Early bird tickets for Inspire 2023 are now available! Discounted pricing closes on January 31st. Save your spot!

Alteryx Designer Discussions

Find answers, ask questions, and share expertise about Alteryx Designer and Intelligence Suite.

Arima Forecast vs TS Covariate Forecast different?

7 - Meteor

I am running an Arima model with covariates to forecast weekly values based on data going back to 2011.  When I look at the interactive portion stemming from the Arima tool, the fitted line is different from the TS Covariate Forecast fitted line for the same week(s).


For example, when hovering over the interactive portion of the Arima tool, I get a fitted value of 0.89 for the week of October 29, 2018


In the forecast tool, hovering over the same week (Oct. 29, 2018), I get a fitted value of 0.81.  What I am missing?  Below is the screen shot of my set up.  I will work on scrubbing the data and can attach my workflow but curious if anyone know why this would be the case?TS_Arima Forecast.PNG



Alteryx Alumni (Retired)

Hi @ctthornb123,


Thank you for your question! The difference you are seeing in estimates has to do with the covariate values each tool is using to forecast your target variable. When the ARIMA tool creates a time series model with a covariate, it does so taking into account how the values of your covariate relate to the values of your target variable. When this model is later used to forecast values, you need to provide covariate values for the dates you are forecasting your target variable for. The model assumes these dates happen in the future, following the trends established during the training of your model.


By default, the ARIMA tool will create a forecast plot for the interactive output, where the values of your target variable are estimated for 6 periods into the future (you can change the number of periods forecasted by the ARIMA tool in Other options tab of the tool's configuration). Because the tool does not have covariate values for these six future periods, you are asked to specify the values you would like the ARIMA tool to use for the covariate; either the Mean of the covariates or the Estimated change.




The Mean of covariates option means that the tool will take the average of all of the covariate values provided to the model, and use that average as the covariate value for all forecasted periods. The Estimated change option creates an estimate of future values of the covariate by taking the average difference (lag) between adjacent covariate value and adding that average to the most recent covariate value (e.g., where the last known covariate value was 0, and the average lag is 2: future period 1 = 2, future period 2 = 4, future period 3 = 6, etc.)


The values estimated by the ARIMA tool differ from the values estimated by the TS Covariate Forecast Tool because each tool is using different covariate values for forecasting. By providing the TS Covariate Forecast tool with the same covariate values you used to build the model, you are getting estimated values for future periods based on covariate values from the past.


If you selected the Mean of covariates option in the ARIMA tool, you can get the values created by the TS Covariate Forecast Tool to match the values estimated by the ARIMA tool by using a Summarize tool to average your covariate values, renaming the field to have the name of  your original covariate field, and using a Generate Rows tool to create as many rows as you want periods forecasted in to the future. 





You can also estimate the Estimated difference values with a handful of Alteryx tools (workflow for both is attached). 




Please note that as mentioned in the help documentation for the TS Covariate Forecast Tool it is common practice to actually create a model to forecast the values of your covariate into the future, and then use those estimated values to forecast your original target variable. The methods used in the ARIMA tool itself are quick and easy ways to get rough estimates of forecasted values, to help with general model assessment.


Does this explanation make sense? Are there any further questions on this matter I might be able to assist you with? Please let me know!


-Sydney and @HossC (co-author/time series guru)