Efficient Tests of Stock Return Predictability

Empirical studies have suggested that stock returns can be predicted by financial variables such as the dividend-price ratio. However, these studies typically ignore the high persistence of predictor variables, which can make first-order asymptotics a poor approximation in finite samples. Using a more accurate asymptotic approximation, we propose two methods to deal with the persistence problem. First, we develop a pretest that determines when the conventional t-test for predictability is misleading. Second, we develop a new test of predictability that results in correct inference regardless of the degree of persistence and is efficient compared to existing methods. Applying our methods to US data, we find that the dividend-price ratio and the smoothed earningsprice ratio are sufficiently persistent for conventional inference to be highly misleading. However, we find some evidence for predictability using our test, particularly with the earnings-price ratio. We also find evidence for predictability with the short-term interest rate and the long-short yield spread, for which the conventional t-test leads to correct inference. ∗Department of Economics, Harvard University and NBER (john [email protected]); and Department of Economics, Harvard University ([email protected]). First draft: June 26, 2002. For helpful comments we thank Andrew Ang, Markku Lanne, Marcelo Moreira, Robert Shiller, Mark Watson, and participants at the 2002 Econometric Society Australasian Meeting.