Recently economists have more and more focussed on scenarios in which agents' views of the world may be erroneous. These notes introduce the concept of perfect forecasting rules which provide best least-squares predictions along the evolution of an economic system. The framework for nonparametric adaptive learning schemes is developed and it is argued that plausible learning schemes should aim at estimating a perfect forecasting rule taking into account the correct feedback structure of an economy. A link is provided between the traditional rational-expectations view and recent behavioristic approaches.