Future values for related time series dataset

[apurvaGodg]

While forecasting item sales using retail domain, the documentation link (https://docs.aws.amazon.com/forecast/latest/dg/retail-domain.html) mentions that we use –

• webpage_hits (float) – The number of web page hits received by the item at the timestamp. Applies only to ecommerce websites.
• stockout_days (float) – The number of days left before the item goes out of stock. This is an optional field. Provide it only if the data is available.
• inventory_onhand (float) – The number of items in inventory.
• revenue (float) – The total revenue generated by that item’s sales.

Using DeepAR+, it is recommended that we provide related time series data for the forecasting horizon period(training horizon+forecasting horizon), I can provide features like pricing, promotion etc. However, for forecasting horizon, I don’t have access to actual & correct figures of webpage hits, stockout_days, inventory_onhand. Moreover, revenue depends on number of items sold.

1) So, how do I impute those values for the forecasting horizon period?
2) Let’s say I use some imputation method; will it not skew my actual forecast results?

[shimomut]

@apurvagodghase
I understand there are optional data fields which are not under your full control (such as 'webpage_hits' and 'revenue'), and it is difficult to feed them in the related time series data. As they are optional fields, you can choose not including those data field in the related time series dataset, and it is better than including data with quality issue. Another idea is to use those fields for what-if analysis.

[apurvaGodg]

@shimomut : That's what I thought. Thanks for the clarification. Another thing I noticed is:

"A related time series dataset can contain up to 10 forecast dimensions (the same ones in your target time series dataset) and up to 13 related time-series features."

Can I expect that the limit of 13 related time-series features will be increased in the future? I didn't see any such limitation on dynamic features while using simply DeepAR

[shimomut]

@apurvagodghase
Thank you for your input regarding the limitations in number of columns. As we are continiously working on increasing limitations and we will include your feedback in our planning.

[robsannaa]

I have a related question: if I pass a related time series for training, where should I provide the future related time series to get prediction output?

[shimomut]

Hi @robertosannazzaro , the future RTS data has to be included in the RTS dataset, in the same dataset as the past RTS data.