Why Dividend-Growth Regressions Don't Provide Stronger Evidence of Return Predictability
56 Pages Posted: 2 Apr 2019 Last revised: 19 Jan 2020
Date Written: August 31, 2019
When aggregate stock market returns are directly regressed on the dividend-price ratio, the statistical evidence supporting the hypothesis that returns are predictable is mixed. One response in the literature has been to employ indirect return predictability tests, which use an approximate present-value identity to transpose the null hypothesis that returns are unpredictable onto dividend-growth forecasting regressions. I first show analytically that indirect tests are identical to direct tests if approximation error is ignored, and that the presence of this error is likely to reduce the power of indirect tests. I then verify empirically that indirect tests do not provide stronger evidence of return predictability by bootstrapping asymptotically pivotal test statistics, which guarantee more accurate inferences in finite samples. Using the same approach, I also show that long-horizon return predictability tests produce results that are nearly identical, at all horizons, to tests that use one-period regressions. Finally, I make several methodological contributions that are more broadly applicable to predictive regressions.
Keywords: return predictability, empirical asset pricing, forecasting, maximum likelihood, bootstrap
JEL Classification: G12, G17, C22, C32
Suggested Citation: Suggested Citation