Yeild curve forecasting with a large macroeconomic data set using two-step factorization

Abstract

This study emphasizes on improving the performance of term structure forecasting with an appropriate methodology of extracting common factors from large macroeconomic time series. This empirical study has been conducted for the US and German bond markets. We investigate the yield curve forecast performance under the Principal Component Analysis (PCA) and the Two-Step Factorization techniques for extracting information. We assume that the dynamics of the short rate follow a Factor-Augmented Vector Autoregression (FAVAR) model and that the term structure implies no-arbitrage condition. The PCA technique and the Two-Step Factorization will be used to extract common factors from the same macroeconomic data set. Then, the yield curve forecast performance under the different approaches will be compared. The finding shows that the common factors extracted from the Two-Step Factorization outperform those extracted from the PCA technique and a random walk model in forecasting the intermediate yields in particular at the intermediate and long forecast horizons. By extracting factors only from macro variables that can explain short rates, the Two-Step Factorization method leads to better forecast performance for long forecast horizons. However, the performance for intermediate forecast horizons becomes worse. On the other hand, using factors extracted from macro variables that can explain a long term rate instead of the short rate leads to better forecast performance for the long term yields but it lowers the performance for the short term yields for all forecast horizons. This implies that the constraints put to the extracting method for selecting factors may help eliminate some noises, but at the same time they may also eliminate some of the information and hence lower the forecasting performance.