We adapt a genetic-based learning classifier system to a forecast evaluation exercise by making its key parameters endogenous and taking into account the need of convergence of the learning algorithm, an issue usually neglected in the literature. Doing so, we find it hard for the algorithm to beat simpler ones based on recursive regressions and on the random walk in forecasting stock returns. We then argue that our results cast doubts on the plausibility of using learning classifier systems to represent agents process of expectations formation, an approach commonly found into the agent-based computational finance literature.