GOR AG online webinar on "Forecasting for Inventory Management at Wholesalers"

The German OR Society working group in Forecasting is launching a series of webinars leading up to our in-person meeting next year in Munich.

Our 2nd speaker is Claudia Ehrig, PhD candidate at the Christian-Albrechts University Kiel, and Research Associate in the Data Science Group of the Analytics Department for Supply Chain Services at Fraunhofer IIS. Claudia will talk about his research topic as outlined below.

The webinar is open to members of the GOR and all other interrested parties, no membership required!

Please RSVP in the Meetup page!

Talk 1 - Probabilistic Cluster Forecasting Models for Inventory Management: A Case Study on Wholesalers
Since major supply chain disruptions, such as the COVID-19 pandemic and the war in Eastern Europe, have caused supply shortages, wholesalers have built up massive stock levels to prepare for future uncertainties. This strategy becomes expensive, however, due to high inflation rates. Therefore, accurate demand forecasting which considers uncertainties and desired service levels is crucial for effective inventory management. In our case study involving two wholesalers from Germany and Austria, we cluster articles' demand time series according to their intermittency and variance. We then train a probabilistic neural network per cluster. By doing so, information is shared among similar time series through the model parameters, reducing the need to train and maintain individual models for each time series. The resulting forecasts provide a distribution of likely values in the form of sample paths, allowing for a more comprehensive understanding of the joint distribution during the forecast period. The quantiles of the cumulated sums of these paths correspond to the targeted inventory service levels and can be directly incorporated into inventory control optimization.
In a benchmark with the commonly used forecasting model Exponential Smoothing, our approach achieves a 25% and 32% lower MASE on 5000 products from each wholesaler during an unseen test period, respectively. An ablation study further examines the impact of providing weekly forecasts instead of monthly as well as including external regressors such as seasonalities, product groups, hierarchies, and pricing information. While our approach demonstrates improved forecast accuracy based on historical data, further insights will be gained through real-time evaluation at the two wholesalers.
This study is part of the project DOMNet, which also encompasses a reinforcement learning based inventory optimization strategy and a modular software architecture for the deployment.
Authors: Claudia Ehrig
Co-Authors: Nico Beck, Henning Frechen, Georg Schett