Semantic-Enhanced Demand Forecasting: A Multimodal Transformer Integrating Product Descriptions and Customer Purchase History

Abstract

Accurate forecasting of daily demand at the individual customer-product level remains a critical yet challenging problem in retail, hindered by data sparsity, volatile consumer behavior, and the underutilization of unstructured product information. This study addresses this gap by proposing a novel Multimodal Semantic Transformer (MST) framework that integrates semantic product embeddings, derived from Natural Language Processing (NLP) of descriptions, with structured customer purchase history and multi-scale temporal features. Using the Online Retail II dataset, the model was evaluated against benchmarks including LSTM, Gradient Boosting Machines, and a unimodal Transformer. The results demonstrate that the MST framework significantly outperforms all benchmarks, achieving a 15.5% reduction in Mean Squared Error (MSE) compared to the best baseline. Key findings confirm that semantic fusion provides a crucial signal for sparse products and that temporal embeddings with dynamic attention conditioning are essential for modeling complex seasonality and context. The study concludes that deep multimodal integration is a transformative approach for granular demand forecasting, offering a scalable and interpretable solution to enable hyper-personalized inventory management and more resilient, customer-centric retail supply chains.