Implement min_train_series_length() method for the FourTheta model

[rijkvandermeulen]

Is your feature request related to a current problem? Please describe.
The FourTheta model extends the ForecastingModel class which has a class property defining the minimum required length for the training series (min_train_series_length). This property defautlts to 3 but can be overriden for each child class inherenting from ForecastingModel . Now, I'm missing the method that overrides this min_train_series_length in the case of the FourTheta model. I'm pretty sure the Theta model requires at least 2 * seasonality_period in case seasonality is used (i.e. season_mode != SeasonalityMode.NONE).

When you currently try to fit a seasonal FourTheta model when you have < 2*seasonality_period observations, it errors out when it tries to do a seasonal_decomposition in extract_trend_and_seasonality(). Would it be an idea to already capture this in an earlier stage by implementing the min_train_series_length() method

Describe proposed solution
Implement min_train_series_length() method for the FourTheta model overwriting the default of 3. Something like:

if self.season_mode.value: return self.seasonality_period * 2 else: return 3
Might be a bit more nuance to the abovementioned solution since seasonality_period is an optional param and can also be retrieved from the time series. Hence, if seasonality_period=None we need to first retrieve the seasonality_period.

Describe potential alternatives

Additional context

[hrzn]

Thanks, that makes sense. Would you be willing to open a PR with this change?

[rijkvandermeulen]

Sure, let me give that a try!