Demand forecasting in clearance season for fast fashion retail: a residual learning approach

Purpose This study focuses on forecasting demand for markdown pricing in situations where historical data are limited and proposes using a hybrid knowledge-based residual (KRL) network to enhance the accuracy of demand predictions in fast fashion retailing. Design/methodology/approach We use a hybrid KRL network structure to increase forecasting accuracy in fast fashion data by combining artificial neural network (ANN) and theoretical demand models (TDMs) such as linear, exponential and multinomial logit. We used a linear demand model (LDM) as a theoretical demand model, which is one of the popular TDMs in the literature, and combined it with neural networks utilizing the residual network structure. We tested it on real fast fashion data to estimate the next week’s demand at the clearance seasons of 5 years. Findings The results underscore KRL’s capability to derive mutual benefits from both neural networks and LDM, especially in the specific context of limited fast fashion data. Furthermore, KRL outperforms LDM and ANN models when used individually in forecasting accuracy. Originality/value The research paper proposes a scientific, quantitative method for forecasting the sales for markdown settings, combining data driven and TDMs using residual learning thereby resolving the issue of insufficient sales data in the fashion industry.