Predicting By Learning: An Adaptive Rationale

The paper proposes a partial-adjustment mechanism for the learning process of economic agents and justify the use of past information in predicting stock returns from four different perspectives. By making a pair of mild assumptions about how rational investors learn about the fundamental values of returns and dividend yield over time, I show that for one-step-ahead forecast a stable and significant improvement in terms of short-horizon R2 can be achieved by recasting the classical single-equation predictive regression in a differenced form and incorporating information from the recent past. For longer horizons, the relationship reduces to the standard form.