Forward-looking and Backward-looking

Hi class, I found I didn’t really get the concept of forward-looking and backward-looking, and here I have some summary from our class materials. Please point out if anything is incorrect, and I’m glad to hear any understanding on this.

For Forward-looking methods:

  1. entire dataset is used (searched) and models look into both future and historic values; (I read an article saying that forward-looking only uses data that should be unchanged from now on, but I wonder what is some data that would change through time and can be used in modeling and training? Is this saying right? Also I wonder what is future value in this case?)
  2. fitting transformation/interacting/mapping on training data and then only apply on testing data;
  3. perform on rolling basis and extracting the appropriate values;

For Backward-looking methods:

  1. recursive and only takes present and historical values into account;
  2. directly fit on all the data and we don’t have to distinguish between train and test data;
  3. recalculated at every time step to form a time series of data points; (here I didn’t get the difference between this point with rolling bases for forward-looking)

Thanks for any explanation!

I think for feature generation you should more think of it like this: forward-looking uses information from the future, so it can lead to data leakage, as such you need to create the features on a rolling basis or you need to fit on the training dataset and transform on the test dataset (both of these methods would ensure that you don’t leak information from the future into your training dataset).