- we can do transfer learning in forecasting.
- if you don't have future values you can transform your original features into something else, which you could then more easily set (instead of using the direct value). For example, instead of using sunlight directly, you could use the co-variate "absolute/relative change to yesterday".
- another approach is to use mulit-variate forecasting techniques where you forecast everything jointly
- We align timestamps. The idea is that all timestamps within the same range, are represented by the same value. In this case 2019-7-2 is aligned to 2019-06-27.
- I think we should remove make_evaluation_predictions entirely for a better more understandable solution.