An entity embeddings deep learning approach for demand forecast of highly differentiated products

The paper deals with Deep Learning architectures applied to demand forecasting in a complex environment. The focus is on a famous Italian Fashion Company, which periodically performs a sales campaign, to presents its new products' line and to collect customers' orders. Although production follows an MTO strategy, fabrics must be purchased in advance and a forecasting system is required to predict the total quantity sold for each product, at the early stages of the campaign. Due to high product variability, the forecasting system must consider products' similarities and the evolution of customers taste. Additionally, customer and product data are mostly described by categorical variables (hard to reconcile with a predictive task) and, unfortunately, time-series techniques cannot be used because of a sparse dataset. Given these criticalities, we propose an end-to-end approach based on Deep Neural Networks and on Entity Embeddings. A first neural network is trained to predict the total quantity of a given product ordered by a specific customer. Different Embeddings are learned for each customer and product categorical attribute. This gives the network the ability to effectively learn the complex and evolving relationships between products characteristics and customers taste. Next, freezing the learned product's embeddings, a second Recurrent Neural Network is trained to predict the total amount ordered for a given product, incorporating real-time data of customers' orders of the ongoing sales campaign. Ten years of sales have been analyzed and the approach, tested on unseen sales campaigns, has outperformed the forecasting algorithm currently adopted by the fashion firm.