0

I have a series of 700 possibles variables to explain (Y), all of them are continuous variables. Those 700 variables are lagged and new lagged variables for the last 12 months are calculated, so from the initial 700 we end up with 8200 possible variables to explain (Y) process. There are 110 obs available, and only 98 are used after elimnating N/A data.

To perform adaptive lasso I first need to calculate the Best Beta coeficients from the Ridge Regression, and used them as extra penalty on Lasso applied after the Ridge. This process can be done through glmnet package.

But , Glmnet CV function seems not to be the right one for time series, as some of the information involved in lag variables maybe broken once applied CV on some of the Folds. It might be more appropiate to conduct CV on a rolling window and in the in-sample data or leaving one out.

I have found a few packages from R that perform some of those requirements, but the ones I have found so far fail on the CV process needed for time series.

This are some of the drawbacks I have came acrros.

  • Glmnet, seems to use CV not adecuate for time series, as some of the lag information/properties are lost in CV folds.

  • HDeconometrics, seems to have the correct CV process, but I have failed to install it in R from github, and I have not seem to many using it.

    install_github("gabrielrvsc/HDeconometrics") FAILED

  • Msgps : seems to operate Adaptive Lasso directly but there is no option to control over CV, and I am not sure what type of CV is performing.

  • BigVar, seems to use fantastic CV process, but is more designed for VAR. It´s possible to use CV in a rolling-window strategy or Leave one out.

Is there any one, that knows other packages or possible solutions through the packages I have mention?

Thanks in advance for your help

Luis Carmona Martinez
  • 119
  • 12
  • Seems too broad for SO's coding focus as well as being an off-topic request for package rec, but perhaps would be a better fit for CrossValidated.com or the QuantFinance forum whose name I forget at the moment. – IRTFM Apr 12 '18 at 15:59
  • I agree with @42 but I would also suggest that you consider just using glmnet and code the cross validation yourself, then you'll have full control over it and know exactly what you are doing. – AidanGawronski Apr 12 '18 at 16:05

0 Answers0