On the Choice of Optimization Routine in Estimation of Parsimonious Term Structure Models: Results from the Svensson Model

Abstract

Objective function in term structure estimation with price errors is not only nonlinear but also non-convex in parameters. This makes the final results sensitive to both the choice of the optimization routine as well as to the starting guess. This study looks at the impact of the choice of the optimization routine to final parameter estimates for the Svensson model. While results are expected to differ numerically across routines, what is of interest is the economic impact. Using eleven different routines over a range of starting parameter values, it is found while there is significant variation in the final objective function value across routines, for the most part, implied short-rates and long-rates have low standard deviation. Also, while grid-search seems unavoidable, popular quasi- Newton methods allowing for linear constraints seem quite adequate for the task at hand.