Testing the Predictive Power of the Term Structure Without Data Snooping Bias

Abstract

It is well documented that the term structure of interest rates has predictive power for real economic growth. Although various models have been constructed to test the predictive power, there is no consensus on which model captures most information about future states of the economy. Consideration to data snooping arises when the data set is reused to test many term spread models for the predictive power. Applying the stepwise superior predictive ability test of Hsu, Hsu, and Kuan (2009) to examine 900 alternative models of interest rates and term spreads, we find that there are some models statistically superior to the benchmark, the univariate autoregression model, in predicting the U.S. real growth, and these models contain both a short-term rate and a term spread. In particular, in predicting the annual growth rates, the best model consists of the short-term rate of 3-month Treasury bill and the term spread between this short rate and the yield of 5-year Treasury note.