Determinants of Bond Risk Premia : A Machine-Learning-Based Resolution of the Spanning Controversy

There is no consensus on whether macroeconomic fundamentals have any predictive power for bond risk premia, either unconditionally or conditionally over bond yields. Using Adaptive Group LASSO, a machine learning algorithm, we are able to construct a new, parsimonious macro variable that is weakly correlated with bond yields yet has significant predictive power for excess bond returns. Results from finite-sample analysis based on both regressions and macro-finance term-structure models (MTSMs) reject the hypothesis that this macro factor is spanned by bond yields. Unspanned MTSMs outperform spanned models in both capturing the dynamics of forward term premia and forecasting future yields