APBRmetrics

I decided to try to predict the NBA playoffs using the xRAPM values that J.E. provided on ESPN because those are supposed to be predictive.

O rating, D rating, Total per 100 possessions
Spurs 6.6 6.9 13.4
Mavs 6.6 0.5 7.1

Thunder 6.1 4.1 10.2
Grizzlies 1.7 5.6 7.3

Clippers 10.0 2.4 12.4
Warriors 4.4 7.4 11.9

Rockets 5.8 3.7 9.5
Blazers 6.6 0.2 6.8

The Clippers look to have the best offense by this measure while the Warriors are the best defense even though I gave no minutes to Andrew Bogut. By RAPM the Spurs are the clear top team followed by Clippers, Warriors, and Thunder.


Here are the MPG values I used. Feel free to give me suggestions so that I can update these Minutes more accurately.

Eastern Conference playoff rankings

O rating, D rating, Total per 100 possessions
Pacers -0.9 7.0 6.1
Hawks -0.6 0.2 -0.4

Heat 7.4 2.9 10.3
Bobcats -1.6 -0.3 -1.8

Raptors 2.9 0.6 3.4
Nets 3.1 2.1 5.1

Bulls -0.8 5.1 4.3
Wizards 0.5 3.1 3.6


Miami is the favorite by a wide margin based on RAPM. The Bulls-Wizards series should be a good one

Top teams in the playoffs by RAPM

1 Spurs 13.4
2 Clippers 12.4
3 Warriors 11.9
4 Heat 10.3
5 Thunder 10.2
6 Rockets 9.5
7 Grizzlies 7.3
8 Mavs 7.1
9 Blazers 6.8
10 Pacers 6.1
11 Nets 5.1
12 Bulls 4.3
13 Wizards 3.6
14 Raptors 3.4
15 Hawks -0.4
16 Bobcats -1.8


MPG values for Eastern conference playoffs

Speights has been killing GSW all season long. His minute forecast will have a significant impact on their overall strength forecast

The predicted score differences are probably a little too wide, given what we now know about the 'effect of being up X', which the RPM ratings are adjusted for. (Also, the ratings are per 200 possessions, but each game only has about ~190 on average)

I'd guess the only series where this differs from the regular forecast (using Sagarin, point differential, SRS) is Toronto v. Brooklyn. Should be interesting

Nando de Colo won't be commuting between Tor and SA. He's a Raptor.

Compared to the current vegas lines (pinnacle) for all first games of each series most of the predictions posted by colts18 are similar

The largest differences are:
Clippers by 7. Colts18's method says 4 when factoring in HCA
Heat by 9.5 (15)

Their BRO@TOR line almost exactly matches what colts18 has posted (Toronto by 2)

take it for whatever it's worth, but I think the Raptors are a great play. I do not understand the series price on that one at all, I have a much much different view on it. pretty excited for the postseason!

Does anyone know how to convert SRS into win% or odds so that can I create a simulation for the playoffs?

colts18 wrote:Does anyone know how to convert SRS into win% or odds so that can I create a simulation for the playoffs?

Pythagorean method: http://www.basketball-reference.com/about/glossary.html

W Pyth
Pythagorean Wins; the formula is G * (Tm PTS14 / (Tm PTS14 + Opp PTS14)). The formula was obtained by fitting a logistic regression model with log(Tm PTS / Opp PTS) as the explanatory variable. Using this formula for all BAA, NBA, and ABA seasons, the root mean-square error (rmse) is 3.14 wins. Using an exponent of 16.5 (a common choice), the rmse is 3.48 wins. (Note: An exponent of 10 is used for the WNBA.)

Have you built lagged weights into the model? The Nets are almost certainly a stronger team now than they were for the first 1/3 or so of the season. That doesn't mean that we should ignore the first 1/3 of the season, but it probably would be appropriate and potentially more accurate to give less weight to the earlier games and more to the later games. A similar principle to setting Bogut's minutes to 0, i.e. update in accordance with new information.

