The Use of Predictive Regressions at Alternative Horizons in Finance and Economics

When a k period future return is regressed on a current variable such as the log dividend yield, the marginal significance level of the t-test that the return is unpredictable typically increases over some range of future return horizons, k. Local asymptotic power analysis shows that the power of the long-horizon predictive regression test dominates that of the short-horizon test over a nontrivial region of the admissible parameter space. In practice, small sample OLS bias, which differs under the null and the alternative, can distort the size and reduce the power gains of long-horizon tests. To overcome these problems, we suggest a moving block recursive Jackknife estimator of the predictive regression slope coefficient and test statistics that is appropriate under both the null and the alternative. The methods are applied to testing whether future stock returns are predictable. Consistent evidence in favor of return predictability shows up at the 5 year horizon