Forecast Of China's Macroeconomic Growth Based On MF-VAR Model

When using the traditional econometric model to study the problems like the growth of macro-economic,the variables in the regression analysis must have the same frequency and the same length.In many cases,only using the observation at same time will lead the reduction of variables.Especially that when studying the factors affecting the change of an index in the analysis,we usually does not put enough factors contained in the same model for the restriction of the sample data structure.Based on the basic requirement of the traditional econometric model,many researchers studying the multifactorial model have started to try to change the data structure of the research variables to make the model more convincing.Methods of changing the data structure,including:addition,substitution,and the filling method by using different interpolation techniques.Addition and substitution is mainly used to transfer the data structure with a high frequency of some stock or flow index tothe data structure with a low frequency data structure.Using various filling methods to fill the data is mainly used in transferring the low frequency data structure to high frequency data structure.Although the methods of changing the data structure is feasible and usually has the advantages of simple calculation,but artificially changing data structure may undermine the relationship between the original variables,because of the data with different time interval havedifferent economic meaning in a certain period.Besides,they have different trends and fluctuations.Therefore,although these methods are simple,they are only suitable for small changes in the data,andnot suitable for time series data with large number.For example,time series data filled by filling methods does not trulyrepresent the original fluctuation properties and distributioncharacteristics when there are many non-random missing values.With the prosperityof economic market,the continuous improvement of statistical technology,China has accumulated a large number of statistical data.The need of studying the relationship between the different data structure becomes more urgent.In recent years,with the improvement of data processing ability of the computers,the studying of econometric model by using the computer technology has made a great breakthrough and some new econometric modelscare founded.And some traditional econometric modelhave been improved including the mixed-frequency data vector autoregressive model(MF-VAR)which is a kind of mixed-frequency data model.From the modeling perspective,mixed-frequency data model includes variables with different frequency into a model without changing the original data structure,it uses the sample information within different variables to study an index or actual problems.The mixed-frequency data model mainly includes two kinds of models:the mixed-frequency data vector autoregressive model(MF-VAR)and the mixed data sampling model(MIDAS).These two kinds of models are the improvement of the original traditional econometric model.Among them,the MF-VAR model is animprovement of traditional VAR model by adding the mixed-frequency data,the MIDAS model is an improvement of distributed lag model by adding the mixed-frequency data.There are many studies of the MIDAS model and many derivative models of the MIDAS model.In addition,the MIDAS model is based on the estimation of Calman filter,this estimation method has some deficiencies in the hypothesis condition and the processing of irregular frequency data.This paper mainly discusses another mixed-frequency data model which is MF-VAR model.The MF-VAR model is the improvement of the traditional VAR model,it uses BMF method based on Gibbs sampling to estimate the parameter and missing values.This estimation method is based on the distribution of the original data,and it guarantee the data's original economic significance andavoid the drawbacks of artificial intervention to some extent.The principle of the model construction,the prior distribution,the selection of the hyper parameters and the estimation procession of the parameters are introduced in detail in the second chapter,In the comparative analysis of the MF-VAR model for exploring the impact of high-frequency data information on the forecast of low-frequency variables,the estimation of the benchmark model still use the traditional least squares method.It is not inequitable to use two kinds ofestimation methods,it cannot reflect the effect of high-frequency information on the forecast oflow-frequency variables.From this point,this paper makes a comparative analysis of the MF-VAR model,and uses the Bayesian method to estimate the two models.In addition,the effects of hyper parameters on the model effect is relatively large,many scholars did not do too much research on the hyper parameters,they just use a simple set of hyper parameter or compare few hyper parameters.Based on this point,this paper uses the grid search to optimize the parameters.Through the analysis of empirical results,this paper found that the MF-VAR model using missed-frequency data can significantly improve the prediction accuracy of the traditional VAR model with the same frequency in the short term.But in the long-term forecast period,there is no obvious advantage of MF-VAR model.The empirical part in the third chapter compared the MF-VAR model and MIDAS model.The results show that,in most cases,the prediction accuracy of MF-VAR model is better than that of MIDAS model,but the performance of each forecast period is different.In the long-term forecast period,there is no significant difference between the two models.