I converted these values into a win% vs the opponent. This is odds on a neutral court. The Spurs would win 70.3% vs the Mavs on a neutral court. My numbers have the Clippers-Warriors series as a dead heat. The Clippers are expected to win 50.03% of neutral court games vs GSW. Thats 3 extra wins per 10,000 games


Spurs 0.703
Mavs

Thunder 0.586
Grizzlies

Clippers 0.500
Warriors

Rockets 0.590
Blazers


Pacers 0.706
Hawks

Heat 0.827
Bobcats

Raptors 0.444
Nets

Bulls 0.525
Wizards


Here are those values adjusted for HCA. I used 3.5 points for HCA. The values next to the team is their expected home win% vs the Opponent. For example the Spurs are expected to win 79.3% of the time at home vs the Mavs while the Mavs are only expected to win 40.5% of the time when they play the Spurs at home.

Spurs 0.793
Mavs 0.405

Thunder 0.698
Grizzlies 0.535

Clippers 0.618
Warriors 0.618

Rockets 0.699
Blazers 0.528


Pacers 0.798
Hawks 0.406

Heat 0.885
Bobcats 0.253

Raptors 0.564
Nets 0.669

Bulls 0.645
Wizards 0.597

Series odds based on 1,000 simulations.

Spurs 90.8%
Mavs

Thunder 73.9%
Grizzlies

Clippers 55.9%
Warriors

Rockets 74.9%
Blazers


Pacers 91.2%
Hawks

Heat 98.0%
Bobcats

Raptors 44.2%
Nets

Bulls 62.6%
Wizards

A few thoughts:
-Your minutes projections for the Spurs' stars seem low. Last year, Duncan, Kawhi, and Parker each saw large increases in min/g in the postseason, ending at around 36 each. Some of that was due to the Spurs playing a total of 7 overtimes in their 21 games (adding 1.4 minutes per game, or 7 additional player-minutes) or due to them having slightly less depth last year, but I'd bump up the minutes of each of those three, as well as Ginobili. Popovich spoke about Kawhi in particular here, saying,

“We want to up his minutes,” Popovich said. “He’s going to play more minutes (in the playoffs) than Tim Duncan does probably, more minutes than Manu Ginobili probably. This is his stretch run and he needs to be in shape for it. He’s never really been able to do this because it was a lockout season or we had to limit his minutes last year. This is the first time he’s been able to lay it out there.”

I'd project Kawhi at no less than 37 min/g and probably more, especially considering that if seeds hold, the Spurs' road would go through James Harden, Kevin Durant, and LeBron James. The extra minutes for the stars would come from Belinelli, Diaw, Mills, and Splitter at the back end of the rotation, I'd expect. Good lord, the Spurs are deep.

-Jerryd Bayless is no longer with the Grizzlies, having been traded to Boston in the deal that acquired Courtney Lee. I'd expect most of his projected minutes to instead go Mike Miller's way.

-Adding xRAPM values up like that may make sense, but calling them SRS values isn't passing the smell test. The Spurs, Clippers, and Warriors all would come in as the #1 overall team of all time were they to keep those values up for an entire season, and we'd have 10 teams with an SRS north of +6. If each value was multiplied by .6, the magnitudes would seem more reasonable, but it seems like the "effect of being up X" variable is having quite a large effect. Maybe there's some function we can apply to these totals, so that we know that when the Spurs are given a score of 13.4, due to the extra difficulty in adding each additional point of efficiency differential, that translates out to the difference between 0 and +8, or +9 or +7 or whatever it is.

