Suppose I want to make predictions of a response from predictors but I have some autocorrelation in the response variable. Under OLS this would be a problem as the residuals would have autocorrelation. What if I just want to predict the response and I use regularized least squares, like lasso or ridge or elastic net? I don't care about variances of the coefficients or anything of that nature as I'm not testing any hypotheses but I feel like I might be missing something.
$\begingroup$
$\endgroup$
1
-
$\begingroup$ I have some autocorrelation in the response variable. Under OLS this would be a problem as the residuals would have autocorrelation. No, not necessarily. If you find a suitable model, the residuals will be autocorrelation-free. $\endgroup$– Richard HardySep 28, 2017 at 5:30
Add a comment
|
1 Answer
$\begingroup$
$\endgroup$
I don't care about variances of the coefficients or anything of that nature
In this case OLS will not have an issue too with coefficients of your model.
The problem is with things like seasonality, which is going to be the same in OLS or regularized regression unless you take care of it.
