This dissertation includes three chapters. The first two chapters are co-authored with Naoki Takayama. The first chapter presents a model of business cycles driven by shocks to agents' beliefs about economic fundamentals. Agents are hit both by common and idiosyncratic shocks. Common shocks act as confidence shocks, which cause economy-wide optimism or pessimism and consequently, aggregate fluctuations in real variables. Idiosyncratic shocks generate dispersed information, which prevents agents from perfectly inferring the state of the economy. Crucially, asymmetric information induces the infinite regress problem, that is, agents need to forecast the forecasts of others. We develop a method that can solve the infinite regress problem without approximation. Even though agents face a complicated learning problem, the equilibrium policy can be represented by a small number of state variables. Theoretically, we prove that the persistence of aggregate output is increasing in the degree of information frictions and strategic complementarity, and there is a hump-shaped relationship between the variance of output and the variance of the confidence shock. Quantitatively, our model with confidence shocks can match a number of the key business cycle moments. The second chapter develops a general method of solving rational expectations models with higher order beliefs. Higher order beliefs are crucial in an environment with dispersed information and strategic complementarity, and the equilibrium policy depends on infinite higher order beliefs. It is generally believed that solving this type of equilibrium policy requires an infinite number of state variables (Townsend, 1983). This paper proves that the equilibrium policy rule can always be represented by a finite number of state variables if the signals observed by agents follow an ARMA process, in which case we obtain a general solution formula. We also prove that when the signals contain endogenous variables, a finite-state-variable representation of the equilibrium may not exist. For this case, we develop a tractable algorithm that can approximate the solution arbitrarily well. The key innovation in our method is to use the factorization identity and Wiener filter to solve signal extraction problems conditional on infinite observables. This method can be used in a wide range of applications. We demonstrate its strong practicability by solving several classical models featuring higher order beliefs. The third chapter is co-authored with Jose-Victor Rios-Rull. We build a variation of the neoclassical growth model in which both wealth shocks (in the sense of wealth destruction) and financial shocks to households generate recessions. The model features three mild departures from the standard model: (1) adjustment costs make it difficult to expand the tradable goods sector by reallocating factors of production from nontradables to tradables; (2) there is a mild form of labor market frictions (Nash bargaining wage setting with Mortensen-Pissarides labor markets); (3) goods markets for nontradables require active search from households wherein increases in consumption expenditures increase measured productivity. These departures provide a novel quantitative theory to explain recessions like those in southern Europe without relying on technology shocks.