We analyze the classical asset pricing model assuming non fully rational agents. Agents forecast future prices cum dividend through an adaptive learning rule. This assumption provides an explanation of some anomalies encountered in the empirical analysis of asset prices under full rationality: Returns are serially correlated (positively over a short horizon and negatively over a longer horizon) and the dividend yield predicts future returns (positive correlation). Considering the continuous time limit process, the same regularities are established analytically for price increments