Machine learning for inventory management: forecasting demand quantiles of perishable products with a neural network

Thumbnail Image
Journal Title
Journal ISSN
Volume Title
School of Business | Master's thesis
Degree programme
Information and Service Management (ISM)
Accurate demand forecasting is a crucial component in building an efficient supply chain. Forecasting is a major determinant of inventory cost. Several methods and models for forecasting have been studied extensively over the last decades. In recent years, there has been a growing interest in the capabilities of Machine Learning algorithms in forecasting, and specifically in Neural Network models. Despite the expanding research on forecasting with Neural Networks, there have been only few studies focusing on the specific ramifications for forecasting demand of perishable products at the Stock Keeping Unit (SKU) level. Forecasting SKU-level demand for perishable products is a challenging task: time series for demand are volatile, skewed, subject to external factors, and frequently consist of only a few observations. Furthermore, SKU-level demand forecasts are typically used for inventory management, which imposes additional requirements on the forecasting procedure. This study examines how to design Neural Networks that address the specific ramifications of inventory management for several thousand SKUs. This work identifies central issues in the field and compiles successful approaches to overcome them. Next, a Neural Network architecture is suggested that takes these special requirements into account, building on insights from the literature. Namely, it learns from multiple hundred time series, incorporates external data into the prediction, and provides quantile forecasts of cumulative demand. In a large-scale experiment, the model forecasted the demand for several hundred SKUs in the fresh product segment of a German wholesale company. These forecasts were subsequently used for simulating the inventory development at the company for three months under close-to-real-life conditions. This study shows that Neural Networks are a promising approach to deal with large-scale forecasting problems for perishable products. The main finding of this study is that within the experimental setting, the base form of the suggested model for accurate daily demand forecasting yielded superior results to an array of competing baselines. In terms of inventory performance, the results are mixed, but present exciting directions for further research.
Thesis advisor
Malo, Pekka
Kuosmanen, Timo
Fügener, Andreas
neural networks, machine learning, inventory management, demand forecasting, simulation, deep learning, supply chain management
Other note