Predicting the Yield Curve Using Forecast Combinations

We examine the statistical accuracy and economic value of modelling and forecasting the term structure of interest rates using forecast combinations. We adopt five alternative methods to combine point forecasts from several univariate and multivariate autoregressive specifications, as well as from factor models for the yield curve such as the dynamic versions of the Nelson-Siegel and Svensson specifications. Moreover, we conduct a detailed performance evaluation based not only on statistical measures of forecast accuracy, but also an economic criteria like Sharpe ratios of optimal mean-variance fixed income portfolios constructed based upon forecasts from individual models and their alternative combinations. Our empirical application based on a large panel of Brazilian interest rate future contracts with different maturities shows that combined forecasts consistently outperform individual models in several instances, specially when economic criteria are taken into account