This paper develops a theory of expectations-driven business cycles based on learning. Agents have incomplete knowledge about how market prices are determined and shifts in expectations of future prices affect dynamics. In a real business cycle model, the theoretical framework amplifies and propagates technology shocks. Improved correspondence with data arises from dynamics in beliefs being themselves persistent and because they generate strong intertemporal substitution effects in consumption and leisure. Output volatility is comparable with a rational expectations analysis with a standard deviation of technology shock that is 20 percent smaller, and has substantially more volatility in investment and hours. Persistence in these series is captured, unlike in standard models. Inherited from real business cycle theory, the benchmark model suffers a comovement problem between consumption, hours, output and investment. An augmented model that is consistent with expectations-driven business cycles, in the sense of Beaudry and Portier (2006), resolves these counterfactual predictions.