Previous literature indicates that stock returns are predictable by several strongly autocorrelated forecasting variables, especially at longer horizons. It is suggested that this finding is spurious and follows from a neglected near unit root problem. Instead of the commonly used t test we propose a test that can be considered as a general test of whether the return can be predicted by any highly presistent variable. Using this test no predictablility is found for US stock return data from the period 1928-1996. Simulation experiments show that the standard t test clearly overrejects while our proposed test controls size much better.
