Empirical Analysis Of Combination Forecast Of Macroeconomic Indicators

Predicting significant macroeconomic indicators has become an important basis for government leaders and economic managers to formulate and implement economic policies and conduct scientific management,which plays an important role in economic construction and sustainable development in daily lives.What's more,the formulation and implementation of effective macroeconomic policies have also become an important part of the work of government departments and central banks.However,due to the time lag of economic policies,it will take a period of time for the current economic policies to take effect,and the future economic environment may change with external conditions.Therefore,the implementation of macroeconomic policies requires forecasting the macroeconomic.Only by making correct judgments on future economic trends can the role of economic policy be effectively brought into play.If the important macroeconomic variables can be accurately predicted,the government and central bank can adopt appropriate macroeconomic policies to achieve the purpose of policy implementation.In order to achieve the purpose of forecasting important macroeconomic indicators,taking some effective time series forecasting models is essential.When forecasting key macroeconomic indicators,many researchers have proposed a variety of different forecasting methods,including ARIMA models,error correction models,vector autoregressive models,mechanism conversion models,mixing models and dynamic factor models.These single time series forecasting models can produce better forecast results under certain conditions,but sometimes the forecast errors are also large.Therefore,many single time series forecasting models have their applicability.Each forecasting method owns a specific field and environment.If only a single time series forecasting method is used,the information contained in the data not only cannot be fully mined,but sometimes the forecast results are unstable,and the model is greatly affected by abnormal points or outliers.Combination forecasting method can solve the problem of unstable and low accuracy of a single time series method,which has been widely used in weather forecasting,energy forecasting,geographic information system construction and financial product yield forecasting.This paper selects two macroeconomic indicators including GDP growth rate and CPI growth rate as the modeling objects,and constructs a combined forecasting model for these two macroeconomic indicators.The data set is the quarter from the first quarter of 1998 to the fourth quarter of 2019.In order to establish a combined forecasting model for macroeconomic indicators,this paper firstly establishes six single time series models including exponential smoothing model,ARIMA model,BP neural network model,Theta model,STL model and Tbats model.The corresponding time series model is tested and the residual error is diagnosed.The combined forecasting model is constructed on the basis of establishing each single time series forecasting model.Unlike many documents,this article does not only establish the forecasting model that contains the most combined methods and compares it with six single forecasting methods.The methods can be combined arbitrarily,getting 57 different prediction method.When performing different combinations of a single prediction method,this paper proposes three combined prediction methods with different weight settings,including equal weight method,inverse root mean square error method,and cross-validation inverse root mean square error.The equal weight method refers to simply average the prediction results according to the root mean square error in the train set.The inverse root mean square error applies different weights to different single prediction models based on the root mean square error,and the prediction error is inverse to the weight.The basic idea of cross-validation inverse root mean square error method lies in combining the cross-validation of time series method and the inverse root mean square error results,re-weighting different single forecasting models.In order to effectively evaluate the three different weighting methods and observe the advantages of the combined forecasting model compared to a single forecasting model,this paper draws different error figures such as root mean square error,average absolute error and average percentage error,which can avoid single model evaluation indicators.The horizontal axis of the error graph is the model error value,and the vertical axis is 63 prediction models including single and combined prediction methods.The horizontal axis can compare the advantages and disadvantages of the combined prediction methods under three different weight settings,and the vertical axis can be used to compared the advantages of combined forecasting methods over single forecasting methods.Finally,this paper draws an important conclusion:Compared with a single forecasting method,the combined forecasting method will make the forecasting results more robust with the increase of method numbers.The forecast accuracy of the six combined forecasting methods will be better than most forecasting methods.but it is not truth that the more the number of combination methods,the higher the accuracy of the combination prediction.When horizontally comparing the combined forecasting models under three different weight settings,the cross-validation inverse root mean square error method has the great advantage in the prediction result,and the inverse root mean square error combination forecast result is the worst.Therefore,the cross validation of the time series method combined with the root-mean square error can significantly improve the prediction accuracy of the combined prediction model,which is also an important innovation of this paper.