The Information Content in Bond Model Residuals: An Empirical Study on the Belgian Bond Market

Abstract

We estimate both one-factor Vasicek and CIR bond pricing models and the cubic spline model using Belgian government bond data on each trading day over 1991-1992. We observe humped zero-yield curves with the two economic models but not the spline model during the period. The CIR model scores better than the Vasicek model in terms of goodness-of-fit (an average Std.Err of 12.4 bp versus 13.5 bp), but the curve-fitting spline model easily beats the two economic models with much smaller model residuals (8.0 bp). To make economic sense, we further test whether bond model residuals contain useful information for forecasting bond holding period returns. First, we find that the abnormal bond holding period returns, which are measured by three alternative benchmarks, are negatively related to once- and twice-lagged bond pricing errors. Second, we form trading strategies that exploit the mispricing based on alternative bond models. We find that buying underpriced bonds and selling overpriced bonds generates significant abnormal returns even when the trade is delayed by up to five days after observing the mispricing. Nevertheless, the magnitude of abnormal returns decreases dramatically with the delayed days. In light of that bond pricing errors contain both information and noise, further tests reveal that large model residuals are more likely caused by mis-specification and/or -estimation than small or medium-sized residuals. More profound, the curve-fitting model seem to overfit the data because it shows the least ability to detect mispricing.

Note: This is a description of the paper and is not the actual abstract.

JEL Classification: G12, G14