Misspecification and Identification Robust Predictability Testing with Panel Data

We propose a new method-of-moments based toolkit to test for (return) predictability in the panel data setting with general predictors and cross-section dependence. Our proposed toolkit is based on overidentified method-of-moments procedures that accommodate settings with both strong and weak instruments. We argue that for the panel setting, any proposed procedure should be able to generate empty confidence intervals when several potential predictors are considered, thereby indicating misspecification of the postulated model. We prove the validity of the proposed procedure using asymptotic theory for potentially non-stationary large dimensional panel data. The proposed procedure is illustrated using both historical and simulated data. Our empirical results mostly confirm previously documented evidence for predictability of excess stock returns by short-term interest rates and term spreads for developed countries.