主管:中国科学院
主办:中国优选法统筹法与经济数学研究会
   中国科学院科技战略咨询研究院

Chinese Journal of Management Science ›› 2022, Vol. 30 ›› Issue (12): 63-76.doi: 10.16381/j.cnki.issn1003-207x.2021.2635

• Articles • Previous Articles    

Forecasting Online Retail Sales of China Based on Splitting-filling-decomposition-ensemble Model

ZENG Neng-min1, 2, ZHANG Ming1, YU Le-an1, 2, 3   

  1. 1. School of Economics and Management, Harbin Engineering University, Harbin 150001, China;2. Key Laboratory of Big Data and Business Intelligence Technology Harbin Engineering University, Ministry of Industry and Information Technology, Harbin 150001, China;3. Business School, Sichuan University, Chengdu 610065, China
  • Received:2021-07-27 Revised:2022-06-24 Published:2023-01-10
  • Contact: 余乐安 E-mail:yulean@amss.ac.cn

Abstract: In recent years, China’s online retail industry has developed rapidly. Accurate prediction for online retail sales is the basis for government to formulate retail policies, as well as the foundation for ecommerce and logistics companies to determine operation strategies. However, no existing research has focused on macro online retail sales prediction driven by data characteristics. Forecasting the monthly online retail sales of China is a great challenge because the dataset has the characteristics of small sample size, high volatility, large holiday influence. In addition, the China’s online retail sales data has a unique data missing phenomenon: the total number of January and February is known but the monthly value is missing, which is caused by the relevant regulations of the National Bureau of Statistics. Motivated by these, a splittingfillingdecompositionintegration (SFDE) prediction framework is proposed. Specifically, firstly, the data set of online total retail sales of China is split into two parts, i.e., physical retail sales data and nonphysical retail sales data. Secondly, in the light of the incompleteness of online physical retail sales data, a revised spline interpolation approach (i.e., the hybrid approach of spline interpolation and dichotomous adjustment) is proposed to fill the missing value of the data. Meanwhile, considering that nonphysical retail data has different trends at different stages and increasing fluctuations, another revised spline interpolation approach (i.e., the hybrid approach of piecewise linear function fitting and spline interpolation) is proposed to fill the missing value of the data. Thirdly, based on the different characteristics between the physical retail data and nonphysical retail data, two hybrid ensemble forecasting approaches are proposed to predict the above two series, where the first one integrates multiplication decomposition, ARIMA and moving average, and the second one integrates STL decomposition, BP neural network and gray waveform forecasting. Finally, the prediction results of the above two series are integrated to get the predicted value of online total retail sales of China. In our experiments, the monthly data of China’s online retail sales from 2015 to 2019 are selected to verify the model performance. The results obtained in this study show that the revised spline interpolation approaches based on data characteristics are able to solve the problem of mentioned missingdata filling effectively. In addition, the combination of the revised spline interpolation approaches and the hybrid ensemble forecasting approaches achieved significant performance improvements over single model. Furthermore, the SFDE framework of combining the strengths of the conventional and deep learning methods provides a robust modelling framework capable of capturing the nonlinear nature of the complex online retail sales series and thus producing more accurate forecasts. On the whole, the proposed hybrid framework enriches the research of missing value filling methods, and have tremendous scope for application in a wide range of areas for achieving increased accuracies in complex time series forecasting.

Key words: online retail; forecasting; missingdata filling; decomposition ensemble

CLC Number: