Demand Forecasting Using Deep Learning for Resilient and Agile Supply Chain Networks

Supply chains around the world have been becoming more prone to various forms of disruptions from
pandemics to geopolitical tensions, and it has revealed the inadequacy of existing forecasting methods.
Successful demand forecasting is the foundation of a resilient supply chain, allowing for strategic inventory
management and capacity planning. This study develops an advanced demand forecasting system based on
deep learning that leverages a Temporal Fusion Transformer (TFT) and multiple sources of external data such
as weather information, economic statistics, social media trends, and supply chain disruptions. Using the data
collected for five years (2019-2025) from the multinational retail supply chain that amounts to 50 million SKUlocation-weeks, the TFT model demonstrates WAPE = 12.4% when making forecasts for four weeks ahead,
surpassing other forecasting models (ARIMA - 24.8%, XGBoost - 18.2%, and LSTM – 15.6%). Moreover, the
developed system features a unique disruption-aware training process that increases forecast precision during
disruptions by 28%. When tested in conjunction with a multi-echelon inventory management system, the
forecasting system was able to cut the amount of safety stocks by 19% and improve on-time delivery
performance by 31%.