This paper sets out a model to predict quarterly real GDP growth in the current and the next quarter. Towards this end, a parsimonious quarterly model that can predict quarterly real output growth based on monthly indicators is constructed. The monthly indicators that will be used to forecast real GDP are selected on the basis of (i) in-sample performance, (ii) out-of-sample forecast performance, and (iii) availability of the data relatively early in the quarter. Our finally chosen model contains only two variables: industrial production index (IP) and service production index (SVR).
But in order to produce forecasts of real GDP growth we should generate forecasts of the monthly indicators themselves. We use a bivariate Bayesian Vector Autoregression (VAR) model to forecast monthly predictors. The Bayesian VAR models overcome over-parametrization and over-fitting problem of VAR model by imposing ‘Bayesian priors.’ A widely used Minnesota prior is implemented through dummy observations. A comparison of pseudo-out-of-sample forecast performance of bivariate VAR and BVAR model shows that BVAR (12) outperforms VAR (12) in terms of Root Mean Squared Error (RMSE) and Mean Absolute Error (MAE).
Our model, combining quarterly forecast model of real GDP growth and monthly forecast model of IP and SVR, shows an excellent performance. The Figure 1 demonstrates fan chart of predictive distribution for the real GDP growth for 2011. The third quarter and the fourth quarter based on estimated model using sample covering from January 2001 to August 2010. The Table 1 shows how the predictive density for current and next quarter real GDP growth changes as available monthly indicators increase.
But in order to produce forecasts of real GDP growth we should generate forecasts of the monthly indicators themselves. We use a bivariate Bayesian Vector Autoregression (VAR) model to forecast monthly predictors. The Bayesian VAR models overcome over-parametrization and over-fitting problem of VAR model by imposing ‘Bayesian priors.’ A widely used Minnesota prior is implemented through dummy observations. A comparison of pseudo-out-of-sample forecast performance of bivariate VAR and BVAR model shows that BVAR (12) outperforms VAR (12) in terms of Root Mean Squared Error (RMSE) and Mean Absolute Error (MAE).
Our model, combining quarterly forecast model of real GDP growth and monthly forecast model of IP and SVR, shows an excellent performance. The Figure 1 demonstrates fan chart of predictive distribution for the real GDP growth for 2011. The third quarter and the fourth quarter based on estimated model using sample covering from January 2001 to August 2010. The Table 1 shows how the predictive density for current and next quarter real GDP growth changes as available monthly indicators increase.
Hwang Jong-ryul