【摘要】 为弥补传统预测变量及预测技术的不足，本文基于深度学习长期和短期时间序列网络(LSTNet),结合网络搜索数据与政府统计指标，构建LSTNet&BI模型开展浙江省及地级市社会消费品零售总额的预测研究。研究发现：(1)引入网络搜索数据能够有效提高LSTNet模型的预测性能与预测精度；(2)LSTNet&BI模型具有较好的泛化能力，对浙江省社会消费品零售总额的短期和长期预测效果较稳定，其预测性能与预测精度均优于其他基准模型；(3)LSTNet&BI模型具备较强的稳健性，对杭州市、绍兴市和衢州市社会消费品零售总额的预测效果也较好。更多还原

【Abstract】 Total retail sales of consumer goods is a key indicator of social consumption demand, and can provide a relatively comprehensive picture of consumers’ comprehensive judgement of the current and future economic situation, market fluctuations and personal income expectations. At present, the forecast of total retail sales of consumer goods is mostly based on traditional government statistics, which is not ideal due to the single system of forecasting indicators and the lack of micro information reflecting the consumption of social groups. In the context of big data, the introduction of web search data can effectively break through the limitations of traditional data such as small capacity and poor timeliness, and help improve the performance of the forecasting model.In order to make up for the shortcomings and deficiencies of traditional forecasting variables and forecasting techniques, this paper constructs an LSTNet&BI model for forecasting total retail sales of consumer goods based on the Long-and Short-term Time series Network(LSTNet) model, which combines Baidu search data with government statistical indicators. This paper focuses on the following four aspects. First, based on the existing literature, we construct a comprehensive library of web search keywords related to the total retail sales of consumer goods from four aspects: macroeconomy, financial environment, consumer demand and consumer goods supply, and adopt dynamic time regularization and mRMR algorithms to select the forecast variables. Second, we introduce the selected Baidu index and related socio-economic indicators into the LSTNet&BI model and further optimize the hyperparameters in the model to accurately forecast the trend of total retail sales of consumer goods in Zhejiang province. Third, based on the evaluation indexes of the forecasting performance, a comprehensive and systematic comparison and analysis of the forecasting effect of the LSTNet&BI model and other benchmark models is conducted. Fourth, the LSTNet&BI model is used to carry out a study of the forecast of total retail sales of consumer goods at the prefecture-level city to enhance the value of the model and verify its robustness.It is found that:(1)The introduction of web search data can effectively improve the forecasting ability of the LSTNet model for total retail sales of consumer goods.(2)The forecasting error of the LSTNet&BI model does not change significantly with the length of the forecasting period, and the model has strong stability.(3)The LSTNet&BI model has better forecasting performance and accuracy for both short-term and long-term total retail sales of consumer goods, and the forecasting results are significantly better than those of other benchmark models.(4)The LSTNet&BI model possesses strong robustness and has good forecasting performance for total retail sales of consumer goods in Hangzhou, Shaoxing and Quzhou. The research results show that the LSTNet&BI model has certain practical value, and the method provides a new way of thinking for total retail sales of social consumer goods forecasting, enriching the research on the application of machine learning in the field of macroeconomic indicators forecasting.更多还原