Product Price Prediction and Inventory Management for Cross-Border E-Commerce Based on Neural Architecture Search and Global Optimization

This study proposes a collaborative approach integrating neural architecture search (NAS) and global optimization for cross-border e-commerce price prediction and inventory management. First, multidimensional data was integrated into a 3D tensor centered on time periods and spatial market features. A hybrid neural network search space (including time-series branches, spatial feature branches, and fusion layers) was designed via NAS; reinforcement learning optimized the network structure, while the differential evolution algorithm tuned parameters synchronously. The inventory module built a cost model (holding, stockout, ordering costs) and solved the optimal ordering strategy using a genetic algorithm based on Ito process-modeled demand fluctuations. Experiments show a 4.72% mean absolute percentage error for price prediction, 19.3% lower total inventory cost than the particle swarm optimization algorithm, and 40% shorter NAS convergence time than baselines, providing quantifiable solutions for dynamic pricing and inventory optimization.