seems a bit excessive.

[Mike G]

You just want to account for all the minutes. Some teams have a lot of minutes going to replacement level players, or players who aren't really adding wins. Others have very few of these types on the floor.
The Warriors have strong players 1 thru 15. The Bucks have just a handful before they are scraping bottom.
Raps are also very deep. Thunders are top heavy with little underneath.

Code: Select all

avg abs err   RMSE      avg abs err   RMSE
sbs.   4.32   5.19      GK5.   4.86   6.03
RyRi   4.36   5.18      sndi   4.95   5.68
vegas  4.57   5.52      Nath   5.06   6.17
shad   4.58   5.41      kmed   5.16   6.01
Josh   4.59   5.61      ATCt   5.32   6.53
AJBk   4.59   5.63      ncs.   5.39   6.79
cali   4.65   5.52      Mike   5.45   6.26
lnqi   4.67   5.34      emin   5.49   6.37
538_   4.71   5.85      Crow   5.49   6.48
gold   4.80   5.65      knar   5.59   6.71

[Crow]

Anytime including at the end the predictions can be compared to actual wins or expected wins by Pythagorean method. Until the end they can be compared to combination of actual wins to date and expected wins the rest of way. The average "errors" are the variances but one could say they are the sum of variance of prediction to expected and expected team performance and actual performance, right? Did anyone explicitly attempted to model variance of expected team performance and actual performance? I subjectively adjusted for it. In the end the contest will be decided by variance from actual but I'd like to final variances from expected too. How much appears to be each kind of variance? Does anyone plan to model the actual vs. expected variance further, for insight on that and better future predictions? How does player, coach, team performance variation compare to predictor variation? Who is more "responsible" for the error between prediction and actual?

[jgoldstein34]

Here's how my model has it after tonight's games:

[Mike G]

Yikes.

Code: Select all

avg abs err   RMSE      avg abs err   RMSE
sbs.   4.17   5.04      Nath   4.65   5.85
RyRi   4.19   5.03      sndi   4.80   5.42
shad   4.29   5.20      GK5.   4.85   6.02
vegas  4.40   5.33      kmed   4.89   5.80
lnqi   4.41   5.15      emin   5.07   5.96
cali   4.43   5.42      Crow   5.08   6.16
538_   4.47   5.70      ATCt   5.17   6.34
AJBk   4.52   5.53      ncs.   5.29   6.75
gold   4.58   5.50      knar   5.48   6.54
Josh   4.64   5.54      Mike   5.55   6.32

At the 4th power, I rank 6 spots higher.

Feb 6, new front runner

Code: Select all

avg abs err   RMSE      avg abs err   RMSE
RyRi   4.01   4.92      Josh   4.52   5.51
sbs.   4.01   4.94      GK5.   4.61   5.81
538_   4.20   5.47      Nath   4.61   5.74
shad   4.25   5.18      kmed   4.71   5.56
gold   4.26   5.36      emin   4.90   5.66
lnqi   4.27   5.03      ATCt   5.02   6.19
cali   4.30   5.32      Crow   5.06   6.07
AJBk   4.31   5.43      ncs.   5.18   6.68
vegas  4.38   5.29      knar   5.26   6.42
sndi   4.40   5.14      Mike   5.36   6.20

[shadow]

Only have my sheet setup to calculate absolute error but here are the standings through games on 2/15 based on my win projections, which essentially align with B-R's.

Code: Select all

Entry  Error  Rank
RyRi	4.21	1
538_	4.36	2
sbs.	4.37	3
cali	4.42	4
shad	4.46	5
lnqi	4.47	6
sndi	4.53	7
gold	4.57	8
AJBk	4.69	9
kmed	4.72	10
GK5.	4.77	11
Nath	4.84	12
emin	4.87	13
Josh	4.93	14
Crow	5.07	15
ATCt	5.08	16
knar	5.13	17
ncs.	5.31	18
Mike	5.52	19

[Mike G]

co-leaders

Code: Select all

avg abs err   RMSE      avg abs err   RMSE
sbs.   4.42   5.49      AJBk   4.79   5.91
RyRi   4.43   5.46      Nath   4.79   5.92
lnqi   4.50   5.56      emin   4.90   5.77
538_   4.51   5.76      GK5.   4.95   6.40
cali   4.52   5.68      ATCt   4.97   6.44
sndi   4.63   5.47      Crow   5.01   6.18
shad   4.64   5.90      Josh   5.13   6.11
vegas  4.67   5.79      knar   5.19   6.46
kmed   4.72   5.87      Mike   5.37   6.37
gold   4.77   6.01      ncs.   5.44   6.82

The separation between top and bottom has shrunk to 1.02 in abs err. Two weeks ago it was 1.42