Long-horizon stock return predictability test with a nonlinear nonparametric bootstrap method

A nonparametric bootstrap procedure with an LSTAR modeling of the valuation ratio is applied to the continuously compounded real stock return and the log of the price-dividend process. The empirical distribution of the test statistics shows that the evidence for a stock return predictability weakens when we take care of nonlinearity dynamics in the regressor. We split the sample into two regimes and implement the long-horizon predictability tests. Results show that the stock return is predictable in the stationary regime, while the test statistic under the null of unpredictability is insignificant in the non-stationary regime.