Community Spring Cleaning week is here! Join your fellow Maveryx in digging through your old posts and marking comments on them as solved. Learn more here!

Alteryx Designer Desktop Discussions

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

Time Series Analysis Not Predicting Subsequent Periods

frank6773
8 - Asteroid

Hello,

 

Question on time series analysis. I cannot figure out why the attached workflows using both ARIMA and ETA are not providing me the subsequent time periods for forecasts.  The input is through week 25 of this year, yet the predictions start at week 31.  I would like them to start at week 26 and cannot figure out what I'm doing wrong.  I know there is some missing periods and not all are accounted for but would this cause the model to predict further out in the future rather than the closest period, not sure.

7 REPLIES 7
Sntrada
11 - Bolide

Hi @frank6773 ,

 

Are you able to provide this as a packaged workflow file instead? It's appearing as an app file, and I'm getting errors when I try to open it. 

frank6773
8 - Asteroid

Sorry about that, I guess that would work better.  I have not attached it as a package. Thanks in advance for taking a look!

Sntrada
11 - Bolide

Hi @frank6773 ,

 

What I deduced is that you need to set the target field frequency to "other", in the tool setup node, screenshot attached. The weekly option is not working well because you don't have a complete date in your dataset. After making this switch you will get a forecast from record ID (period) 187, for eight periods. I attached the output of this as well. This output now includes predictions from week 26, AKA record ID 187. 

 

I'm a bit surprised by the forecast from the ETS model. The 8 forecasts all have the same value which is highly unusual. I'm looking into this, and I'll post if I can find out why. 

 

 

 

Sntrada
11 - Bolide

FYI, the lack of forecast from week 26, when using the weekly target field frequency in the ETS model, is not due to a lack of a date. I computed the date and reran the model using the weekly frequency selection, we are still having that issue with the forecast starting from 31. I attached the new data file. 

 

This is weird. Still looking into why we have the same forecast for the ETS. 

 

frank6773
8 - Asteroid

Thanks for looking into it, I really appreciate it. As I am not as familiar with ETS or time series general (and am learning as I go), I was not even aware that it is strange to see the same value for the prediction as I had been getting that consistently with other datasets I was modeling as well.  I cannot figure out why I wouldn't get any prediction for the current upcoming weeks either, could it be just due to not having a data point for all weeks?  It seems odd it could start predicting at week 31 though.

Sntrada
11 - Bolide

Hi @frank6773 , 

 

Yeah, it only predicts from week 26 if you select that "other" frequency, in the tool configuration. But it is weird that both Arima and ETS are starting from week 31 if we do a weekly frequency. At least with Arima, the predictions seem plausible. 

 

Missing data could throw things off, but I don't think it should cause the same prediction. I tried using different subsets of the data, and I still got the same prediction using ETS. I'm thinking it has something to do with seasonality, but I am still exploring this.  

Sntrada
11 - Bolide

Hi @frank6773

 

The ETS model isn't predicting varied values because it is not detecting seasonality or trend in the data. This is why it is giving that constant value, which is pretty close to the median of the data. ETS is able to predict a range for your data in the confidence interval, but that's about it! 

 

Since there is no trend or seasonality, ETS isn't the way to go for this data type. But it still good practice to build two models and go with the better one, i.e the one with lower error figures, and the one that actually predicts!

 

To explore if your data has trend or seasonality you can run the TS Plot tool. This tool works best when you have actual dates though. 

 

If you look at the ETS tool report, at the top of the first graph, you can see that the ETS model which R decided was the best fit is the one with no trend or seasonality, ETS (M,N,N). M stands for multiplicative error, N stands for no trend, the second N stands for no seasonality. I confirmed this by looking at the TS Plot decomposition graphs, specifically the seasonality, and trend graphs (after I generated dates for the data). 

 

Let me know if you have any questions!  

 

Labels