Context
I am running an unobservedcomponents model from statsmodel api in python
The data is daily in nature , and these are the parameters I have set for my model:
I have a day of week pattern in the data which is why the seasonal term is set to 7.
Issue For some series , the forecast seems to start at a very high level (please see image below) . This is mostly due to the seasonal component.
The entire history that I am using to generate this forecast looks like this :
Sure, the actuals have some really spikey data that could be causing this seasonal component to be high. I tried to solve for this by removing the spikey data (treat as outliers) and running the model , but it didnt work. In order to solve for this or even just to understand why I get such results I would like to know the following : a) Is the seaonal component (in this case day of week pattern) indeed being affected by the spikey data? b) How is the seasonal component computed in UC model. Is it the average of the seasonal variation over the entire history ? c) Any ideas on how to tackle this?
