To implement the relation in expression (6) in an empirical setting, it is still necessary to specify how observable data are used to form expec
tations about future abnormal earnings. This becomes the key step in the research design and, ultimately, the step that distinguishes one study from another. However, this step can be accomplished with assumptions much more plausible than those invoked in studies based on the dividend dis count formula. For example, define financial assets as those that represent zero net present value projects, so that, only operating assets are expect ed to generate abnormal returns. Then, assume abnormal earnings exhib it reversion over time to a (possibly nonzero) mean. That is a sensible assumption in the face of competitive pressures and one that is consistent with evidence in Penman (1991) and Bernard (1994). More specifically,
assume abnormal earnings follow a simple autoregressive process. This is
restrictive, but good enough to capture the first-order effects of the behav
ior of abnormal earnings. These assumptions, when combined with the
relation in expression (6), produce the following valuation model, which also appears in Feltham and Ohlson (1995):