Power Ranking

[EvanZ]

The CHI-WAS line only moved two points when it was known that Rose wouldn't play. My Vegas ratings had that line at -14. The opening line was -13.5, so pretty close. The closing line was -11.5. (Funny enough, the actual result turned out to be -14 in an extremely low scoring game.) Assuming Vegas thinks Rose plays 36 minutes, can we infer that his "Vegas +/-" is about 2.7, give or take?

[J.E.]

It's probably a little more complicated than that. You'd have to consider that Watson also wasn't playing, John Lucas was the starting guard and the Bulls just signed Mike James who also played in that game. I think, Vegas is saying Rose+Watson are 2.7 better than Lucas+James, but you don't know the individual numbers. To complicate matters, Hamilton also didn't play

[Mike G]

I think, Vegas is saying Rose+Watson are 2.7 better than Lucas+James, but you don't know the individual numbers. To complicate matters, Hamilton also didn't play

I can't believe Rose and Watson and Rip are only 2.7 points better than some replacement/D-League players.
Maybe Vegas figures the Wizards will not only fail to rise to the occasion, but they'll actually sink to the occasion?

[EvanZ]

Another possibility is they thought Rose would only play 26-30 minutes in a blowout anyway.

[mystic]

Well, I did a couple of things after seeing that Vegas has a clear smaller RMSE than my ranking can provide. My guess was that Vegas is just closer to the mean for all teams, which explains the RMSE. Thus, I just took of 40% of the difference to the mean for each team and calculated a new RMSE, the result: 12.26. So, in fact, if the ratings for the teams are all closer to the mean the RMSE will decrease.

I now checked the ability to predict the correct winner. My prediction based on the Power Ranking + HCA gives me 61 correct winners out of 78 games, Vegas has 59. That made me think that Vegas might be just better in terms of RMSE, because the closer to the mean approach gives them less error in case they have the incorrect winner. And indeed, if I only use the 61 games in which my Ranking predicted the correct winner, the RMSE goes down to 9.11, while Vegas in their 59 have 9.79. My "new" closer to the mean approach has 9.70 in also 59 games with the correct predicted winner.
When we now look at the wrong winners, my Ranking has a RMSE of 21.97 in those 17 games. Vegas has 14.69 in their respective 19 games.

Code: Select all

Sat, Jan 7, 2012   Chicago Bulls          94  Atlanta Hawks            109
Mon, Jan 9, 2012   New Orleans Hornets    94  Denver Nuggets           81
Sat, Jan 7, 2012   New York Knicks        103 Detroit Pistons          80
Tue, Jan 10, 2012  Miami Heat             106 Golden State Warriors    111
Wed, Jan 11, 2012  Atlanta Hawks          84  Indiana Pacers           96
Tue, Jan 10, 2012  San Antonio Spurs      103 Milwaukee Bucks          106
Fri, Jan 6, 2012   Cleveland Cavaliers    98  Minnesota Timberwolves   87
Wed, Jan 4, 2012   Memphis Grizzlies      90  Minnesota Timberwolves   86
Wed, Jan 11, 2012  Philadelphia 76ers     79  New York Knicks          85
Wed, Jan 4, 2012   Charlotte Bobcats      118 New York Knicks          110
Tue, Jan 3, 2012   Portland Trail Blazers 103 Oklahoma City Thunder    93
Fri, Jan 6, 2012   Portland Trail Blazers 77  Phoenix Suns             102
Wed, Jan 11, 2012  Orlando Magic          107 Portland Trail Blazers   104
Sat, Jan 7, 2012   Denver Nuggets         117 San Antonio Spurs        121
Fri, Jan 6, 2012   New Jersey Nets        97  Toronto Raptors          85
Wed, Jan 11, 2012  Sacramento Kings       98  Toronto Raptors          91
Tue, Jan 10, 2012  Toronto Raptors        78  Washington Wizards       93

Those are the 17 games in which my Ranking predicted the wrong winner. Can anyone see a common theme here? Any idea?

[EvanZ]

All the games took place in January.

[Jeff Fogle]

Some of those on the list triggered memories of some three-point extremes. So, I went back to check that.

17 losers: 100 of 323 on treys for 31%
17 winners: 135 of 314 on treys for 43%

A reminder at least that treys can skew results. Even good teams have some off nights. Even bad teams have some good nights. Sometimes Orlando goes 16 of 27 as a dog. Sometimes favorites go 2 of 11, or 4 of 20, or 2 of 19.

Might be part of what differentiates stat stuff from the market in short samples. Maybe the market largely dismisses extremes in either direction, but stat measures are influenced (partially polluted?) by them temporarily. Not sure if that's what you're looking for or not mystic, but I figured I'd type it up since I noticed the trey differentials...

[mystic]

All the games took place in January.

I would be worried, if one of the games would have happened in another month.


Jeff Fogle, I'm interested in any ideas and yours seems to be a pretty good one. So, it might be related to teams which are more dependent on their 3pt shooting or have a bigger variance from game to game in terms of converting their attempts.

