yforecasting / leadingindicators Goto Github PK

Hi Yves,

I am very appreciative of your leadingindicators package, it's coming in really handy. I have a question regarding the selection of lambda within the build_lasso function. I noticed that the cv.glmnet is conducting cross-validation to train a LASSO model with no specification for lambda. So cv.glmnet computes a sequence of lambda values to use based on the data (number of observations and number of variables). It produces both lambda.min, the value of lambda that gives the lowest mean cross-validated error and the lambda.1se, the largest value of lambda such that error is within 1 standard error of the minimum. The lambda fed into build lasso isn't called until the call to capture the coefficients, in which we capture the coefficients according to our original lambda. My question is, how come you choose to set up the function in this way? Thinking about it, maybe it's because you want to determine the accuracy of the specified lambda on the subsequent forecast, is it correct that that is the case? If so, specifying the set of lambda values to test and passing it to cv.glmnet may save a significant amount of time when searching for hyperparameters.

Curious on your thoughts about outputting the model results from 'labda.1se' and extracting those coefficients as an alternative to the specified lambda solution? Seems like it could be reasonable, but I'm sure these are things that you've thought about and thus I'd be interested in your perspective.