Predicting Winning Percentage

[jaebradley]

Anyone know of any methods that successfully predict a team's winning percentage in the next season within a reasonable accuracy? Or successfully predict a team's winning percentage for the rest of the season after, say, 10% of the regular season?

[DSMok1]

It's tough. Here is this season's predictions thread:

viewtopic.php?f=2&t=8068

[jaebradley]

Here's what I did:

Do you know the Pythagorean Expectation Formula?

If not, here is the wiki link: http://en.wikipedia.org/wiki/Pythagorean_expectation

Basically, it estimates winning percentage using points scored and points scored against. I project points scored/scored against per 100 possessions by using data from the first month of every NBA season from 2001-2002 to 2010-2011, and then OLS regressing the first month data on final regular season points scored/scored against per 100 possessions from each of these seasons. This regression estimates what a team's final regular season points scored/scored against per 100 possessions is given data from the first month of the NBA season. Then, using the estimated equation, I estimated final regular season points scored/scored against for 2011-2012 and 2012-2013 given data from the first month of the 2011-2012 and 2012-2013 seasons. Then, I used the Pythagorean Expectation Formula to predict winning percentage (according to the wiki link and other sources, such as Basketball Reference, the exponent used the Pythagorean Expectation Formula is 13.91) for the past two seasons.

I couldn't find readily available Vegas over-under numbers for the 2011-2012 NBA season, but for the 2012-2013 season, using my relatively simple (and relatively crude methodology) I would have been predicted the correct over-under 20 times and been wrong 10 times. This isn't too bad considering that Schoene's predicted the correct over-under 17 times and got it wrong 11 times (Schoene and the over-under had the exact same value twice, so I didn't count two predictions). John Hollinger would have gotten 22 teams right on the over-under and 8 wrong.

Obviously, it's easier to predict the NBA season once it's started, but after only a month, you can use a very crude method to find a relatively decent predictive model.

How does all this sound? Any questions? Comments? Etc.

Why did I use only a month of data versus, say, the first 15 games each team played?
I was using data from NBA Stats and they didn't have this data category readily available. However, I could search for statistics within a certain time period. So I decided to obtain data for the first month of the season, even if each team didn't play the same number of games.

[DSMok1]

If you have the first month of data, its WAY easier to predict the season, other than injuries.

Predicting beforehand is much harder.

[jaebradley]

Right. I said that. I think my main point was that even after a month (which is still a pretty small sample) you can still predict relatively accurately given a relatively crappy model.

[talkingpractice]

DSMok1 knows the Pyth Formula .

[colts18]

I couldn't find readily available Vegas over-under numbers for the 2011-2012 NBA season.

Here are 2 sources I found:

http://thepaintedarea.blogspot.com/2011 ... tions.html

http://www.sportsmemo.com/content/print ... og_id=8694

[Neil Paine]

Right. I said that. I think my main point was that even after a month (which is still a pretty small sample) you can still predict relatively accurately given a relatively crappy model.

The first month isn't that small of a sample in the NBA. 12 games is all you need to regress W-L records halfway to the mean (that is, half of observed record is skill and half is luck). That means 12 NBA games convey roughly the same amount of info about team talent as 11 NFL games, or 70 MLB games.

So, to use baseball as the comparison, we won't reach the MLB equivalent of the first month of the NBA season until mid-to-late June (another 3 weeks or so).

When put that way, it obviously will be a LOT easier to predict the final baseball standings 3 weeks from now than it was 2 months ago. And that's a bit what it's like to predict the NBA standings after just one month of the season (although teams do have fewer "banked" wins after a month of NBA than 3 months of MLB).