The ability of the yield curve in forecasting output gaps

Several studies (Litterman, 1986; Estrella et al., 2003; Wright, 2006, Christiansen, 2012) provide evidence of the forecasting ability of the yield curve on GDP. The yield curve is a graphical representation of the relationship between the time-to-maturity and the yield-to-maturity of bonds at a specific point in time. Short-term rates reflect the currently implemented monetary policy while long-term rates reveal the market’s expectations on future economic conditions. The great majority of the studies related to this topic, focus on the slope of the yield curve, i.e. the spread between interest rates of different maturities. A positive slope of the yield curve indicates normal economic activity and an inverted (negative slope) yield curve is considered as an indicator of an upcoming economic slowdown. In this paper we attempt to forecast out-of-sample the deviations of real GDP below its long-run trend (output gaps). The innovation of our study is that a) instead of using only data on the slope of the yield curve we also incorporate information on the arc of the yield curve and b) the methodology we use is both a classic probit model and also we employ from the field of Machine Learning a Support Vector Machines (SVM) framework. The SVM model paired with Kernel Methods is considered the state-of-the art in supervised classification. In doing so, we start with an exhaustive testing of all possible pairs of interest rates. Next, we employ a similar exhaustive testing strategy of all possible triplets of interest rates in terms of one short-term, one medium-term and one long-term interest rate in order to take into account the information provided by the arc of the yield curve. Finally, we test a model where all eight interest rates are employed simultaneously. In an effort to examine whether additional macroeconomic variables can improve the forecasting ability of our models, we include twenty nine additional variables to the models with the best performance in both the pairs and triples combinations. Our training data span the period from 1976Q3 to 2006Q1 and the out-of-sample data cover the period from 2006Q2 to 2013Q3. The optimal overall models reach an out-of-sample accuracy of up to 90%. The forecasting performance of our models corroborates the existing evidence that the yield curve can be a useful predictive tool of economic activity.