preprocessing

[m-muecke]

Check the needed preprocessing, date features, box-cox, log, transformations etc.

[m-muecke]

try to implement pipe op for creating lags as an alternative to creating lags inside forecast learner. Then lags only have to be created once in benchmark.

lag transformation pipe op could also be then run sequentially instead of passed as arg

[m-muecke]

Now there're two forecast learner wrapper versions implemented.

1. Builds the lags inside the wrapper prior to training and stores the needed data of the tasks or the lags. For lag transformations this would require an arg like lag_trafo to provide a list of trafos.
2. The wrapper doesn't create any lags and requires a lag pipe op to be run prior. The pipe op stores the relevant data it needs. Note: here the order of the graph is essential.

Note: Both implementations only work for recursive prediction. Direct prediction requirements would have to be checked.