To solve the problem that the traditional retail market is limited to single-dimensional data analysis of sales when predicting commodity sales, which ignores the long-term trend, seasonality and holiday within the sales data, making it difficult to fully capture the data characteristics and accurately predict sales. This paper constructs a Prophet-LightGBM combined machine learning model. Firstly, the Prophet model is used to automatically decompose data characteristics and flexibly adjust their impact on sales. Then, the LightGBM model is used to build an efficient prediction model for the multi-dimensional features identified by Prophet to improve the reliability of sales forecasts. Experiments have shown that the prediction R² of the combined model has reached 0.85, which is better than the prediction effect of a single model. It is proved that the proposed combined model has better performance in commodity sales forecasting, and provides strong theoretical data support for enterprises to accurately grasp market demand, optimize production planning and inventory management.