Time Series Forecasting

[sakaria97]

Hi,

I'm trying to forecast product sales for each of 66 products using two years worth of historic data. I have grouped the days up by weeks, and my weeks format is 1, 2, 3 , (a single number). The key thing is that , our working calendar starts on Friday, so this poses problems for any forecast. 

I created my own week lookup and would like the ARIMA model to lookup my week column for the forecast.

When selecting ARIMA model, i must specify, my column for sales and a period to group by, I want to select week, but it wont work.

I've posted a snapshot of my data, in excel format, i have a unique identifier for product, but i haven't included that in the file.

Any help will be greatly appreciated.

Thanks.

[BenMoss]

Can you share your workflow too, that would be helpful.

Also what exactly you mean by 'it won't work'

Just a knowledge dump for you in regards to the forecasting tools.

It will produce a forecast along the entire dataset, if you want to do a forecast for subsets of this data, such as 'Products' then you should look at the TS Factory tools that exist in the alteryx gallery (https://gallery.alteryx.com/#!districts/56327e37aa690a17f0760bdc) that allow you to achieve this.

Your data should be aggregated at the correct level and filled with any missing weeks too. So you end up with one row per week per product.

Ben

[BenMoss]

If you mean it freezes, thats probably because you havn't aggregated your data to the correct level, so it's doing a forecast thinking it has 38,818 weeks of data in, which isn't the case and this is making alteryx go a bit crazy!

So make sure you use a summerize tool, and as mentioned if you then want a forecast per group then you must use the TS Factory tools on the gallery.

Ben

[sakaria97]

Hi Ben,

Please see workflow below for my time series forecast.

[BenMoss · Accepted solution]

Yes, it looks like the error is caused for the reasons outlined above.

With the standard tools you can only do a single forecast where the data should be aggregated at the level required for the forecast.

I.e. if you are doing a weekly forecast and have 2 years of historical data then you should have 104 rows.

If you want to do it via a group then please use the TS factory tools as per the link in my first post.

Ben