Demand forecasting in cross-border live streaming commerce: An explainable multimodal AI framework for market segmentation

Cross-border live streaming e-commerce serves as a powerful driver of international trade growth. However, critical gaps remain in current research: (1) inadequate theoretical modelling of multimodal information fusion, (2) limitations in the interpretability of predictive algorithms, (3) single-modality approaches restricting generalisability, and (4) unclear mechanisms linking cultural variation to live stream sales cycles. To address these gaps, we propose a novel Multimodal Dynamic Prediction Framework that integrates Temporal Fusion Transformers (TFT), Bidirectional Encoder Representations from Transformers (BERT), and Contrastive Language-Image Pretraining (CLIP) for capturing multimodal dynamics in live-commerce forecasting. Empirical evaluation demonstrates that our framework outperforms baseline approaches significantly. Key findings include: (1) Elucidation of a novel dual-track evolution of KOL influence, characterised by high sensitivity and rapid decay in emerging markets versus gradual penetration and sustained impact in mature markets. (2) Identification of core cyclical patterns underpinning the live streaming restructuring of traditional retail. (3) Temporal variation in multimodal feature contributions confirms an inverted U-shaped relationship between cognitive load and information digestion efficiency. Our framework provides actionable insights for precision marketing and dynamic KOL allocation, advancing both theory and practice in global e-commerce.