Let clickstream talk: A graph neural network approach to sales forecasting

Abstract This article proposes ForecastClickGraph, a deep learning framework that extracts cross-product relationships and demand information from clickstream data for probabilistic sales forecasting. ForecastClickGraph models stages in consumer shopping journeys as a dynamic graph, where nodes represent individual product-stage combinations, such as product A viewed or added to cart, and directed edges capture the transitions between stages. A customized graph neural network learns product representations that incorporate both product-level temporal patterns and cross-product associations. Another graph-based module further detects demand spikes using early signals from related products. ForecastClickGraph cross-learns demand patterns across large-scale products and estimates demand distributions through quantile regression. Extensive experiments on a real-world dataset show that ForecastClickGraph outperforms state-of-the-art benchmark models by 10%–28% in forecast accuracy, with particularly strong performance in predicting promotional sales bursts. It also yields superior probabilistic forecasts with substantially lower quantile loss and improved calibration relative to the benchmarks.