We reconsider the derivation of the traditional capital asset pricing model (CAPM) in the discrete time setting for a portfolio of one riskless asset and many risky assets. In contrast to the standard setting, it is assumed that agents are heterogeneous in their conditional means and covariances of the risky returns, and that their beliefs about future returns - based on statistical properties of past returns - induce expectations feedback. A Walrasian auctioneer scenario is used for the determination of the market clearing price. In this framework we first construct a consensus belief to represent the aggregate market beliefs about means and variances/covariances of returns, and derive a heterogeneous CAPM which relates aggregate excess return on risky assets with aggregate excess return on the market portfolio via aggregate beta coefficients. We then use this result to establish a dynamic {\em market fraction\/} model in which agents are grouped according to their beliefs. In particular, we focus on three classical heterogeneous agents types - fundamentalists, trend followers and noise traders - and investigate how some of the key agent characteristics affect the time varying behaviour of market returns and beta coefficients
