Research on the application of deep learning in E-commerce dynamic pricing and demand forecasting

The separation of e-commerce dynamic pricing and demand forecasting often leads to the accumulation of forecast bias and the lag of pricing strategy, as the two are in fact a bidirectionally coupled closed-loop system. We propose a collaborative forecasting-pricing framework: the demand forecasting module based on Transformer outputs the point estimate of demand and price gradient, and the Proximal Policy Optimization (PPO) pricing module optimizes the continuous price accordingly. At the same time, the actual demand feedback generated by the pricing decision continuously updates the forecasting model, forming a closed loop of mutual promotion. On UCI Online Retail and M5 datasets, the RMSE of demand forecasting (on dataset-specific normalized scales) reaches 15.7 and 22.3, respectively, which is 16.5% and 13.9% lower than the optimal baseline Informer, respectively. The cumulative revenue of dynamic pricing is 883 (normalized simulator metric, not real money, for relative comparison only), 5.0% higher than that of the separated scheme. The ablation experiment shows that the price gradient feedback and prediction module continuously update the revenue gain of 2.8% and 4.8% respectively. The two modules work together to reduce the prediction error by 26.0%, which effectively alleviates the impact of demand drift. The framework breaks the traditional separation bottleneck through the closed-loop feedback mechanism, and the gain is particularly significant on high-elastic goods, which provides a feasible joint optimization path for e-commerce real-time pricing decisions. Not applicable. This study does not involve clinical trials or human/clinical intervention.