From product sequence to sales: A prediction framework incorporating multi-source social cues in livestream e-commerce

As livestream e-commerce becomes increasingly competitive, accurate sales prediction is crucial for merchants' operational decisions. During livestreams, consumers are typically exposed to multiple products sequentially, with information from previously shown products potentially providing additional predictive signals for subsequent product sales. However, existing sales prediction research often treats products as relatively independent observational units and overlooks the role of product sequences in prediction. To address this gap, we develop a dynamic sequential sales prediction framework for livestream e-commerce by integrating multi-source social cues. The framework examines the informational value of two sources in the product sequence, namely product attributes and bullet screens, for predicting subsequent product sales. Furthermore, we distinguish between product-focused discussion and background atmosphere in bullet screens. We also propose a sentiment analysis method for background atmosphere based on large language models (LLMs). Our experiments are conducted using a real-world dataset collected from Douyin livestreams. The results show that information from preceding products improves subsequent sales prediction, while sentiment-classified product-focused discussion and background atmosphere further enhance predictive performance. These findings provide practical support for livestream planning and dynamic operational decision-making in livestream e-commerce.