You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
We had question about computing r-squared and loglike for out-of-sample data for cross-validation. (I don't find any related issues right now)
Problem is that r-squared requires assumption on the intercept, or more generally params of the null model.
I think now that we can get a consistent interpretation of R-squared and similar measures if we evaluate the new exog at the null model for comparison. To be able to do that, we need to keep the null model or at least some sufficient statistics like params around.
In the simple OLS rsquared case, this would be using the intercept estimate from the training/estimation data, i.e. model.endog.mean(), or zero if no constant has been added.
Some statistics are in tools.eval_measures
The text was updated successfully, but these errors were encountered:
We had question about computing r-squared and loglike for out-of-sample data for cross-validation. (I don't find any related issues right now)
Problem is that r-squared requires assumption on the intercept, or more generally params of the null model.
I think now that we can get a consistent interpretation of R-squared and similar measures if we evaluate the new exog at the null model for comparison. To be able to do that, we need to keep the null model or at least some sufficient statistics like params around.
In the simple OLS rsquared case, this would be using the intercept estimate from the training/estimation data, i.e.
model.endog.mean()
, or zero if no constant has been added.Some statistics are in
tools.eval_measures
The text was updated successfully, but these errors were encountered: