0
$\begingroup$

I'm sorry I don't have reproducible code, but I have a pretty specific question that I can't find an answer to.

I'm using randomForest to project NBA statistics. Vegas-odds are incredibly useful because it's provides the wisdom of the crowd. Intuitively I feel like they need to be standardized for analysis, but maybe randomForest is good enough.

The reason why I feel it needs to be standardized is because it's disjoint. If a team has a moneyline of -125, that means that you must pay \$125 to win \$100 (payout of \$225). If a team has a betting line of +110, that means you need to bet \$100 to win \$110 (payout of \$210). Therefore, it's disjoint in that there would never be scores in (-100, 100) since +100 or -100 are both even odds.

With that said, would you recommend reshaping the vector in some way so that the random forest can learn "better"? E.g. -125 is a (125/(100 + 125) 55.6% chance of winning and a +110 is a (100/110+100) 47.6% chance of winning. Would changing the moneylines to percentages help performance? I know the only surefire way to check would be to run models, but I really don't have time for it at the moment, and this question will help me to determine in general if/when standardizing is necessary.

$\endgroup$
0
$\begingroup$

This is directly related to the idea of calibrating probability values produced by a random forest. Aside from the link, there is substantial literature on how to do this. The simplest approach basically amounts to fitting a logistic regression on the outputs of the random forest to change the response surface into a logistic form.

Once the predictions have this form, you've demonstrated that you have the knowledge to turn the probability estimate into Vegas odds, so that should be a straightforward process for you.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.