At all scales, economies undergo substantial fluctuations instead of changing smoothly over time. There exist two alternative views about these fluctuations. According to the first view, fluctuations are caused by exogenous noise. Absent noise, economies would return to stable stationary states. The second view argues instead that such stationary states are unstable, and economies endogenously fluctuate following limit cycles or chaotic attractors. Endogenous fluctuations would occur even without noise. It is not yet clear which of these two interpretations most closely describes economic reality, but it has fundamental intellectual and policy implications. This thesis presents theoretical arguments in favor of the second view, in both micro- and macro-economic systems. In the first part, we study the smallest scale at which these two alternative views can be compared. This is a normal form game repeatedly played by two players who learn from their mistakes and form beliefs about their opponents' future play. Under the exogenous view, players would converge to a Nash equilibrium, and strategic fluctuations would only be caused by behavioral noise. The endogenous view argues instead that strategic dynamics may fail to converge to equilibrium. Here, we completely characterize convergence in games with two players and two actions per player under a learning rule known as Experience-Weighted Attraction. We also introduce a theory that we call "best-reply structure" to approximately characterize convergence in games with two players and an arbitrary number of actions. Our results show that non-convergence and so endogenous fluctuations are more common in games than what might be thought. The second part of this thesis studies the largest scale of economic fluctuations, namely business cycles in macroeconomic systems. The two views propose different mechanisms to explain correlation, or comovement, of business cycles across economic sectors and across countries. The first view argues that comovement is caused by the propagation of shocks exogenous to the economy, such as political decisions and natural catastrophes. According to the second view, comovement is instead at least partially caused by the synchronization of non-linear dynamics that describe the business cycle. Here, we introduce a demand-driven model of endogenous business cycles in which synchronization occurs through input-output linkages. We develop a theory that mathematically illustrates the interplay between synchronizing non-linear dynamics, shock propagation and network structure. While exogenous business cycle models have difficulty to generate a level of comovement that is as high as in the data, synchronization of endogenous business cycles generates much stronger comovement, potentially solving long-standing puzzles in macroeconomics.