Term structure forecasting. Does a good fit imply reasonable simulation results?

In this paper I question whether term structure models of interest rates used for pricing derivative instruments are suited to use in a simulation based context like an asset liability management (alm) study. I consider the Vasicek, Cox-Ingersoll-Ross (cir) and Nelson-Siegel model. First, I discuss the models and their merits and drawbacks. Second, I estimate the parameters of all three models. Estimates are based on monthly Canadian zero-coupon yields from 1986 up until and including 2009. I estimate the Vasicek model twice: once using time series of the two-year yields and once using cross-sectional data (i.e., the different maturity times). I calibrate the cir model using time series of the two-year yields and the Nelson-Siegel model using cross-sectional data. Third, I simulate the models. I simulate the Vasicek and cir short rate processes by their discrete representations, whereas I simulate the factors of the NelsonSiegel with a first-order vector autoregressive (var(1)) model. I judge the term structure models’ empirical fit as well as dynamics in the alm study and find that the Nelson-Siegel model outperforms the Vasicek and cir models.