[Crow]

Anybody want to put together a blend of the various power ratings (those here, Hollinger's, maybe others) and compare the performance of an equal blend to the unique ones? Or use recent metric performance to constantly refine the blend over the season? I think that would be interesting.

[mystic]

Cleveland won at Phoenix. The Cavs on the road made so far only 28% of their 3pt attempts, in this game they went 10 of 19 from behind the arc. They exceeded expectations by around 14 points with their 3pt shooting. But the Suns also shot better than in previous games, making 4 points more than expected from 3pt range. That still counts for a 10 pt swing. Vegas had the Suns winning by 7, I had them winning by 8.4. They lost by 11. The 3pt shooting alone can't explain the difference here.

I suspect some sort of inconsistency in the performance level for the participating teams. It is not something really related to b2b games, because neither the Suns nor the Cavs had a game the day before. Also, in other cases b2b games weren't a factor.

In my personal quest vs. Vegas: 12.99 vs. 11.33, I closed the gap a tiny little bit. I'm still ahead by 2 in terms of picking the right winner. Vegas and myself were only wrong about the winner of the Cavs@Suns last night. That makes 65 of 83 correct for me and 63 of 83 for Vegas.

[DSMok1]

Another site's take on the Vegas Power Rating: http://bettingmarketanalytics.blogspot. ... -2012.html

(Originally created for the NFL at AdvancedNFLStats and then adapted for the NBA).

[EvanZ]

I think this is really interesting. If I take my pre-season RAPM+ezPM blend (what Mike is calling ez2 I believe), the correlation (R^2) between that and my current "Vegas Ratings" is 0.86. Not only that, but the actual slope of the regression is very, very close to 1 (0.95). This seems almost too good to be true, but there it is.



Also, it makes me think, going forward, this might be a very good way to calibrate player/team metrics.

[J.E.]

Anybody want to put together a blend of the various power ratings (those here, Hollinger's, maybe others) and compare the performance of an equal blend to the unique ones?

I think Hollinger's spread is too large, as he does not regress to the mean (there's a 25 point difference between first and last place).

I think EvanZ's Vegas rankings are probably the best. Maybe they need some (small) adjustment for recency but those are definitely the ones I would trust the most

[EvanZ]

Jerry makes a great point about Hollinger's ratings. JH is not the only one with a rating system that seems, um, not that useful. Here's one from RealGM that came out today that got an even more egregious variance than JH's:

Code: Select all

1. Philadelphia 76ers: 24.6  
2. Miami Heat: 18.5  
3. Chicago Bulls: 17.4  
4. Denver Nuggets: 14.5  
5. Portland Trail Blazers: 10.1  
6. Atlanta Hawks: 9.5  
7. Los Angeles Lakers: 8.8  
8. Los Angeles Clippers: 7.7  
9. Indiana Pacers: 6.1  
10. Oklahoma City Thunder: 6.0  
11. Orlando Magic: 5.0  
12. San Antonio Spurs: 4.7  
13. Dallas Mavericks: 3.2  
14. Phoenix Suns: 1.7  
15. Boston Celtics: 1.2  
16. Utah Jazz: 0.7  
17. New York Knicks: 0.5  
18. Cleveland Cavaliers: -0.4  
19. Toronto Raptors: -3.5  
20. Memphis Grizzlies: -5.5  
21. Houston Rockets: -5.6  
22. Milwaukee Bucks: -5.6  
23. New Orleans Hornets: -5.9  
24. Golden State Warriors: -7.4  
25. Minnesota Timberwolves: -7.8  
26. Charlotte Bobcats: -15.4  
27. Washington Wizards: -17.5  
28. Sacramento Kings: -20.4  
29. New Jersey Nets: -21.8  
30. Detroit Pistons: -27.8

Does this mean if Philly plays Detroit, that Philly should be favored by almost 60 points?!

[mystic]

Jerry makes a great point about Hollinger's ratings. JH is not the only one with a rating system that seems, um, not that useful.

Why is that not useful? The average scoring margin in real games is bigger than the average point spreads so far. Pulling the numbers all closer to the mean will likely only reduce the RMSE due to a smaller error, if the wrong winner is picked, not so much about how good a ranking is at picking the correct winner. My Ranking has also around 25 as difference between top and bottom. My average predicted scoring margin is 8.96 for the last 83 games, Vegas has only 6.39 while the reality has 11.99!

Did anyone tested the ability of predicting the correct winner instead of RMSE of the scoring margins? As I said before I tested it by pulling also all teams closer to the mean. The RMSE got lower (and I'm sure I can get that even smaller, if I go even further), but the Ranking got worse at predicting the correct winner.

As far as I see it RMSE is not necessarily a good indicator of the predicting power of a method when it comes to predicting the correct winner.

Btw, SRS on b-r.com has also 24 as difference.


Anyway, I'm surprised that the most rankings have the 76ers clearly as their best team so far while my ranking is putting them at 6th.