How to calculate SRS?

[mopper8]

I am curious about how one goes about calculating SRS. Seems like it could get endlessly iterative if you adjusted every games MOV, but then once had a prelim SRS go back and REadjust based on opponent SRS rather than opponent avg MOV, and repeat the process over until the numbers stabilize. But it seems like there must be an easier way.

I'm curious about comparing Miami's SRS before Wade went down with injury, while he was out with injury, and since he's returned. It seems like their SRS and other advanced metrics (Ortg/Drtg) have been steadily improving since he got back.

Thanks for any help

[SideshowBob]

I was actually searching for this yesterday. I'm pretty certain that Neil had linked the formula on BBallRef a while ago, but I've yet to find it.

[SideshowBob]

I posted the question on the APBRmetrics board.

http://apbr.org/metrics/viewtopic.php?f=2&t=7898

SRS = MoV + SoS

SoS does not include HCA.

I'm pretty sure there was a different formula posted on BBallRef a few years ago, but for now, this formula does work, given BBallRef's listed SOS and MOV ratings for each team.

http://www.basketball-reference.com/leagues/NBA_2012.html#misc::none

For example, Chicago's current SRS = 9.12 + (-2.41) = 6.71

I'm not sure this helps with what you're trying to do Mop, but its a start.

FWIW, here's the link that I'm sure had the formula earlier, that seems to be dead now

http://www.pro-football-reference.com/blog/?p=37

[mysticbb]

SRS is OLS (ordinary least square). You can do this with R, if you like. But for your excerise just take the average SRS of the opponents as SOS and the average MOV. That will come already pretty close. And there is no HCA, which is bad. Adjust that for HCA (take a value from 2.8 to 3.6, I recommend 3.2) as the HCA. With HCA the prediction becomes better.

So, in the 9 games without Wade the Heat had 6 home games and 3 road games, that alone would give an expected MOV of 1.07, if an average schedule is played by an average team. The average SRS of the opponents was -0.17. So, the schedule was pretty easy with -1.24 SOS (adjusted for home court). The MOV was 12. That makes a SRS of 10.76 during those 9 games. That leaves 7.36 SRS in those 23 games with Wade.

Since Wade's return the Heat played 8 games on the road and 6 at home, makes it -0.46 as the expected MOV due to HCA. The average SRS of the opponents was -0.79, that makes the SOS overall -0.33. MOV was 11.07. So, overall it is 10.74 SRS. They played basically as strong with Wade after his return as they played in the 9 games without him.

[mopper8]

Awesome, thank you, makes perfect sense, and jives with the eye test. They basically played badly (relatively speaking) at the start of the season with him in the lineup but injured, and have played phenomenally ever since. ~10.75 SRS for the last 23 games and still seem to be improving...that's pretty good stuff.

[SideshowBob]

Would there be any advantage of using efficiency differential in place of MOV?

[SideshowBob]

Without adjusting for homecourt, Miami's been at a 19 SRS over the last 6 games

[mysticbb]

Would there be any advantage of using efficiency differential in place of MOV?

No. The point differential gives a better prediction. Which is completely understandble, because a team plays usually as their usual pace. The efficiency differential would be just adjusting that to a pace of 100. Well, it is more likely that a team with pace 92 plays the next game at that pace than playing now the upcoming games at a pace 100.

And I think you have the SOS backwards. The average SRS of the opponents for the last 6 games was -1.54. The Heat had a MOV of 17.67, that makes a SRS of 16.13 without the adjustment for HCA during those 6 games. Now, the HCA would push that to 18.26, given the fact that they played 5 on the road and only one at home. Very impressive stretch.

[EvanZ]

I wrote a post on my blog a while back about doing this for football (but it's obviously the same technique):

http://thecity2.com/2011/10/23/roll-you ... uby-and-r/

Then I think I described the process a little better in my stats primer:

http://thecity2.com/advanced-stats-primer/

I’m going to use football as a starting point. Imagine we want to know the strength (power rating) of every NFL team. The reason we want to know such a thing is so that we can predict the result when two teams face each other. To do this, we first assume that each team can be described by a single rating. This sounds obvious, but what it means is that regardless of the matchup, we can always use the same rating for each team. This is the assumption of linearity. An alternative non-linear assumption would be that teams have different ratings versus different teams. It’s not that the latter assumption is technically wrong; however, given the limited amount of data (i.e. games) in a single season, it is not a practical solution. Therefore, we stick with the linear assumption and assume that each team has a single rating. Given our assumption, let’s say that San Francisco is hosting Arizona and we want to predict the outcome:

MOV = HFA + SF - ARI = 3.0 + 5.0 - 3.5 = 4.5

This equation says the margin of victory (MOV) is equal to the home field advantage (usually around 3 pts) + San Francisco’s rating (which is close to roughly 5 at the time I’m writing this) minus Arizona’s current rating (roughly -3.5). In this case, the equation tells us that the predicted outcome is SF winning by 4.5 points.

But how do we get those ratings? Basically, we reverse the process mathematically and use the actual results from previous games. For example, in the recent “Harbowl”, Baltimore beat SF 16-6:

10 = HFA + BAL - SF

See, now we know the MOV from that game, but we don’t know any of the three values. We can set up an equation like this for every game that has been played. Once we’ve done that, we then use linear regression to calculate for us what the ratings were that most closely replicate the data for each game. Of course, there is a lot of error involved. It’s obvious that for each game, the results are not going to go exactly according to the ratings. Sometimes, the higher rated team loses, after all. However, on average, the weights/ratings determined from the regression give us the lowest error out of all the possible linear weights we could imagine. This is often called BLUE (best linear unbiased estimate).

[SideshowBob]

No. The point differential gives a better prediction. Which is completely understandble, because a team plays usually as their usual pace. The efficiency differential would be just adjusting that to a pace of 100. Well, it is more likely that a team with pace 92 plays the next game at that pace than playing now the upcoming games at a pace 100.

Makes sense. Thanks

And I think you have the SOS backwards. The average SRS of the opponents for the last 6 games was -1.54. The Heat had a MOV of 17.67, that makes a SRS of 16.13 without the adjustment for HCA during those 6 games. Now, the HCA would push that to 18.26, given the fact that they played 5 on the road and only one at home. Very impressive stretch.

Actually I think I just had a slip up in calculating the average MOV. My average SRS was -1.54