Retail Demand Forecasting Using Temporal Fusion Transformer

Demand forecasting is a vital problem that affects nearly every aspect of supply chain management operations in the retail industry. A retailer would need to precisely anticipate the upcoming demand to efficiently manage their inventory, form profitable pricing strategies, and handle logistics operations in time. Novel techniques in machine learning and deep learning literature have also been important in demand forecasting domain as they thrive at extracting complex relationships from data. In this study, a multivariate hierarchical time series forecasting problem is tackled for a leading retailer that operates with hundreds of stores in Turkey. A dataset consisting of 50 items for e-commerce channel sales starting from 2016 is used. A well-known transformer-based Deep Learning algorithm, Temporal Fusion Transformer has been compared against numerous state-of-the-art architectures such as DeepAR, N-Beats, and NHITS and a classical time series analysis technique, ARIMA. We observe that TFT significantly outperforms benchmark algorithms and can work well with a relatively smaller dataset.