Macroeconomic forecasting using model averaging

Abstract

Recently, there has been a broadening concern on forecasting techniques that are applied on large data sets, since economists in business and management want to deal with the great magnitude of information. In this analysis, the issue of forecasting a large data set by using different model averaging approaches is addressed. In particular, Bayesian and frequentist model averaging methods are considered, including Bayesian model averaging (BMA), information theoretic model averaging (ITMA) and predictive likelihood model averaging (PLMA). The predictive performance of each scheme is compared with the most promising existing alternatives, namely benchmark AR model and the equal weighted model averaging (AV) scheme. An empirical application on Inflation forecasting for five countries using large data sets within the model averaging framework is applied. The average ARX model with weights constructed differently according to each model averaging scheme is compared with both the benchmark AR and the AV model. For the comparison of the accuracy of forecasts several performance indicators have been provided such as the Root Mean Square Error (RMSE), the Mean Absolute Error (MAE), the U-Theil’s Inequality Coefficient (U), Mean Square Forecast Error (MSFE) and the Relative Mean Square Forecast Error (RMSFE). Next, within the Granger causality framework through the Diebold & Mariano (DM) test and the Clark & McCracken (CM) test, whether the data-rich models represented by the three different model averaging schemes have made a statistically significant improvement relative to the benchmark forecasts has been tested. Critical values at 5% and at 10% have been calculated based on bootstrap approximation of the finite sample distribution of the DM and CM test statistics. The main outcome is that although the information theoretic model averaging scheme is a more powerful approach, the other two model averaging techniques can be regarded as useful alternatives.