Implement forecast decomposition for Holt-like models

[alex-hse-repository]

🚀 Feature Request

Implement forecast decomposition for models from holt_winters.py

Proposal

In _HoltWintersAdapter implement method

def predict_components(df: pd.DataFrame) -> pd.DataFrame
def forecast_components(df: pd.DataFrame) -> pd.DataFrame

Details:

• implement in-sample decomposition in predict_components
• implement out-of-sample decomposition in forecast_components
• use model fit results to obtain level, trend and seasonal components for training set
• for forecast use componetns from train to estimate corresponding decomposition
• use target_component_ as prefix for component names
• raise error for multiplicative models
• if use_boxcox for HoltWintersModel is on, do the inverse transformation for components like in inverse_transform method of our Transforms

Test cases

For models from holt_winters.py

• Test components sum up to target for different parameter sets
• Test damped trend for HoltWintersModel and HoltModel
• Test that the component names are correct for all possible parameters combination
• Test components transformed along with target when use_boxcox set in HoltWintersModel
• Test that model decomposition contains only additive components and raises corresponding error otherwise

Additional context

Models reference

See implementation in ETNA-1480 folder