About creating target time series data when obtaining weekly forecasts with Forecast

Amazon Forecast allows you to specify the forecast frequency when creating a predictor. This determines the frequency of the predictions in the forecasts. For weekly forecasts, you can set the forecast frequency to "Weekly".

Your target time series data can be <= your forecast frequency. Amazon Forecast has the ability to aggregate the data up to your specified forecast frequency. When aggregating data to the weekly forecast frequency, Amazon Forecast will align the data based on the default time boundary, which is the most recent Monday. You can also specify a custom time boundary if needed, such as the start of the week on a different day.

For the target time series data, you can create timestamps on a weekly basis, aligning with the forecast frequency and time boundary you choose. If you have missing data for certain weeks when a product was not available, you can indicate this in the data, rather than using zeros. This can help the model better understand the underlying patterns.

Once your predictor is ready, you can use the Amazon Forecast SDK or CLI to generate weekly forecasts. You can specify the forecast types (e.g., p50, p70, p90 quantiles) and any additional filtering criteria when creating the forecast. The forecasts can then be exported to an S3 bucket for further analysis.