Amazon Forecast now supports new automated data imputation options for the related and target time series datasets

Posted on: May 14, 2020

Amazon Forecast is a managed service that uses machine learning (ML) to generate accurate forecasts, without requiring any prior ML experience. Amazon Forecast is applicable in a wide variety of use cases, including product demand forecasting, inventory planning, workforce and resource planning, energy demand forecasting, and cloud infrastructure usage forecasting.  

Today, we are excited to announce support for automated filling of missing values in your related time series dataset for both the historical and forecast time periods. In Amazon Forecast, related time series includes data such as promotions, prices, or weather, that correlates with the target value (e.g. product demand) and can often improve the accuracy of the forecast. Until now, Amazon Forecast customers were expected to provide related time series data with no missing values, which can be challenging at times (e.g. providing price data for products for the entire historical and forecast time periods). With this new feature, customers can now use several missing value options (such as value, median, min, max and mean), depending on the specific use case, for their related time series dataset. Additionally, we are also expanding support for existing missing value filling options (beyond ‘0’ and ‘NaN’) for the target time series dataset. You can leverage the FeaturizationConfig in the CreatePredictor API to use this new feature. Please visit the Amazon Forecast developer documentation for more detailed information here.

This expanded missing value functionality is now available in US East (N. Virginia, Ohio), US West (Oregon), Europe (Ireland, Frankfurt),and Asia Pacific (Tokyo, Singapore, Seoul, Sydney, Mumbai).