Tests of Return Predictability: A Comparative Study

Abstract

Evidence of return predictability based on dividend yields remains inconclusive. Predictive regressions are subject to subtle econometric issues rendering standard inference inappropriate. As a consequence, there now exists a broad range of statistical tests. We review the tests from Stambaugh (1999), Lewellen (2004), Campbell and Yogo (2006), and Cochrane (2007) and complement our analysis with standard maximum likelihood as well as restricted maximum likelihood estimation as discussed by Chen and Deo (2009). When applied to two samples from the literature, we find that p-values range from almost 0% to more than 20%. We reconcile the conflicting evidence and identify two main sources: First, we attribute part of it to the two different sample periods. Second, we urge to go beyond the p-value and find that some of the tests are severely size-distorted, whereas others lack power. To demonstrate this, we use an extensive simulation study to assess the size (falsely rejecting a true null) and power (correctly rejecting a false null) of the selected tests.