Price-dependent demand forecasting of consumer packaged goods: a hybrid Prophet-LSTM approach

This paper deals with the demand forecasting of consumer packaged goods. While Prophet and LSTM are established tools in time-series forecasting, known for their reliability and effectiveness over traditional techniques, the main goal achieved in this paper is the development of a hybrid Prophet-LSTM framework that outperforms both Prophet’s and LSTM’s capabilities, bridging the gap between time-series data mining and economic analysis. The methodology has been assessed on a large scale using a local modelling approach, outperforming Transformer-based architectures as well, with results confirmed by well-suited statistical tests. Furthermore, integrating the proposed framework with the PCHIP algorithm has led to a method to forecast demand curves of goods over a medium-term horizon. In addition, two reliability criteria have been assessed for the generated forecasts: an a posteriori qualitative criterion derived from the law of demand from microeconomics, and an a priori quantitative criterion based on the coefficient of variation metric. Accordingly, two model selection approaches have been introduced. Each involves switching to a more suitable deep learning model when the corresponding reliability criterion classifies Prophet’s forecasts as not very reliable. Finally, the economic perspective has clearly emerged as a key factor guiding the choice between the proposed model selection strategies.