E-commerce enterprise sales volume prediction model based on feature construction and feature selection

Abstract

E-commerce enterprises’ sales volume prediction suffers from inaccuracy under the influence of uncertainty factors, which directly affects the enterprises’ resource allocation and market competitiveness. Based on this, the study aims to optimize the linear mixed model through feature construction and feature selection in order to improve the accuracy of sales volume forecasting. Feature engineering methods are used to construct features related to seasonality, promotional activities and historical sales, and the Akaike information criterion and Bayesian information criterion are combined to select the optimal covariance structure of the model. The experimental results show that the feature-optimized linear hybrid model outperforms the traditional method in terms of prediction accuracy, with an average absolute percentage error of 0.040 and a root-mean-square percentage error of 0.074. In addition, the model’s prediction error during the Double 11 period is maintained at −8% to 2%, which is highly practical and effective. The model is highly practical and effective. This study provides new ideas for e-commerce enterprises’ sales volume prediction, and effectively improves the enterprises’ decision-making ability in the complex market environment.