Demand Forecasting with Recurrent Neural Networks
Director of the project:
Jens Burger, PhD
July 2018 - June 2019
In production and operations management demand forecasting is an important method as it helps to develop better approximations of future operations under the presence of uncertainty. Forecasting extracts mathematical relations from past data that can be used to inform future decision making. In supply chain management, efficient coordination of resource acquisition, production and warehousing strongly depends on accurately predicting future product demand in particular and market dynamics in general. Accurate demand forecasting therefore reduces investment risks in uncertain environments. The challenges of demand forecasting lie in the complexity of demand dynamics. We investigate the application of reservoir computing (RC) to product demand forecasting. RC utilizes a randomly initialized recurrent neural network that implements finite memory and generalization. Under these conditions it should be sufficient to reduce training complexity to only a single linear output layer and achieve accurate forecasting results. The output layer is therefore able to derive a simple linear relationship between the input data and its projection into a higher-dimensional feature space.