Due to the characteristics of e-commerce sales data—high dimensionality, diverse types, and nonlinear relationships—traditional models exhibit insufficient adaptability. This study aims to construct a deep ensemble prediction model to provide scientific support for enterprise inventory and marketing decisions. The model integrates convolutional neural networks (CNN) with an improved weighted deep forest (WDF). A multi-granularity scanning mechanism enhances perception of local spatiotemporal features; a binary adaptive differential evolution algorithm dynamically selects key features; and Bayesian optimization precisely tunes model hyperparameters. Experiments are conducted using large-scale data from mainstream e-commerce platforms. Results show that the model achieves a mean absolute percentage error (MAPE) of only 6.8% on the test set, outperforming baseline models such as ARIMA, random forest, and XGBoost in capturing nonlinear trends and long-term dependencies. This adaptive ensemble model effectively addresses feature redundancy and model mismatch issues, significantly improving the stability of sales prediction and providing reliable technical support for precise e-commerce operations.