The learning of optimal discrete investment rules is analysed and related to the problem of forecasting financial returns. The aim is twofold: to characterize some 'good' learning methods for agents using investment rules of this form and to explain why many observed investment rules such as technical trading rules are discrete. A consistent estimator for discrete investment rules is used and it is shown, using simulations, that direct estimation of investment rules is preferable to the estimation of forecasting models to be used in such rules. This model and the associated results indicate there are a number of reasons why it may be easier to learn a good discrete investment rule than to learn a continuous rule; this provides a partial explanation of why discrete investment rules are used so widely.