Kernel-based Multistep-ahead Predictions of the US Short-term Interest Rate

This paper presents a comparison of prediction performances of threekernel-based nonparametric methods applied to the U.S. weekly T-bill rate.Predictions are generated through the rolling approachfor the out-of-sample period 1989-1993. We compare the multistep-aheadprediction performance of the conditional mean, the conditionalmedian, and the conditional mode with the performanceof the benchmark random walk model. Using four predictionevaluation criteria, it is shown that two of the threepredictors are superior -- or at least equal --to the random walk at prediction horizons 1 - 5.In addition, by combining two of the three predictors, a significantimprovement in prediction accuracy is obtainedat all prediction horizons. Also the combined predictions resultin substantial improvement at predicting the direction of change.Further, we propose two prediction intervals based on the estimatednonparametric conditional distributionfunction. These intervals are useful when the predictivedistribution underlying the time series process is asymmetricor multi-modal. Finally, we assess the choice of the bandwidthin the kernel-based prediction methods through a recently proposedmethod for evaluating the estimated prediction densities.