Dynamic Yield Curve Forecasting with Regime-Switching Nelson-Siegel-Svensson Models

Abstract

The United States Treasury yield curve, a prominent economic indicator with the inherent ability to convey investors' expectations, is well-understood and has been modeled extensively in academic literature. Yield curve forecasting is an essential activity performed by asset managers and banks in the private sector, as well as central banks tasked with making policy decisions. Improving forecasting technology is essential for policymakers interested in stabilizing economies as well as institutional or sovereign asset managers looking to responsibly invest for long-term time horizons. The PCA-based Nelson-Siegel model of 1987 uncovered the \textit{Level-Twist-Curvature} framework, a widely studied phenomenon used for yield curve modeling and forecasting. Several years later, the extended Nelson-Siegel-Svensson model was introduced, adding complexity to this new branch of analytics.

This paper will expand upon current research by introducing a Markov chain-based regime switching routine in an attempt to provide increased econometric flexibility in term structure forecasts. First, a dynamic version of the Nelson-Siegel model isvimplemented, allowing for robust yield curve forecasts under the assumption that yield curve factors are time-dependent and follow an AR(1) process. A factor model, including global equities indices and econometric indicators, are be introduced into the analysis to capture the joint behavior of the yield curve and the macroeconomy. A Hidden Markov Model identifies broad "regimes" under which the yield curve is expected to behave predictably. A Markov switching routine is used to forecast yield curve factors following an AR(1) process with regime-conditional means and heteroscedasticity in the factor dynamics, with the ultimate goal of improving yield curve forecasting techniques. This analysis concludes that the inclusion of yield curves in this analysis is particularly useful, improving forecasting ability of the Nelson-Siegel-Svensson model by over 20%.