Regressions for predicting long-term stock returns often use moving averages of earnings as the earnings trend. We show that the earnings trend can be directly estimated using unobserved components models. The estimated trends improve the fit of predictive regressions.