I am wondering if anyone can provide some clarity about feature selection with time series of panel data. Specifically I'm doing this with R.
Lets say I have a time series of financial panel data with a monthly frequency. Say K number of factors on 1000 stocks every month end and I have 20 years of this monthly data. I want to build say a random forest or xgboost model to predict Y using 7 years of history, so a 7 year panel. This data gets updated at the end of every month. If I want to build a model for backtesting, I'd have models for every month, which lookback 7 years. How do I go about feature selection?
Say I use a Boruta algorithm for example. Do I just use the entire 20 year history to select the factors, and build a new model every month using those?
Since their is both a cross-sectional and a temporal component to this, I'm unsure of the best course of action.