Subsampling-Based Tests of Stock-Return Predictability

We develop subsampling-based tests of stock-return predictability and apply them to U.S. data. These tests allow for multiple predictor variables with local-to-unit roots. By contrast, previous methods that model the predictor variables as nearly integrated are only applicable to univariate predictive regressions. Simulation results demonstrate that our subsampling-based tests have desirable size and power properties. Using stock-market valuation ratios and the risk-free rate as predictors, our univariate tests show that the evidence of predictability is more concentrated in the 1926-1994 subperiod. In bivariate tests, we find support for predictability in the full sample period 1926-2004 and the 1952-2004 subperiod as well. For the subperiod 1952-2004, we also consider a number of consumption-based variables as predictors for stock returns and find that they tend to perform better than the dividend-price ratio. Among the variables we consider, the predictive power of the consumption-wealth ratio proposed by Lettau and Ludvigson (2001a, 2001b) seems to be the most robust. Among variables based on habit persistence, Campbell and Cochrane's (1999) nonlinear specication tends to outperform a more traditional, linear specification.