Forecasting of yield curves using local state space reconstruction

Abstract

We examine models of yield curves through chaotic dynamical systems whose dynamics can be unfolded using non-linear embeddings in higher dimensions. We refine recent techniques used in the state space reconstruction of spatially extended time series in order to forecast the dynamics of yield curves. We use daily LIBOR GBP data (January 2007–June 2008) in order to perform forecasts over a one-month horizon. Our method outperforms random walk and other benchmark models on the basis of mean square forecast error criteria.