-It also seems strange that Warriors-Clippers ended up that close in your model. How did you set that up? If you used the 12.4 and 11.9 values from before, a half-point difference is not insignificant, and should result in a swing of more than 3 games per 10,000, right? Am I missing something there? It may be that the model just minimizes differences as these scores grow. Strange behavior with large inputs like that wouldn't surprise me much considering how extreme of an outlier these are when interpreted as SRS scores, but I don't see how that would have that effect if you just generated predicted PPG for each team and plugged it into Pythagorean W/L. I might just have misunderstood what your inputs are, though.

ryannow wrote:A few thoughts:
-Your minutes projections for the Spurs' stars seem low. Last year, Duncan, Kawhi, and Parker each saw large increases in min/g in the postseason, ending at around 36 each. Some of that was due to the Spurs playing a total of 7 overtimes in their 21 games (adding 1.4 minutes per game, or 7 additional player-minutes) or due to them having slightly less depth last year, but I'd bump up the minutes of each of those three, as well as Ginobili. Popovich spoke about Kawhi in particular

My MPG projections were based on regular season MPG, adding minutes for starters, and looking at last years 1st round MPG. The Spurs are hard to project because no one played 30 MPG for them this season.

-Adding xRAPM values up like that may make sense, but calling them SRS values isn't passing the smell test. The Spurs, Clippers, and Warriors all would come in as the #1 overall team of all time were they to keep those values up for an entire season, and we'd have 10 teams with an SRS north of +6. If each value was multiplied by .6, the magnitudes would seem more reasonable, but it seems like the "effect of being up X" variable is having quite a large effect. Maybe there's some function we can apply to these totals, so that we know that when the Spurs are given a score of 13.4, due to the extra difficulty in adding each additional point of efficiency differential, that translates out to the difference between 0 and +8, or +9 or +7 or whatever it is.

That's because I gave a lot of minutes to star players and less to scrubs. Imagine a hypothetical world where Miami is healthy every game With no game to game fatigue and they play all out in that game without using scrubs, they are probably a +9 or +10 team in that scenario. The Spurs +13.4 xRAPM value is similar to the 96 Bulls +13.4 efficiency differential. If they played healthy and all out every game, they get around 70+ wins and challenge the 96 Bulls.

-It also seems strange that Warriors-Clippers ended up that close in your model. How did you set that up? If you used the 12.4 and 11.9 values from before, a half-point difference is not insignificant, and should result in a swing of more than 3 games per 10,000, right? Am I missing something there? It may be that the model just minimizes differences as these scores grow. Strange behavior with large inputs like that wouldn't surprise me much considering how extreme of an outlier these are when interpreted as SRS scores, but I don't see how that would have that effect if you just generated predicted PPG for each team and plugged it into Pythagorean W/L. I might just have misunderstood what your inputs are, though.

I converted SRS into a Pythagorean W-L%. While the Clippers had a higher SRS, they are virtually tied in the pythag% because Pythagorean tends to value better defenses if two teams are the of the same quality. The Clippers have a much better offense so they get punished a little by it. A team that wins 95-85 gets a higher pythagorean value than a team that wins 105-95 even though they both have a +10 PPG differential.

colts18 wrote: My MPG projections were based on regular season MPG, adding minutes for starters, and looking at last years 1st round MPG. The Spurs are hard to project because no one played 30 MPG for them this season.

True, and if it's just the first round, that makes more sense, as I was reading these as being for the entirety of the playoffs. My comment wasn't meant as a criticism (and I don't think it would change the values much since Kawhi comes out surprisingly low in xRAPM), I was just responding to "feel free to give me suggestions so that I can update these Minutes more accurately." I'd be very surprised if Kawhi remained at 30 mpg.

-Adding xRAPM values up like that may make sense, but calling them SRS values isn't passing the smell test. The Spurs, Clippers, and Warriors all would come in as the #1 overall team of all time were they to keep those values up for an entire season, and we'd have 10 teams with an SRS north of +6. If each value was multiplied by .6, the magnitudes would seem more reasonable, but it seems like the "effect of being up X" variable is having quite a large effect. Maybe there's some function we can apply to these totals, so that we know that when the Spurs are given a score of 13.4, due to the extra difficulty in adding each additional point of efficiency differential, that translates out to the difference between 0 and +8, or +9 or +7 or whatever it is.

