I currently have a model which is trained on all possible input features that seem to be good predictors. The issue however, is that these features are available throughout time, at different stages. So my idea is to perform a prediction,
n times, for each of these stages in an incremental way, that is, to accumulate the features already available up to a given time
t-1, and include the newly available ones at time
t and return a prediction, which will hence be progressively refined.
So my thoughts at this point are:
Are there any approaches to deal with such scenario with with a single trained model? So to have only a ML model (currently a XGBClassifier), and to have it return a prediction, even though
X_testcontains only a subset of the columns in
X_train? My thoughts on this is that it is not possible. But maybe I'm missing something, such as using feature weights on the test sample, to ignore those features? Not sure.
The alternative I've thought of, is to build n models, each of which includes the the features available up to a given time
t. Resulting in my case in about 6 or 7 pre-trained models.
Any thoughts/advice on this would be much appreciated.