A Note on the Robustness of Diebold and Li's Forecasting Results

The paper by Diebold and Li (2006) has become a benchmark in the yield curve forecasting literature, mostly owing to its excellent out-of-sample results. In this note we investigate the robustness of these outcomes in two different ways: (i) in terms of the arbitrary choices in their forecasting experimental design, and (ii) in updated samples up to 2015. In both cases, the fragility of their out-of-sample results became evident. In addition, we propose a novel decomposition method, which reveals that their positive results are unlikely to be caused by any intentionally built feature within their model. Overall, the evidence suggests that the forecasting ability of Dielbold and Li's model is questionable.