Our simulation results show that tests for long-horizon (i.e.. multi-year) abnormal security returns around firm-specific events are severely misspecified. The rejection frequencies using parametric tests sometimes exceed 30% when the significance level of the test is 5%. Our results are robust to many different abnormal-return models. Conclusions from long-horizon studies require extreme caution. Nonparametric and bootstrap tests are likely to reduce misspecification.

We find reliable evidence that both book-to-market (B M) and dividend yield track time-series variation in expected real stock returns over the period 1926 91 (in which B M is stronger) and the subperiod 1941–1991 (in which dividend yield is stronger). A Bayesian bootstrap procedure implies that an investor with prior belief 0.5 that expected returns on the equal-weighted index are never negative comes away from the full-period B/M evidence with posterior probability 0.08 for the hypothesis (0.14 with the impact of the 1933 outlier tempered). Although this raises doubts about market efficiency, the post-1940 evidence is consistent with expected returns always being positive.