[Ganchrow]

As a model progress the correlation lowers naturally.

I don't think I understand what you mean here.

Can you clarify what is a "low correlation" in this context?

I'm just referring to a correlation coefficient as close to zero in magnitude as possible.

[Data]

I don't think I understand what you mean here.

I was talking about an ever evolving model. The better it becomes, the lower that correlation.

[Ganchrow]

I was talking about an ever evolving model. The better it becomes, the lower that correlation.

Gotcha.

[RockyV]

As a model progress the correlation lowers naturally. Can you clarify what is a "low correlation" in this context?

Sorry to bump such an old thread, but he is basically referring to the Orthogonality Principle (http://en.wikipedia.org/wiki/Orthogonality_principle).

Basically, we have some random variable X (e.g., the true spread for a game), and some observed variable Y (e.g., all the information publicly available to us at the time we place our bet.)

If our estimator of X (call it X_hat) is optimal, then the error variable E (defined as X-X_hat) should be statistically independent of the data Y. In other words, optimal estimators suck out all of the information available to us at the time, and use all that information as efficiently as possible.
One implication of this is that the error variable E also has to be statistically independent of any other estimators that are functions of Y. And since the line right when you place your bet is also a function of Y, then the error E associated with an optimal estimate is also independent of the line at that time.

Of course, testing independence statistically is impossible...so you'd instead probably want to use some sort of correlation coefficient as a surrogate.

I guess the cool thing about this is another way to evaluate the quality of an algorithm....you track your algorithm's line, the Vegas line at the time you placed your bet, and also what the final game spread was. You compute your error E for each bet (difference between your predicted spread and what the final spread was), and then the correlation coefficient between the vector E and the Vegas lines.

If your algorithm is bad, this correlation is probably high. If very small, then probably a very good algorithm.

Anyway, this is my understanding of things...please feel free to correct me if I am wrong.

[Data]

This is the finest thread I have ever read at SBR or any other forum and I'm guessing it has been read by fewer than 30 people and understood by fewer than 15.

As Wheell (one of the sharpest minds of the forum) said, this was THE THREAD. Look at the participants, undeniably, the sharpest bunch of SBR assholes, no squares in this thread (arguably, except yours truly). RockyV, congrats on bumping this up. You are on the right track.

[Pokerjoe]

It is an interesting thread, but didn't start out that way, to me, and would only have been a real winner if Wheel had been more explicit (understandably he wasn't).

Data wrote: What is the meaning or significance of calculating ROI this way? From economical standpoint, you have to count all the money you post up no matter whether you bet them or not.

If you put 20k offshore and a year later it's 40k, your ROI is 100% for the year.

When sportsbettors (some of them) talk about ROI, they mean something along the lines of, on that same 20k for example, having made a million dollars worth of bets and profited 20k for a 2% ROI. That's fine, but that isn't what the phrase ROI conventionally means, which was Data's point, imo.

And for predictive value--which is what sportsbettor's want, and for which purpose some of them have adopted ROI, in that they want to know how likely they are to continue their past returns--ROI isn't the best measure.

So really, how did the habit of using the investment concept ROI come into play in sportsbetting in the way it has? That is, in a way that isn't very useful? Somebody must have started it, LOL.

Oh, well, IMO.

Anyway, nice thread, guys.