Real-time Prediction Of CPI Based On Mixed Frequency Stream Data

The report of the 19 th National Congress pointed out that the new development concept should be used to lead the construction of a modern economic system,and macroeconomic monitoring and forecasting is the core content of economic system construction and regulation.Among them,CPI is an important economic indicator reflecting the level of inflation,and its forecast can provide an important basis for government's macroeconomic regulation.With the advent of the era of big data,real-time,interactive,discretized and unstructured massive data contains various leading indicator signals for the development of economy and society,and data processing technology and processing power are greatly improved accordingly.The introduction of high frequency data appears more important and feasible for CPI prediction.On the basis of the research of CPI prediction based on macroscopic low-frequency data,this paper introduces high-frequency big data suitable for CPI prediction,that is,the Baidu index,and apply the mixed frequency data model and the method of stream data paradigm to predict CPI.Thus,this paper solves the past problem of prediction lag caused by insufficient clarification of the influencing factors and the frequency gap between high-frequency data and low-frequency data and at the same time expands the mixed-econometric econometric method of the stream data type to realize the real-time dynamic prediction of CPI.This paper firstly uses the Baidu index data to construct the search heat index,which eliminates the natural growth of search volume caused by the rapid development of the Internet and the fluctuation of search volume caused by the change of netizens' behavior and thus only the influence of social and economic events is retained.It is found that there is a strong correlation between the search heat index of some key words and CPI.Further,the time relation analysis method is used to select the strong-correlated leading indicators with predictive effect.The correlation coefficients of the search heat in the time relation analysis are used as the weights to synthesize the high and low frequency search index,and then the synthesized search indexes are used to construct the low-frequency data model and the mixed-data model to predict the CPI,which are compared with the benchmarkmodel.In-sample prediction and out-of-sample prediction are carried out in the process of prediction.The in-sample prediction of the MIDAS model further adopts the h-step forward-going method to predict the CPI in real time,while the out-of-sample prediction further back-tests the robustness of the constructed search index and stability of the model,in which static prediction,dynamic prediction and rolling window prediction catering to the characteristics of data stream.The following conclusions drawn:First,in the in-sample prediction,the prediction accuracy of MIDAS model is significantly better than the benchmark model and the low-frequency model,indicating that the full and lossless use of high-frequency information improves the prediction effect and avoids the data information loss caused by the traditional low-frequency data model due to the fact that the frequency is artificially summed down.Thus,the original information is used to the maximum extent via MIDAS model which performs more realistically and reliably on estimation and prediction.Second,using the h-step forward-going mMIDAShpARK)(),,(-model to predict CPI in real time can precede the official data by 29 days.If out-sample dynamic prediction is applied,it can offer the forecast prior to the official data by 34 days thus further improving the timeliness of the forecast.The advantages of the mixed frequency data method as against the low-frequency model are not only manifest in the accuracy of the prediction but also in the timeliness of the prediction and the updatability of the result.Compared with the benchmark model,the superiority of the MIDAS method is mainly reflected in that the full use of high frequency information to the latter part of the month significantly improves the prediction accuracy.Third,in the dynamic out-sample forecast,the mixed frequency data model has significant superiority in prediction accuracy and trend fitting,and can grasp the trend and directional changes of CPI in a timely manner in practical application,which has great practical value.Out-sample dynamic prediction not only verifies the robustness of the search indexes,but also further validates the dual advantages of the mixed frequency data model in prediction accuracy and timeliness.Fourth,this paper uses the rolling window technology which is more suitable forstreaming data to implement the rolling window prediction of CPI.The predicted values make good fitting for the CPI change trend from June 2016 to July 2018 and are consistent with the static out-sample forecast result,which indicates that the models are relatively stable and no obvious concept drift happened during the entire sample interval.