Exponential Smoothing & Seasonality I am relatively new to stats and forecasting and was hoping to tap the wisdom of this community in regards to a question.
One of my colleagues insists that exponential smoothing presumes seasonality.  I have not seen that anywhere in my investigations of the topic.  To be clear, I'm looking for an actual document that states clearly "exponential smoothing assumes seasonality".  I've seen reference to exponential smoothing with seasonality but that implies that there is exponential smoothing without seasonality.
So my basic question is this: 


*

*Do all forms of exponential smoothing (simple, double and triple) assume that the underlying data are seasonal?  


If the answer to the above is yes, then:


*

*what happens when those techniques are used on non-seasonal data?

*How strong a pattern of seasonality is required?


If the answer to the above is no, then:


*

*what happens when those techniques are used on seasonal data?


Several applications (such as Tableau) seem to use exponential smoothing by default when forecasting time series data.  This led me to believe that seasonality cannot be a requirement of the technique, or that the requirement for seasonality is so small that it does not materially impact the forecast when there is no seasonality present.  Otherwise, why use that technique as a default?
Any assistance this community can provide is most appreciated.  Thank you.
 A: 
Do all forms of exponential smoothing (simple, double and triple) assume that the underlying data are seasonal?

No. Only triple exponential smoothing does. 

what happens when those techniques are used on non-seasonal data?

Ideally, since triple exponential smoothing decomposes the series into 3 parts, level, trend and seasonality, it should simply choose parameters for the seasonal component such that the model is equivalent to a non-seasonal model. In practice, this will depend on the optimization method and software package being used. 

How strong a pattern of seasonality is required?

Again depends on the optimizer and software being used. In the most basic versions of triple exponential smoothing, you need to specify beforehand what the seasonality is (12 months? 52 weeks? 7 days? etc...) and then it will try to fit the best seasonal model it can find for the specified seasonality. 
If the seasonal pattern is not strong, then  ideally the seasonal component in the model will have a smaller contribution to the overall model. However, this also depends on the software and the optimization method used. 

Several applications (such as Tableau) seem to use exponential smoothing by default when forecasting time series data.  

I don't know specifically about Tableau, but most automated forecasting tools will try multiple models, both seasonal and non-seasonal, and then return the model which best fits the data. So it is not that seasonality is not a requirement of the technique, but more that the process of choosing seasonal or non-seasonal has been automated. 
