Predicting NBA Playoffs using RAPM (updated with Finals)

[colts18]

So far the first 4 games seem like vindication for xRPM predicting that the Spurs would dominate the Heat. I don't think any other system had the Spurs as a 80% favorite.

[Crow]

This was a worthwhile test. If I read right, 3 first round series misses, then one in second round and it looks likely to right the rest of the way. Pretty good.

I had wondered if the estimated RAPM errors for players might be unevenly distributed among teams and if that might hurt this effort. One test is not enough to conclude a lot definitively but one test is better than none.

[J.E.]

A test that's a little more meaningful would be to not look at which team won the series, but average point differential for the series, and whether the RAPM based forecast did better or worse (at predicting point differential) than simple team point differential or SRS. You definitely want to look at a larger sample though - the forecasts from SRS and RAPM are often going to be very similar, thus making it harder to compare the two

[Crow]

Good points.

[Crow]

I looked at the offensive and defensive RPM estimates for the starters on the 4 conference finalists.

The average RPM values were:

Position Off Def

PG 2.8 0.5
SG 1.1 1.4
SF 4.8 1.2
PF 1.7 2.5
C -1.9 2.5

Perkins had a ridiculous -7.5 on offensive RPM. All the centers were negative though (using Haslem for MIA).

OKC had 4 of the worst of these 10 estimates. Miami with 6 of the best. Spurs with one best, two worst. Bench carried them.

[mystic]

So far the first 4 games seem like vindication for xRPM predicting that the Spurs would dominate the Heat. I don't think any other system had the Spurs as a 80% favorite.

Well, I had the Spurs with +5.5 going into the finals. Using the minute distribution and the updated values after those 5 games, it becomes +6.0. Though, the difference was even bigger than that. If you need some sort of "proof" for that, please click: http://sportforen.de/showthread.php?590 ... st50557370 (at the end I wrote: Im bisherigen Verlauf der Playoffs spielten die Heat im Schnitt wie ein +8 Team, die Spurs wie ein +13.5 Team, das ist schon ein erheblicher Unterschied. Rough translation: Heat played like +8, Spurs like +13.5.)

Against Vegas I was 4-1, with the lines:

G1: -8.5 vs. -4.5
G2: -8.5 vs. -4
G3: +2.5 vs. -4
G4: +2.6 vs. -5
G5: -8.8 vs. -5.5

Not a particular great RMSE (was better last season), but still clearly better than Vegas. The interesting thing for me was the pretty obvious gambler's fallacy in G4, when the Heat all of the sudden were thought to be more likely to win that game after the game 3 loss. I used a -3 HCA, if we apply that to Vegas, they actually had the Heat in average for the first 4 games as +0.5 over the Spurs on neutral court.

As point of comparison: Last season's finals I had the Spurs with +1.8, they ended up being +1.2, I went 4-3 against Vegas, two times lost due to last minute actions, where luck was a big factor for the Heat. Vegas had the Heat in average at +2.9.

[Statman]

So far the first 4 games seem like vindication for xRPM predicting that the Spurs would dominate the Heat. I don't think any other system had the Spurs as a 80% favorite.

Well, I had the Spurs with +5.5 going into the finals. Using the minute distribution and the updated values after those 5 games, it becomes +6.0. Though, the difference was even bigger than that. If you need some sort of "proof" for that, please click: http://sportforen.de/showthread.php?590 ... st50557370 (at the end I wrote: Im bisherigen Verlauf der Playoffs spielten die Heat im Schnitt wie ein +8 Team, die Spurs wie ein +13.5 Team, das ist schon ein erheblicher Unterschied. Rough translation: Heat played like +8, Spurs like +13.5.)

Against Vegas I was 4-1, with the lines:

G1: -8.5 vs. -4.5
G2: -8.5 vs. -4
G3: +2.5 vs. -4
G4: +2.6 vs. -5
G5: -8.8 vs. -5.5

Not a particular great RMSE (was better last season), but still clearly better than Vegas. The interesting thing for me was the pretty obvious gambler's fallacy in G4, when the Heat all of the sudden were thought to be more likely to win that game after the game 3 loss. I used a -3 HCA, if we apply that to Vegas, they actually had the Heat in average for the first 4 games as +0.5 over the Spurs on neutral court.

As point of comparison: Last season's finals I had the Spurs with +1.8, they ended up being +1.2, I went 4-3 against Vegas, two times lost due to last minute actions, where luck was a big factor for the Heat. Vegas had the Heat in average at +2.9.

I also did a quick finals prediction recap based on my regular season player ratings. The general regular season team rating had the Spurs +2.8, after seeing how playing time was allotted for the finals, +3.6. I didn't use any playoff data, which obviously would have made the Spurs an even bigger favorite - I was just trying to show some of the validity of the regular season player ratings. I talk about Kawhi Leonard some too:

http://hoopsnerd.com/?p=595

[Mike G]

After the fact "predictions" are something else. Postdictions, perhaps.
An exchange in another thread ("2014 Finals") that may be more appropriate here:

... I used a minute and pace prediction, got the Heat with +8 per 100 poss, the Bobcats with -4 per 100 poss with about 87 pace for +10.5...

