Choosing the weighting coefficients for estimating the term structure from sovereign bonds

Estimates of the term structure of interest rates depend heavily on the quality of the market data from which it is constructed. Estimated rates can be incorrect due to observation errors and omissions in the data. The usual way to deal with the heteroskedasticity of observation errors is by introducing weights in the fitting procedure. There is currently no consensus in the literature about the choice of such weights. We introduce a non-parametric bootstrap-based method of introducing observation errors drawn from the empirical distribution into the model data, which allows us to perform a comparison test of different weighting schemes without implicitly favoring one of the contesting models - a common design flaw in comparison studies. We use government bonds of several countries to show that realistic observation errors can distort the estimated yield curve. Moreover, we show that using different weights or other modifications of accounting for observation errors in bond price data doesn't always improve the term structure estimates, and often only worsens the situation. Based on our comparison, we advise to either use equal weights or weights proportional to the inverse duration in practical applications.