A sequence to sequence long short-term memory network for footwear sales forecasting

• Main Author: Santos, Luís
• Publication Date: 2022
• Other Authors: Matos, Luís Miguel, Ferreira, Luís, Alves, Pedro, Viana, Mário, Pilastri, André, Cortez, Paulo
• Language: eng
• Source: Repositórios Científicos de Acesso Aberto de Portugal (RCAAP)
• Download full: https://hdl.handle.net/1822/86238
• topic: Time series forecasting; ARIMA; Prophet; Deep learning; Science & Technology

Footwear sales forecasting is a critical task for supporting product managerial decisions, such as the management of footwear stocks and production levels. In this paper, we explore a recently proposed Sequence to Sequence (Seq2Seq) Long Short-Term Memory (LSTM) deep learning architecture for multi-step ahead footwear sales Time Series Forecasting (TSF). The analyzed Seq2Seq LSTM neural network is compared with two popular TSF methods, namely ARIMA and Prophet. Using real-world data from a Portuguese footwear company, several computational experiments were held. Focusing on daily sales, we analyze data recently collected during a 3-year period (2019-2021) and related with seven types of products (e.g., sandals). The evaluation assumed a robust and realistic rolling window scheme that considers 28 training and testing iterations, each related with one week of multi-step ahead predictions. Overall, competitive predictions were obtained by the proposed LSTM model, resulting in a weekly Normalized Mean Absolute Error (NMAE) that ranges from 5% to 11%.