Dual-Stream Neural Network for Real-Time Sales Forecasting and Inventory Cost Optimization

This paper proposes a Dual-Stream Neural Network (DSNN) for real-time sales forecasting and inventory cost optimization in supply chain analytics. The model integrates two parallel processing streams—one for temporal features such as stock levels and lead times, and another for static features including supplier information and transportation modes. An attention-based fusion mechanism combines these representations, followed by multi-task learning to jointly predict product sales and inventory costs. Experiments were conducted on a realworld dataset from the beauty and cosmetics domain, consisting of of 100 product records with 24 attributes. The proposed model achieved an RMSE of 52.34 for sales prediction and 45.21 for for inventory cost estimation, outperforming baseline models such as Random Forest and XGBoost by 13-20%. Additional ablation studies confirmed the contribution of both data streams and the attention layer, while robustness tests demonstrated stability under input noise. These results indicate that DSNN offers an effective and scalable solution for data-driven decision support in dynamic supply chain environments.