Study On Prediction Of Consumer Confidence Index Based On Web Search Data

From 2011 to 2019,consumption always contributed more than 50% to China’s economic growth,steadily became the primary driving force of economic growth.Hit by the COVID-19 epidemic,China’s GDP fell by 6.8% in the first quarter of 2020 compared with the same period last year.In response,local governments issued a number of policies to boost consumer confidence and promote sustainable and stable growth of consumption.Consumer confidence index(CCI)is an economic index that reflects the strength of consumer confidence and willingness to consume,and plays an important role in forecasting and warning consumer trends and economic development situation.At present,the release of CCI is not timely,and the accuracy and stability of relevant forecasting studies are difficult to guarantee.How to timely and accurately forecast CCI and understand the changing trend of consumer confidence in advance are issues worthy of attention and research.In this paper,time-leading web search data was used,different dimensionality reduction methods were used to screen prediction variables and reduce redundant features,and optimization algorithm was introduced to build prediction models,in order to improve the timeliness,accuracy and stability of prediction results.Firstly,this paper analyzed the internal relationship between Internet search behavior and consumer confidence from a qualitative perspective,built a theoretical framework,and selected web search data from six aspects of economic situation,living conditions,price level,employment,investment and house purchase according to the constitution principle of CCI,combined with relevant literature.Secondly,quantitative analysis was carried out on the selected web search data,data preprocessing was used to reduce the interference of uncertain factors,Pearson correlation analysis and time difference correlation analysis were used to eliminate the influence of noisy data,and keywords with strong correlation and leading changes were selected.Then,three dimensionality reduction methods,namely Lasso algorithm,stepwise regression and principal component analysis,were used for dimensionality reduction analysis of keywords from different perspectives to obtain prediction variables.Finally,the prediction variables selected by different dimensionality reduction methods were put into the constructed LSTM and SVR prediction models,and the prediction results were compared and analyzed;then the SSA algorithm was introduced to construct SSA-LSTM and SSA-SVR prediction models,the prediction variables selected by the optimal dimension reduction method were used for prediction,and the prediction effect and accuracy before and after the introduction of SSA algorithm were compared and analyzed;and then the stability analysis of the prediction model was carried out,and the optimal prediction model was adopted for three periods of future prediction to obtain the predicted value of CCI monthly data,which was compared and analyzed with the actual value.The results show that:(1)there is a significant correlation between web search data and CCI,and the use of web search data to predict CCI has the advantages of leading and low cost.(2)In the same prediction model,the prediction effects of different dimensionality reduction methods on the prediction variables are significantly different,and the prediction results using the composite index are better than those using single keywords as the prediction variable.(3)Compared with the benchmark model,the prediction accuracy,prediction effect and stability of the SSA-LSTM model and the SSA-SVR model constructed by the introduction of SSA algorithm are significantly improved,and the prediction performance of the SSA-LSTM model is significantly better than the other three models,with stronger prediction ability of turning point and strong stability.(4)Constructing SSA-LSTM model with network search data to predict CCI can not only improve the prediction accuracy,but also improve the timeliness of prediction.The predicted value is one month ahead of the official data.