Add `params_to_tune` for exponential smoothing models

[Mr-Geekman]

🚀 Feature Request

Add params_to_tune for exponential smoothing models.

Proposal

Sources:

• pycaret
  • exponential smoothing
  • ets
• kats

Suggested grid for HoltWintersModel:

{
    "trend": CategoricalDistribution(["add", "mul", None]),
    "damped_trend": CategoricalDistribution([False, True]),
    "seasonal": CategoricalDistribution(["add", "mul", None]),
    "use_boxcox": CategoricalDistribution([False, True]),
}

Suggested grid for HoltModel:

{
    "exponential": CategoricalDistribution([False, True]),
    "damped_trend": CategoricalDistribution([False, True]),
    "use_boxcox": CategoricalDistribution([False, True]),
}

Suggested grid for SimpleExpSmoothingModel:

{
    "use_boxcox": CategoricalDistribution([False, True]),
}

To discuss:

• Can we do something with seasonal_periods parameter in HoltWintersModel? The problem is that it depends on the data. For example, we can just keep the value that was put by the user and don't change it.
• What should we do about use_boxcox=True if there are negative values in data? In this case the model will fail probably.
• What should we do about trend="mul" if there are negative values in data?
• Should we touch smoothing levels during tuning?

Test cases

Look at https://github.com/tinkoff-ai/etna/issues/1184.

Additional context

No response

[martins0n]

• seasonal_periods - we can skip this
• In case of negative values I guess user could add transformers or change default grid. So it's ok to leave it this way.
• trend="mul" - Do we really have any problems here?
• Should we touch smoothing levels during tuning? - It's turning by default as I could se

[Mr-Geekman]

Closed by #1209.