I have an observation X every hour => Time serie 1. I want to predict an event Y such as Y =f(X) => Time serie 2 : Y_pred. For the past periods I have both X and Y (Y_true) For future periods I have X
I have tried different models and found so far that gradient boosted trees perform the best. However I can notice that for never seen volumes of observation X, the results plateau. This is as classic issue of extreme event prediction.
I saw the paper from Uber Engineering on extreme event prediciton: https://eng.uber.com/neural-networks/
From what I understand, LSTM typical usage is to predict the "future features", but in my case I have these features (X) and only need to find the right scaling factor for each time unit between X and Y.
What LSTM approach would you suggest in my case?