Issue with output from TS Forecast Tool

Question

Hi,

I have been trying to do a forecast of daily customer collection using daily collection data for the past 365 days. I have built a workflow which is attached.

After running the workflow I had a couple of queries which I have not been able to resolve:

* Actual data as per TS forecast tool does not match with output data as per Input file: When I run my workflow and view the output in the interactive anchor of the TS forecast tool, the actual data in the charts do not match with the actual data in the historical daily collection file. Any idea why is this happening? How can we trust the forecast if the actual data itself is not matching with the historical data file?
* Forecast start period: My historical collection data is till 1 May 2019 and I configured the ETS and TS Forecast tool to generated a forecast for the next 30 periods i.e. T+30. So the forecast should be till 31 May 2019. However, when I opened the output anchor of the TS forecast tool I can see that it is showing actual data for 24 May! Additionally, the forecast starts from 8 June? Why is this happening?

Am I Missing something

Collection_31March2019.PNG

Actual data screenshot.PNG

Forecast start date.PNG

Screenshot of historical collection data_1.PNG

Screenshot of receipt forecast workflow.PNG

rsomani005 · Accepted Answer

I think there is a bug in the ETS and TS tool. I consulted some people within the team who can read the code behind the ETS and TS Forecast tool and this is what they could figure out:

* The issue is with the assumptions made by the TS and ETS tools. Since you can't specify where your data starts from or what your date field is, it assumes that your data starts Jan 1. 
* The forecast generated only has "periods" instead of dates. Alteryx makes some sort of assumption as to the dates that it inserts. I tried following some of the code in the ETS and TS functions but would need more time to figure it out. 
* Another issue that I noticed is that, if you check the interactive plot on the ETS macro, the decomposition plots start at Jan 1, but the forecast plot starts at Jan 8.

So, to summarize, don't trust those dates... perhaps if you are OK with eliminating some of the older data and having your model start on Jan 1, you may be able to get slightly more accurate dates on the charts, but otherwise, you can report this as a bug to Alteryx.

Better use the ETS and TS Tools with a pinch of salt. Currently, I got better results when I started my data with 1 Jan but still not a full blown solution.

AndrewKramer · Accepted Answer

As @rsomani005 said, the issue here is that, for daily data, there is no way to specify the start date of your time series. If you were to aggregate this data into a weekly format, you could fill out the "Series starting period" in the "Other Options" tab. This would render your graph properly. Unfortunately, this is not currently supported for daily data.

Thanks,

Andrew