Testing for Return Predictability with Co-Moving Predictors of Unknown Form

Testing for constant expected returns and forecasting future returns necessitate the information beyond a single predictor. We consider the predictive regression model with multiple predictors which are potentially strongly persistent and cointegrated. Instrumental variables based tests for joint predictability are shown to be robust to cointegration among predictors of unknown form. We also study their behavior under general dynamic and cross-sectional correlation in innovations. Empirical applications to two datasets in the U.S. aggregate stock market show that the results substantially widen the choice set of potential predictors for valid inference of return predictability