Abstract: Consumption is one of the “three carriages” driving economic growth. As China’s economy has entered the new normal, our economic mode is gradually shifting to the consumption-driven growth, making consumption the primary driver of economic expansion. The total retail sales of social consumer goods, as a direct representation of domestic consumption demand, is an important economic indicator for measuring the level of market prosperity. Therefore, scientifically and effectively forecasting its development trend is helpful in understanding social consumption levels and consumption growth trends, and can provide important references for the formulation of macroeconomic policies by the national government. Currently, domestic and foreign scholars have conducted extensive research on the forecasting of total retail sales of social consumer goods, and the proposed forecasting methods have played a significant guiding role in practical work. However, there are still some limitations. The total retail sales of social consumer goods rely on dynamic data from the consumption market and exhibit non-linearity and non-stationarity characteristics, making it difficult for linear forecasting methods to reveal its underlying patterns. In addition, the prediction error contains valuable information, which is often ignored by the existing forecasting methods.
Therefore, this paper follows the modeling idea of “decompose first, then integrate” and introduces the error correction method to construct the error correction combination prediction model based on VMD-DE-LSSVM. Firstly, the original sequence is decomposed by Variational Mode Decomposition (VMD), and then the decomposed sub-sequences are predicted by using the least squares support vector machine (LSSVM) optimized by the differential evolution algorithm (DE), and the predicted results are integrated to obtain the preliminary predicted values. Then the VMD-DE-LSSVM model is used to synchronously predict the initial prediction errors. Finally, the initial prediction value is corrected by the error prediction value. The study focuses on the year-on-year growth rate data of total retail sales of social consumer goods in Anhui Province from January 2010 to December 2020. In terms of selecting forecasting variables, relevant studies are referenced, and indicators are selected based on comprehensiveness, accessibility, and reliability. The selected indicators cover various aspects such as residents’ income, price levels, consumption, industry development, and fiscal conditions. There are 15 economic indicators specific to Anhui Province and 28 national economic indicators, making a total of 43 forecasting indicators. To ensure the usability of experimental data, missing values and outliers are handled, and dataset S₁ is constructed. Due to the impact of the COVID-19 pandemic, there are significant fluctuations in the data during February and March 2020, resulting in many outliers. The local outlier factor algorithm (LOF) is employed to detect the outliers, which are then treated as missing values. Subsequently, the K-nearest neighbor algorithm (KNN) is used to fill in all missing values.
According to the experimental results, the following conclusion can be drawn. First, the introduction of VMD effectively addresses the non-linearity and non-stationarity of total retail sales of social consumer goods during the initial forecasting. Second, the combined forecasting model based on VMD-DE-LSSVM with error correction method fully utilizes the valuable information embedded in the error sequences, leading to a significant improvement in the accuracy of the corrected forecast. Third, the forecasting accuracy of this model is significantly better than other control models, indicating that this model can provide a new modeling method and research idea for the forecasting of total retail sales of social consumer goods.
However, the model proposed in this article still has room for improvement. In the preliminary predictions, only publicly available government statistical indicators are considered, which has certain limitations in terms of timeliness and accuracy. Subsequent research can expand to include non-governmental statistical indicators that possess strong timeliness.

Key words: total retail sales of consumer goods, error correction, variational mode decomposition, least squares support vector machine, differential evolution algorithm

摘要： 社会消费品零售总额具有非线性、非平稳性特点,单一模型很难准确预测。此外,预测误差中蕴含着有价值信息,而现有预测方法往往忽略这些重要信息。为此,本文遵循“先分解后集成”建模思路,并引入误差校正方法,构建了基于VMD-DE-LSSVM误差校正组合预测模型。首先通过变分模态分解(VMD)将原序列分解,再利用差分进化算法(DE)优化的最小二乘支持向量机(LSSVM)对分解得到的子序列分别预测,并将预测结果集成得到初步预测值;然后利用VMD-DE-LSSVM模型对初步预测误差进行同步预测;最后通过误差预测值对初步预测值进行校正。实验结果表明该模型预测精度显著优于其它对照模型,不仅能有效解决社会消费品零售总额的非线性和非平稳性,还能充分利用误差序列中蕴含的有价值信息。

关键词: 社会消费品零售总额, 误差校正, 变分模态分解, 最小二乘支持向量机, 差分进化算法