APBRmetrics

back2newbelf · Joined: 21 Jun 2005 Posts: 271

I just did a rather large test on point differential and SRS, how well they predict the winner of a game in all NBA seasons past. I don't have very interesting findings but figured it might be worth a thread. (spare someone else work, maybe)

What i did:
Start predicting the winner of a game, starting with the 1st of January of every season, using the data we gathered(this season) before the date of the game.

1st method: Point differential, pace adjusted, i also subtract that years' homecourt advantage for every home game and add it for every away game a team has played.

2nd: SRS. I tried messing with how far back i should go when i calculate the rating of a team. Example: for predicting games in February, i would calculate the SRS rating of a team using only games played in January and so far in February.

Findings:
SRS does(did) better than point differential but the difference is tiny.
Using only the last X days to calculate SRS rating does better than normal SRS, but again, tiny difference.

# of games that i tried to predict:27779
Correctly predicted by PD: 19103
Correctly predicted by SRS: 19124
Correctly predicted by SRS(using only the last 3 months to calculate a teams's SRS rating): 19158

Mountain · Joined: 13 Mar 2007 Posts: 1527

Huge test, impressive.

Even when you didn't find earth-shattering results so far, better to check and know more.

Any difference in predictive accuracy when point differential is 0-5. 5-10, 10+ and if this in combination with home court matters?

Is your dataset and interest such that you could test for predictive accuracy when pace differential is 0 to x, x to y and y to z. etc. and if this in combination with home court matters?

Testing for significance of 4 Factor differentials might be interesting too.

gabefarkas · Joined: 31 Dec 2004 Posts: 1313 Location: Durham, NC

When you say you adjusted for pace, do you mean you used their pace over the season as a whole or their pace only in games played up to that point, or their pace for that particular game?

back2newbelf · Joined: 21 Jun 2005 Posts: 271

Mountain · Joined: 13 Mar 2007 Posts: 1527

Any difference in predictive accuracy when point differential is 0-5. 5-10, 10+

Haven't tested this yet, but i would assume there is a difference. I'll have an answer in a couple of days

Sounds good.

and if this in combination with home court matters?

I don't understand this part, sorry

I meant does a point differential of 0-5, etc. after home court edge is included at home or away behave differently in terms of prediction success from same differential away.

Testing for significance of 4 Factor differentials might be interesting too.[/quote]
I have game results only, no boxscores

Ok.

94by50 · Joined: 01 Jan 2006 Posts: 499 Location: Phoenix

As long as we're discussing SRS, I am surprised by how often the predicted MOV by SRS is close or equal to the Vegas line. (It might be selective memory - I don't have hard data, just general observation.)

Kevin Pelton · Posted: Fri Jan 23, 2009 2:29 am Post subject:

Wouldn't it be more interesting to compare their performance relatively early in the season? Certainly the first week or so things are going to be random no matter how you go. However, it seems to me that November and December are going to be when you're going to see the biggest schedule disparities, which is what is accounted for by SRS.

I guess it doesn't surprise me that it ends up about the same as point differential later in the season, since SRS should generally match differential besides conference differences, no?

back2newbelf · Joined: 21 Jun 2005 Posts: 271

back2newbelf · Joined: 21 Jun 2005 Posts: 271

predictive accuracy for different ranges of point differential(league average homecourt advantage factored in):

0-5 difference in average point differential: 57%
5-10 difference in average point differential: 71%
10+ in average point differential: 84%

Mountain · Joined: 13 Mar 2007 Posts: 1527

Thanks for this followup.

Harold Almonte · Joined: 04 Aug 2006 Posts: 616

I've tried an empirical formula, which I don't know the logical behind, but with good approachs: ( (teamA pythW% - teamB pythW%)/4)/10 +/- Home Adv.

Note: when an western team goes to east, the HA increase one extra point.

Mountain · Joined: 13 Mar 2007 Posts: 1527

"(teamA pythW% - teamB pythW%)/4)/10 +/- Home Adv"

Harold tell me if I am interpreting this correctly

...
(edited, I misunderstood)

How well does this match with actual results?

Harold Almonte · Joined: 04 Aug 2006 Posts: 616

No. It's not about prediction.

If you use whole percents, forget the 10 or the hundred.

75-50=25,
25/4=6.25

Team A will give about 9 points at home (to team B). Will give about 3 on the road.

The HA is not the same for all teams and situations, you can play with that, but I´ve obtained numbers very close to Vegas lines, even in college hoop (4.5 HA).

How the .... the division by 4 translates percents to points spread? I don't know, but it's strangely too coincidential. However, the Vegas line is just an about 68% straight win predictor like SRS.

back2newbelf · Joined: 21 Jun 2005 Posts: 271

Exponential (moving) average correctly predicts 27462 out of 40417.
Point differential: 27516.

Mountain · Joined: 13 Mar 2007 Posts: 1527

So 70% appears to be the upper limit on predicting W or L by these methods. Maybe some hybrid or new approach does a bit better but there are going to be a share of surprises.