In today’s rapidly changing economy, accurately forecasting demand is essential for optimizing stock levels and supporting better decisions in supply chains. Traditional models fall short in capturing complex and nonlinear consumer behaviors. This study investigates how integrating multimodal data, such as product images, product descriptions, and tabular sales and economic data, can improve demand forecasting models by capturing a wider range of influencing factors. We evaluate two advanced deep learning models, Long Short-Term Memory (LSTM) and Temporal Fusion Transformer (TFT), both capable of learning complex temporal patterns. To incorporate multimodal data, two approaches are proposed, each using different methods to extract visual and semantic features from product images and descriptions. These approaches aim to improve the models’ ability to understand both visual and textual features. We highlight that TFT offers higher accuracy and that it has the added advantage of interpretability through its attention mechanism in comparison to LSTM, which helps identify the key factors affecting demand. Our work demonstrates the effectiveness of integrating multimodal data into deep learning models for improving demand forecasting, offering valuable insights for retailers and supply chain specialists to enhance forecasting precision in the context of essential goods.