Research On CPI Short-term Prediction Based On Web Search Data

The Consumer Price Index(CPI)can measure the degree of changes in the price of service and consumer goods purchased by households over a period of time.The rate of change reflects the level of inflation or deflation.In view of the fact that the National Bureau of Statistics only release last month's CPI around the middle of each month,there was a time lag in the formulation and implementation of macroeconomic policies,which affected the effects of economic policies.Therefore,analyzing the changing laws of CPI itself and predicting it in time has important theoretical foundation and practical significance.Most of the existing CPI forecasting methods are limited to the improvement of traditional models.The data used is based on government statistics.These data have shortcomings such as limited coverage,relatively small quantity,and release time lag,which seriously affect the accuracy and timeliness of CPI forecasting.At the same time,with the in-depth popularization of the Internet to residents,people have changed from passive acquisition of information to active inquiry of information.The rapid improvement of the technical level of search engine websites has made Internet search the main way to take information.The Internet records queries and browsing history of netizens in real time,which can reflect changes in users' interests,needs,and users' future trading behavior.Web search data can effectively make up for the shortcomings of traditional prediction methods in terms of sample limitation and real-time.Based on this background,the paper attempts to explore the short-term trend of CPI by using web search data.This article builds a logical analysis framework for the relationship between commodity price fluctuations and residents' online search behavior.First,take CPI as the research object,determine the initial keywords from the perspective of economic theory and combine with text mining technology,use the long tail expansion method and the demand map method to expand the keywords to form a keyword vocabulary,use the time difference correlation analysis method,unit root test and LASSO algorithm to screen keywords,and establish a CPI short-term trend prediction model based on the selected variables.Secondly,the sample data is divided into a test set and a training set.The training set samples are used to build the model.This paper uses the network search data to construct a CPI prediction model based on the kernel partial least squares regression method,and its fitting effect is compared with the stepwise regression model;use the autoregressive model as the benchmark model and the mixed frequency data sampling regression model is applied to the CPI prediction analysis.Finally,select an appropriate model to make short-term forecasts of CPI.The main conclusions of this paper are as follows: First,the method of word selection is further improved.Compared with the prediction effect of CPI based on web search data in the existing literatures,the keywords selected in this paper have higher prediction accuracy in the model.Second,the multivariate mixing regression model has the best prediction effect among the four models constructed,and can greatly improve the accuracy of CPI in-sample fitting and out-of-sample prediction.The results show that,firstly,compared with the stepwise regression model,the kernel partial least squares regression model has comparative advantages in fitting effect and prediction accuracy,and successfully predicted the "inflection point" of CPI.Second,based on the weekly network search data,mixing regression model not only successfully captures the "inflection point" in the test set,realizes short-term prediction of CPI,but also ensures high prediction accuracy.Compared with the kernel partial least squares regression model,the mixing regression model can effectively improve the prediction effect of CPI.Third,the kernel partial least squares regression model and the mixing regression model can successfully capture most of the "inflection points" in the test set based on the network search data.Fourth,the forecasting method using online search data can provide more accurate CPI forecast results in a timely manner at the end of the month.Compared with the official data of the National Bureau of Statistics,CPI data can be obtained about two weeks in advance,providing timely information for economic policy making and economic decision-making.The contributions of this paper are as follows: First,CPI is obtained by weighting the primary price index of the specification products in China.Different from the existing compilation methods,this article predicts CPI based on web search data,which enriches the theory and methods of short-term CPI prediction,and expands the application scope of statistical index.Second,the improvement of the keyword selection method,analyze the influencing factors of price fluctuations from the perspective of the currency market and the commodity market,and combine text mining technology to select keywords from multiple perspectives to further enrich the keyword selection methods.Third,apply the multivariate mixing regression model to the prediction of CPI.This method fully excavates the laws behind the high-frequency data,which significantly improves the prediction accuracy compared with the traditional regression model.Fourth,Internet search data can reflect fluctuations in the economic environment in time.The mixedfrequency regression model takes advantage of the high granularity of high-frequency data information to provide more accurate "inflection point" prediction capabilities and has certain early warning effect.