How to set up the TS Covariate Forecast Tool

[IraWatt]

Hi all,

I'm wanting to tweak the example TS Covariate workflow along the lines of what Susan outlined in Back to the Future: ARIMA and Forecasting with Cov... - Alteryx Community. I'm unsure how to correctly feed in future predictions for the covariates to aid in the forecast accuracy. My first question is by default the example workflow has 'bookings' as a covariate, does that make sense as that's already what I'm already aiming to forecast? feeding it future values of bookings seems wrong. My second issue is how to feed in future values for the snow '(inches)' feature to aid in the forecast?

my current idea of what the workflow should look like:

Thanks for any help on the topic.

[danilang · Accepted solution]

Hi @IraWatt 

A straight TS model, like the Learn One Tool at a Time example for ETS_ARIMA_etc, determines the relationship between a single Target variable, Bookings and the time period.  By including covariates, like snowfall, the ARIMA tool builds a model that includes a component that shows how the Target variable might be affected by the other variables.    

To get the sample in line with @SusanCS' article, specifically this part

"The forecasted snow time series would need to have the same field name as our historical snow time series used as a covariate to build the original ARIMA model. It also needs to be ordered such that the soonest forecast period is listed first in the series, with the most distant forecast period at the end of the series."

you need to the take the values that come out of the False output of your Filter tool and remove the Bookings field leaving only the RecordID, Year, Month and Snowfall fields.  Feed that into L input of the TS Covariate Forecast tool.   

With this config,  you're asking for a forecast for the next x periods and also including a forecast of what the covariate variable is likely to be.  You don't include the future values of the Bookings field because that's what your trying to forecast

Dan

[IraWatt]

Thanks @danilang, from that I gather the example covariate workflow is wrong in adding bookings as a covariate?

[danilang]

Hi @IraWatt 

Technically it's wrong, but it looks like the tool doesn't actually include the target variable as a covariate if it's checked.  You can verify this by running the sample twice, with bookings checked and not and comparing the O output of the ARIMA tool.  Pass it through a Blob Convert to convert it to a string with UTF8 as the code page.  The object models are identical even if Bookings selected.

Dan