The SRS differential for Mia-Cha is about 5 ppg. You give Mia another expected 6 ppg, based on ... ?
- Jefferson averaging 8 fewer mpg, picked up by McRoberts? Everyone else within 3 mpg of RS for Cha
- Wade was full-time, Ray played less, Battier not at all; Haslem and Jones played more than usual.

SRS differential is meaningless given the severe differences in terms of minute distribution. Gary Neal is a -4 player in my metric, for example, I expected him to get 24 mpg ... I had the Heat clearly better than their RS SRS, the Bobcats being worse. Again, look at the minute distribution, then use a reliable player metric and you get a MUCH better impression about the strength of the teams.

I admit to being somewhat mystified.
After the Bobcats traded Sessions for Neal (et al.) they went from 25-30 (-1.8 ppg) to 18-9 (+2.8)

How does one juggle the minutes to get them to -4 per 100 poss ?
If you go by their season SRS (-0.89) or figure their season-ending run was some 2 points better than that, you would have to really sabotage their rotation to be 3 to 5 points worse.

Guessing at player minutes, and then guessing at player values -- while dismissing SRS as "meaningless" -- I don't see how you'd get improved predictions.

[mystic]

After the fact "predictions" are something else. Postdictions, perhaps.

My lines I played in reality for the Heat vs Bobcats series:

My line was -13.5, -13.5, +7.6, +7.6 (slight change after the first two games from +10.5 to +10.6) vs. Vegas lines of -9.5, -9, +4.5 and +7; I went 3-1 in that series against Vegas.

That really happened. The lines are based on the prediction I made before the fact. Same thing for the Heat series, as you may notice, the post I made on the german message board was on June 2nd, the first game was June 5th!

Guessing at player minutes, and then guessing at player values -- while dismissing SRS as "meaningless" -- I don't see how you'd get improved predictions.

I predicted player minutes based on a model which took RS minutes and the development in average for PS minutes of the past into account. That wasn't just "guessing". I also didn't guess at player values, but I use a model based on my own boxscore-based SPM model and RAPM (combined via OLS) for those values. That model and my minute prediction combined gave me about 60% wins against Vegas over the past 3 seasons (only played this season, just tested it for 2012 and 2013 with PREDICTIONS based on the model)! Matter of fact is: RS SRS is a worse predictor than that.

Just to make that clear: Working with stats and making predictions is not my hobby anymore, it was my everyday job for the past season.

[Mike G]

OK, trade secrets, I guess.

[mystic]

Well, my point all along was that RS SRS isn't as a good as a predictor as a reliable player metric. You may look at this very thread. Anything else is not that important, thus I wouldn't say "trade secrets" is an appropriate answer.

[Mike G]

So, how does a +2 team become a -4 team with your predicted minutes redistribution?

Jefferson averaging 8 fewer mpg, picked up by McRoberts.. Everyone else within 3 mpg of RS for Cha

[mystic]

So, how does a +2 team become a -4 team with your predicted minutes redistribution?

See, it starts with your assumption that a team would be a +2 team based on a smaller sample. Maybe the Bobcats weren't a +2 team to begin with? Because, quite frankly, they weren't. That's why I pointed out before that runs in smaller samples happen, but I would never base a prediction on such thing. But I think we will run in circles here, because you will stick with your belief that the Bobcats were much better than I saw them (or Colts18 prediction in this thread using RPM as the player metric), and I will not go further into details, because, tbh, I don't see how that would be beneficial for me.

[talkingpractice]

So, how does a +2 team become a -4 team?

This is way off the main topic, but it sounded weird to me too, so I went back and checked what we had. I agree that CHA was projected to be a bit worse than in the regular season (and the Heat way better), due in large part to Gary Neal presumably getting mucho minutes, and Al probably getting less than normal minutes. But that makes them maybe a -1.5 or -2 or something (CHA's SRS was -1). I don't see that -4 there anywhere. Al was crippled in that series which isn't accounted for in that -2ish estimate. But I don't see a -4 for that Bobcats team, if you ignore the crippled issue.

SPM's grossly overvalue Al and grossly undervalue MKG and Kemba, so I don't think you'd want to let an SPM loose near the Bobcats. And if the Heat looked particularly dominant in that series, it was probably also mostly the crippled issue.

[mystic]

But I don't see a -4 for that Bobcats team, if you ignore the crippled issue.

From my prediction model for the playoffs (before the 1st game was played):

Code: Select all

Name                 min    v
Kemba Walker          38   1.02
Josh McRoberts        36   0.08
Al Jefferson          35   0.82
Gary Neal             30  -4.21
Gerald Henderson      30  -1.58
Michael Kidd-Gilchri  24   0.02
Bismack Biyombo       13  -2.46
Anthony Tolliver      12   0.45
Chris Douglas-Robert  12  -1.33
Luke Ridnour          10  -2.73

Makes -3.6. I had the Heat with 8.4, which makes +12 per 100 poss, an estimated pace of 87.5 for +10.5. And that was the line I played!