Option to pass "smoothing_level" to ES model

[nitesh4it]

Is your feature request related to a current problem? Please describe.
Yes, we are using ES model to predict the value of the timeseries, now when the next point is being predicted, it is giving a very high weightage to the last point(s). So say if there was a drop in actual data for the last 2 points only, the predicted value goes down

Describe proposed solution
Exponential smoothing model provides the option to pass "smoothing_level" bounds which is not in Darts. If we get the option to pass this smoothing_level to ES via darts. It would probably resolve the issue. we can have a cap to restrict the weightage.

[dennisbader]

Hi @nitesh4it, and thanks for writing.

We support the fit_kwargs at model creation where you can add the smoothing_level. See the Darts ES docs here and the ES fit docs here.

[nitesh4it]

Thanks for the suggestion. On further checking i found that we can pass a value as smoothing_level in the kw_args and that same value is picked
But what i meant to ask was, when i init ES model i do have the option to pass a range to it and an optimised value for smoothing_level is picked from that range. I wanted to have that bounds option in darts as well. Is that already supported? Am i missing something here?

es = ExponentialSmoothing(
timeseries.values(),
seasonal=model.model_params['seasonal'].value,
trend=model.model_params['trend'].value,
seasonal_periods=model.model_params['seasonal_periods'],
freq=timeseries.freq,
dates=timeseries.time_index,
bounds={"smoothing_level": (0, 0.7)}
)

I have the option to pass bounds={"smoothing_level": (0, 0.7)} in it. and an optimal value is then picked from the range

[nitesh4it]

@dennisbader @madtoinou Could you please have a look again

[madtoinou]

Hi @nitesh4it,

Please avoid pinging us with such a short delay after the last message.

The bounds parameter of the statsmodel implementation of Exponential smoothing is not exposed by Darts, but it could easily be implemented.

I am reopening this issue and will open an PR to address it soon.