Implement forecast decomposition for TBATS

[alex-hse-repository]

🚀 Feature Request

Implement forecast decomposition for TBATSModel and BATSModel

Proposal

1. Create class _BATSBaseAdapter(BaseAdapter)

• Move all the methods form _TBATSAdapter there
• Create abstract_methods:
  • def forecast_components(ts: TSDataset) -> TSDataset
  • def predict_components(ts: TSDataset) -> TSDataset

2. Create _BATSAdapter(_BATSBaseAdapter), _TBATSAdapter(_BATSBaseAdapter)

• Implement abstract method forecast_components
• Fix parents for TBATSModel and BATSModel

3. Implementation details:

• Component names: local_level, trend, seasonal_(s={s}), arma(p={p},q={q})
• Components depends on: always exist, use_trend, seasonal_periods, use_arma_errors
• If use_box_cox is on, do the inverse transformation for components like in inverse_transform method of our Transforms

Test cases

• Test components sum up to target
• Test that depending on parameters we obtain the expected set of components(use_trend, seasonal_periods, use_arma_errors)

Additional context

• See the draft implementation in ETNA-1520
• See implementation of forecast here