What Can We Learn from LLMs? Building a Foundation Model for Inventory Management

Abstract

Motivated by challenges of AI adoption and inspired by recent developments in CV and NLP, this paper proposes a Foundation Model for inventory management based on the GPT architecture. Built on a unified model architecture and a streamlined and standardized training process, our model exhibits exceptional versatility and scalability. It can deal with thousands of heterogeneous products, exploit crosslearning opportunities in a large product portfolio, and has zero-shot learning capabilities that make it suitable for managing new products for which only very limited historical data is available. Based on a real-world retail dataset we can show that these benefits can be achieved without sacrificing performance: Our Foundation Model consistently outperforms various state-of-the-art models in terms of inventory distortion costs. Our contributions are relevant for the Information Systems community, highlighting how AI-based decision-making can be effectively implemented in complex business environments.

Keywords: Artificial Intelligence, Foundation Models, Prescriptive Analytics, Inventory Management