Selecting a Bond‐Pricing Model for Trading: Benchmarking, Pooling, and Other Issues

Does one make money trading on the deviations between observed bond prices and values proposed by bond‐pricing models? We extend Sercu and Wu's (1997) work to more models and more data, but we especially refine the methodology. In particular, we provide a normal‐return benchmark that markedly improves upon the Sercu‐Wu ones in terms of noisiness and bias, and we demonstrate that model errors contribute more to the variance of residuals—actual minus fitted prices—than pricing errors made by the market. Trading on the basis of deemed mispricing is profitable indeed no matter what model one uses. But there is remarkably little difference across models, at least when one re‐estimates and trades daily; and with pooling and/or longer holding periods the results seem to be all over the place, without any relation to various measures of fit in the estimation stage. We also derive and implement an estimator of how much of the typical deviation consists of mispricing and how much is model mis‐estimation or mis‐specification. Lastly, we find that pooled time‐series and cross‐sectional estimation, as applied by e.g., De Munnik and Schotman (1994), does help in stabilizing the parameter, but; hardly improves the trader's profits.