Interpretable Slow-Moving Inventory Forecasting: A Hybrid Neural Network Approach with Interactive Visualization

Slow-moving inventory (SMI), which absorbs working capital over long periods of time and pushes up storage cost, urgently requires reliable and intelligible forecasting methods for making business decisions. We suggest the use of the Temporal Fusion Transformer as the backbone to fuse the graph attention layer, to capture the substitution effect and promotion transmission between SKUs. Secondly, multi-scale expansion causal CNNs account for both long-term and short-term seasonal patterns while Bayesian residual branches measure the uncertainty of prediction. Attention-based feature selectors are designed in the training stage, while SHAP interpretation and counterfactual inference are integrated in the inference stage to interpret how price, demand, and logistics signals contribute to SMI prediction. All the results are integrated into the adaptive control chart of the interactive visual display of feature attribution heat map, forecast interval and core KPI Inventory Turnover in real time, and automatically launch early warning and hypothesis testing and scene simulation when anomalies are detected, to help managers to judge whether to advance the replenishment strategy or clearance strategy, to achieve the closed loop of forecasting and decision. Simulations conducted by a multinational consumer electronics retailer showed an increase in inventory turnover of approximately 14.6%.