Create a train/test split at the data prep level. I borrowed a utility function from the AutoML many-models notebook to do the splitting and datastore upload
All transforms are in sklearn pipeline steps to demonstrate how to consistently featurize train and test sets
Model is trained on the full training data (previously, the training step made a train split and only trained on that part - it's best-practice in forecasting to always re-train on the full history before making forecasts)
Added some links to AutoML notebooks and docs for more advanced scenarios
Fixed bugs in the training script (references to static string, 'Quantity') and the copy predictions script (incorrect column order)
The main changes are: