The yield curve and business cycle forecasting in a support vector machines framework
Periklis Gogas, Theophilos Papadimitriou, Efthymia Chrysanthidoy
The forecasting power of the yield curve is established by several studies; see, for example, among others, Neal and Morley (1997) and Ang et al. (2006). Since short-term rates are instruments of the implemented monetary policy and the long-term rates reflect the market expectations on future economic conditions, the two contain significant information for the policy makers and market participants. A positive slope on the yield curve during an output gap indicates a forthcoming economic upturn and a downward slopping yield curve is an indicator of an upcoming recession. The policy importance is apparent, as an above trend GDP puts inflationary pressure to the economy and an output gap is associated with an unemployment rate that exceeds the NAIRU. In this paper, we attempt to forecast out-of-sample the deviations of real GDP from its long-run trend using the information provided by the short and long term interest rates. In this sense, we are not focusing into forecasting recessionary events in the definition of the NBER which are relatively rare, but deviations of real seasonally adjusted GDP from its long-term trend (the potential GDP). To extract the business cycles we use quarterly U.S. data for the period 1976Q3 to 2011Q4 on real GDP. All possible pairs of short and long term interest rates from a set of three short-term Treasury bill rates and four long-term rates are used as explanatory variables. The main contribution of our approach is that in the effort to out-of-sample forecast the business cycle via the yield curve we employ a Machine Learning approach using a Support Vector Machines (SVM) framework. The SVM model paired with Kernel Methods is considered the state-of-the art in supervised classification. The basic idea is that if the dataset is not linearly separable in the initial data space, we project the dataset in the feature space (a rich space of higher dimensionality), where the linear classification is possible. The larger part of our data is used for training (locating the support vectors in the dataset), and the rest is used for testing the performance of the trained model in out-of-sample forecasting. The results show that the best forecasting accuracy is achieved with the combination of a 6 month and a 5 year interest rate at 77.27% total accuracy and 87.5% accuracy in recessionary gap forecasting.