Demand forecasting with AI using limited historical data

The importance of Artificial Intelligence (AI) is increasing rapidly in the business world. In current business environment, the volatility is accelerating even further. By using AI to better predict future sales, Belgian companies can improve their production and capacity planning, but also optimise purchasing policy and inventory management. However, AI usually needs a lot of data to make predictions. While this is not a problem for large companies such as Amazon, Bol.com or Coolblue, this often forms — in addition to model knowledge and IT infrastructure — a major barrier for Belgian companies to get started with AI. With the ever-increasing globalisation, companies are fighting with unequal weapons against the competition from multinationals. This research project wants to apply AI techniques to make good demand forecasts, even when there is little data. We developed a framework that can exploit the hidden structures in the data in combination with data-sparse models and applied it to various business cases. We investigate how we can use data-intensive Deep Learning models within this existing framework. These have a better predictive power and more accurate predictions, which is given the major impact on costs. However, these models normally require (even) more data, more computing power and are less transparent. To ensure a wide adaptation of our framework, we use an open science strategy via open access papers, open-source software packages, incorporation in course material and popular science articles. Our goal is to bring research results closer to academia, professionals, students and the general audience.