Kernel-based multistep-ahead predictions of the US short-term interest rate

This paper presents a comparison of prediction performances of three kernel-based non-parametric methods applied to the US weekly T-bill rate. Predictions are generated through the rolling approach for the out-of-sample period 1989-1993. The multistep-ahead prediction performance of the three predictors is compared with two benchmarks: a random walk (RW) and an AR model. To this end five prediction evaluation criteria are considered including sign accuracy. Further, two prediction intervals are proposed based on the estimation nonparametric conditional distribution function. Finally, the choice of the bandwidth in the kernel-based prediction methods is assessed through two methods for evaluating the estimated prediction densities.