Revenue forecasting for large business organizations is a challenging but important problem. As a multinational business organization, Bosch has an estimated 2,000,000+ time series capturing monthly financial key figures at multiple organizational and product hierarchies, which are forecasted every month into the future 12 month horizon to inform financial and resource planning. To address this challenge, Bosch has developed an in-house forecasting solution serving 20+ forecasting models, while a meta-model is used for selecting a subset of these models best suitable for each individual time series. The framework is designed to be flexible and adaptive to support continuous introduction of new models, from both an algorithmic and architectural point of view.
We show how we can utilize the flexible development environment on AWS to explore a set of neural forecasting models and demonstrate a path for integration into the existing solution at Bosch via REST API. While doing so we place special attention on addressing the challenges brought forth by unexpected global events such as COVID-19. We investigate recent deep neural network (DNN)-based forecasters, which shows promising results for many forecasting problems. More specifically, we include the offthe-shelf Recurrent Neural Networks (RNN) and Convolutional Neural Networks (CNN) models (DeepAR+ and CNNQR) available from Amazon Forecast, as well as a custom-built Transformer model.
The Transformer model incorporates a special module with attention adjustment to handle out-of-distribution COVID-19 period data. Backtest results from the Amazon Forecast models and the Transformer model are used to obtain ensemble forecasts, which lead to robust forecasts over time. The ensemble model serves the forecasting results to Bosch’s in-house product from a forecasting pipeline implemented in the cloud in a modularized manner via REST API, which is deployed and currently in production.
Research areas
