Time series forecasting clarification

pmv

The documentation on time series algorithm states that statistical algorithms will train separately for each time series in input file, whereas the deep learning models will train on all time series in input simultaneously (but produce a separate forecast for each).

• What might be the rationale/underlying assumptions for why the deep learning models are trained on all time series in input file?
• Does this mean that, if one hypothesizes different DL model parameters for each time series then, to compare model performance vs. statistical algorithms, one should perform model selection on each time series separately?

Thanks in advance for the help.