(Un-)Predictability in Bond Returns: Benchmarking Deep Learning Architectures out of Sample

Abstract

We conduct an out of sample study of predictors of excess returns on US Treasuries. We add to
the group of considered predictors by including Artificial Neural Networks (ANN). We find that the best
predictor of excess returns is the historical mean and thus fail to reject the Expectations Hypothesis out of
sample. ANNs converge on the same signal recovered by a linear combination of forward rates, and nonlinear
combinations of yields do not have predictive power in excess of this linear combination. Lagged
forward rates and real time macroeconomic data do not either, and so we do not find a hidden factor in
the yield curve.