Testing term structure estimation methods

Abstract

This paper tests and compares five distinct methods for estimating the term structure. The Unsmoothed Fama-Bliss method is an iterative method by which the discount rate function is built up by computing the forward rate necessary to price successively longer maturity bonds. The Smoothed Fama-Bliss "smooths out" these discount rates by fitting an approximating function to the "unsmoothed" rates. The McCulloch method fits a cubic spline to the discount function using an implicit smoothness penalty, while the Fisher-Nychka-Zervos method fits a cubic spline to the forward rate function and makes the smoothness penalty explicit. Lastly, the Extended Nelson-Siegel method, introduced in this paper, fits an exponential approximation of the discount rate function directly to bond prices. ; The tests demonstrate the dangers of in-sample goodness-of-fit as the sole criterion for judging term structure estimation methods. A series of residual analysis tests are introduced to detect misspecification of the underlying pricing equation relating the term structure to bond prices. These tests establish the presence of unspecified, but nonetheless systematic, omitted factors in the prices of long maturity notes and bonds. ; Comparisons of the five term structure estimation methods using these parametric and non-parametric tests finds that the Unsmoothed Fama-Bliss does best overall. Differences with some alternatives may not be economically significant given the much larger number of parameters this method estimates. Users seeking a parsimonious representation of the term structure should consider either the Smoothed Fama-Bliss or the Extended Nelson-Siegel methods. One method was found to be unacceptable. The Fisher-Nychka-Zervos cubic spline method performs poorly relative to the alternatives, both in- and out-of-sample. Furthermore, it systematically misprices short maturity issues and suffers from instability in the estimated term structure.

Cite this work

No public copy is stored for this record, so only the metadata is cited.

Stored only when both papers are in the corpus — 203 citations exist upstream.