- Title: Timing intermittent demands with time-varying order-up-to levels
- Abstract: Intermittent demands are notoriously difficult to handle in inventory control, because the times between demands and the demand sizes must both be modelled. Parametric methods based on simple exponential smoothing are still the standard forecasting options today, but various case studies already showed that these cannot capture all patterns in demand intervals. The M5 forecasting competition confirmed that intermittent demand patterns and their forecasting needs differ wildy per product. However, inventory models and the demand distributions that they are based on are not suited to match with these patterns. Whereas the normal approximation ignores demand intermittency altogether, (compound) Poisson or binomial models all assume that the probability of a demand occurrence is the same in every period. This assumption is theoretically convenient as it leads to stationary order-up-to levels, but it disregards more tailor-made stocking opportunities in practice. We propose an intermittent demand inventory model where the probability of a demand occurrence may vary with the time since the last demand. As a result, optimal stock levels also dynamically adjust to the demand pattern. We show the computation of the non-stockout probability, volume and order fill rate, by tracking possible evolutions of the position in the demand interval. We propose a greedy heuristic to optimize the time-varying order-up-to levels given a service level requirement, and show that on the M5 competition data this approach increases on-target service level accuracy by up to 5 times compared to the Poisson-geometric model, and up to 7 times for the normal model.
- Bio: Dennis Prak obtained his PhD from the University of Groningen in 2019, after which he joined TU Munich on an NWO Rubicon fellowship. Since 1 October 2020, he combines this stay with an assistant professorship at the IEBIS department of the UT, and from December 2021 onward he will fully join the UT. His research interests are in inventory control, demand forecasting and their interface, and in his current position he a.o. works on data-driven optimization methods for shared mobility concepts as team leader of a collaborative project between TU Munich and DTU Copenhagen. He is privately very interested in aviation and will therefore take over some tasks from Hans Heerkens at the UT, namely the Minor Aerospace Management and the chairmanship of the Platform for Unmanned Cargo Aircraft.
Robert van Steenbergen
- Title: Forecasting demand profiles of new products
- Abstract: In this presentation, we present a novel demand forecasting method called DemandForest, which combines K-means, Random Forest, and Quantile Regression Forest. This machine learning-based approach combines the historical sales data of previously introduced products and product characteristics of existing and new products to make prelaunch forecasts and support inventory management decisions for new products. DemandForest clusters and predicts demand profiles, and predicts quantiles of the total demand during an introduction period. We illustrate and validate our approach using real-world data sets of several companies. Compared to two benchmark methods, DemandForest provides the most accurate predictions, resulting in potential inventory savings of around 15% depending on lead times and service levels.
- Bio: Robert van Steenbergen is a PhD candidate at IEBIS and started in December 2019 on the project Last Mile Drone Logistics for Humanitarian Aid (AIRLIFT). AIRLIFT is a BMS Signature PhD project, a collaboration between the BMS faculty and the Dutch foundation Wings For Aid. The aim is to investigate the possibilities for deploying drones in humanitarian logistics and develop methods to coordinate them. The presentation will be held about the recently published article by Robert and Martijn Mes in Decision Support Systems, which followed from Robert’s graduation assignment at Slimstock last year.