[Dougie Fresh]

Before I seriously dive into a game I look to at PPG and points allowed per game. Mid way through the season I don't like using the season to date numbers because I believe a lot of teams have ups and downs through out the season and significant injuries can skew the numbers.

What are the thoughts on sample size using theses stats? 5 games? 10 games?

[chemist]

Before I seriously dive into a game I look to at PPG and points allowed per game. Mid way through the season I don't like using the season to date numbers because I believe a lot of teams have ups and downs through out the season and significant injuries can skew the numbers.

What are the thoughts on sample size using theses stats? 5 games? 10 games?

You could look at how the prediction error varies with the window size. The usual caveats about back fitting apply.

[dwaechte]

You could look at how the prediction error varies with the window size. The usual caveats about back fitting apply.

Bingo. I think you're most likely to find that a progressive weighted system will work better than a flat-out cutoff.

[chemist]

Bingo. I think you're most likely to find that a progressive weighted system will work better than a flat-out cutoff.

Yeah, I usually use exponential smoothing, which is just

s_t+1 = (1-α)s_t + αx_t+1

α = 2/(T+1) supposedly smooths about as much as a rectangular window of length T. What you do at the start of the season is an open question.

[MonkeyF0cker]

That is why I prefer a weighted system using player stats rather than team stats. How effective your model will be during injury or "ups and downs" is one thing, but how ineffective will it be after a major trade?

[Justin7]

Another thing to look at is the starting rotation. A set of 5 games with the same players might be more consistent than 100 with 10 different rotations.