Leveraging Panel Data in Retailing Research with Machine Learning Analytics: An Empirical Application

This study explores the application of machine learning (ML) techniques, specifically Long Short-Term Memory (LSTM) neural networks, to forecast retail sales using scanner data. We demonstrate that LSTM models significantly outperform other benchmark models in terms of forecasting accuracy. Our empirical analysis, based on Circana ILD weekly data from 9,268 hypermarkets in Italy, shows that LSTM models achieve average prediction errors as low as 10%, outperforming the meta-learning framework by reducing errors by 50%. The proposed LSTM framework simplifies the forecasting process by eliminating the need for complex base forecasters, making it more interpretable and actionable for retail practitioners. Key features influencing the model's predictions include historical sales volumes, promotional prices, and sales volumes without promotions. The results highlight the potential of ML to bridge the gap between retail research and practice, offering a robust tool for accurate sales forecasting in the retail sector.