Deep Learning and Hierarchical Approaches for Sporadic Demand Forecasting

In the retail industry, demand forecasting can play a pivotal role in successful management, given that accurate predictions of customer purchases are essential for inventory planning, resource allocation, and overall operational efficiency. However, achieving precise forecasts in retail, especially within the fashion domain, presents unique challenges. The fashion industry is indeed characterized and influenced by inherent uncertainties in supply and demand, a lengthy and inflexible supply process, a short life cycle, vast product variety, a high propensity for impulsive buying and other exogenous variables. This work then navigates the complex landscape of fashion retail forecasting, addressing the challenges posed by intermittent time series data and, more in general, by taking into account the forecastability of each time series. Specifically, it is proposed a deep learning solution, represented by the Temporal Fusion Transformer (TFT) model, to a real-world scenario, represented by the hosting company data. Moreover, another approach is built on top of the generated forecasts through hierarchical forecasting, by trying to leverage the company's business logic to enhance accuracy further.