Real-time Forecasting Of Macroeconomic Time Series

Currently, most of the domestic macroeconomic forecasting uses the vintage data, which is the final revised data available at the time when the researchers do the work. However, this data does not characterize the real macroeconomic original state. It experienced the regular revision and comprehensive revision after the initial releasing, so contains some interference new information. Real-time data, which has the feature of the immediacy, can eliminate the interference future information on one particular data point in the past. But there are no systematic real-time data sets in China. This paper based on the publishing and revision mechanism of China quarterly GDP, and collected and summed up four columns China’s quarterly GDP data, which is the representation of the real-time data:the initial accounting data, initial verification data, the final verification data and the final data. Meanwhile we selected six monthly unanimous indicators and used two types of mixed frequency models:MIDAS and MF-VAR to forecast quarterly GDP by univariate model, multivariate model and combination forecast. We choose AR model as the benchmark and compute the relative MSE to test the models’predictive validity. The main conclusion as follows:(1) The initial accounting data has the most prediction accuracy of the four columns of real-time data, and the lowest is final data. The result is consistent with the result of revision test:most revision contains new interference information, but not reduce the noise. The model predictive accuracy of initial verification data and the final verification data is in the middle of the formers and there is not much difference between them;(2) AR-MIDAS model is suitable for short-term forecast, while MF-VAR model is suitable for long-term forecast, especially when hm= 8,9; but there is no single model with absolute advantages and disadvantages, it depends on the data generating process as well as the predictor variables and prediction step size selection;(3)The predictive accuracy of the combination forecast is higher than at least 80% of the single-variable models, especially the weighted mean method. Because it takes into account all the predictive performance of each model and gives higher weight to good performance model. But there are also some individual models whose performance is better than combination forecast.