That's because I gave a lot of minutes to star players and less to scrubs. Imagine a hypothetical world where Miami is healthy every game With no game to game fatigue and they play all out in that game without using scrubs, they are probably a +9 or +10 team in that scenario. The Spurs +13.4 xRAPM value is similar to the 96 Bulls +13.4 efficiency differential. If they played healthy and all out every game, they get around 70+ wins and challenge the 96 Bulls.

The numbers still just don't add up. I ran the exact same equation, but with the actual total minutes played for this year's Spurs, and I came out with a value of 10.2. Did it for the Warriors and got 7.2. Both of those are more than 2 points higher than the team's actual SRS in that time period. It makes sense why this would happen: if you add one +4 player to a team at 0, you can expect them to go to around +4, but that doesn't mean a second one would lift the team to +8, or a third one to +12. That's what the "effect of being up x" term is attempting to capture; it's harder to take a team that's already very efficient and boost it by n points than it is to take an inefficient team and boost it by n points. As a result, you can take +10 worth of players, and end up with a +8 team, although I'm not sure exactly what the shape of this curve is in J.E.'s model.

I converted SRS into a Pythagorean W-L%. While the Clippers had a higher SRS, they are virtually tied in the pythag% because Pythagorean tends to value better defenses if two teams are the of the same quality. The Clippers have a much better offense so they get punished a little by it. A team that wins 95-85 gets a higher pythagorean value than a team that wins 105-95 even though they both have a +10 PPG differential.

That makes sense, but I'm still not following 100%. Did you generate a separate pythagorean winning percentage for each team by setting up Team PTS and Opp PTS as (average + sum of O) and (average - sum of D), respectively? In that case, I'm not sure how one combines the two winning percentages to find win% against each other. If you generated one win% for each series by setting, say, LAC pts= average + sum of LAC O - sum of GSW D, then it seems like offensive and defensive strength would cancel out and the Clippers would still be projected a half-point ahead per-game. I'm not sure which approach is better.

ryannow wrote: The numbers still just don't add up. I ran the exact same equation, but with the actual total minutes played for this year's Spurs, and I came out with a value of 10.2. Did it for the Warriors and got 7.2. Both of those are more than 2 points higher than the team's actual SRS in that time period. It makes sense why this would happen: if you add one +4 player to a team at 0, you can expect them to go to around +4, but that doesn't mean a second one would lift the team to +8, or a third one to +12. That's what the "effect of being up x" term is attempting to capture; it's harder to take a team that's already very efficient and boost it by n points than it is to take an inefficient team and boost it by n points. As a result, you can take +10 worth of players, and end up with a +8 team, although I'm not sure exactly what the shape of this curve is in J.E.'s model.

You ran the projection and got +10.2 for the Spurs because though are actual minutes. I'm using playoff strength. The Spurs aren't playing many scrubs in the playoffs so their strength is higher than +10.2. Plus J.E. already said that his xRAPM takes into account the leading by X effect.

That makes sense, but I'm still not following 100%. Did you generate a separate pythagorean winning percentage for each team by setting up Team PTS and Opp PTS as (average + sum of O) and (average - sum of D), respectively? In that case, I'm not sure how one combines the two winning percentages to find win% against each other. If you generated one win% for each series by setting, say, LAC pts= average + sum of LAC O - sum of GSW D, then it seems like offensive and defensive strength would cancel out and the Clippers would still be projected a half-point ahead per-game. I'm not sure which approach is better.

I used the pythagorean win% from DSmok1's posts for both teams. Then I used the log 5 method to generate an expected win%. Log 5 formula:

(home win%/Home loss%)/(home win%/home loss%+road win%/road loss%)