Macroeconomic Forecasting with Independent Component Analysis

This paper considers a factor model in which independent component analysis (ICA) is employed to construct common factors out of a large number of macroeconomic time series. The ICA has been regarded as a better method to separate unobserved sources that are statistically independent to each other. Two algorithms are employed to compute the independent factors. The first algorithm takes into account the kurtosis feature contained in the sample. The second algorithm accommodates the time dependence structure in the time series data. A straightforward forecasting model using the independent factors is then compared with the forecasting models using the principal components in Stock and Watson (2002). The results of this research can help us to gain more knowledge about the underlying economic sources and their impacts on the aggregate variables. The empirical findings suggest that the independent component method is a powerful method of macroeconomic data compression. Whether the ICA method is superior over the principal component method in forecasting the U.S. real output and inflation variables is however inconclusive