Macroeconomic models with endogenous learning

The behavior of the macroeconomy and monetary policy is heavily influenced by expectations. Recent research has explored how minor changes in expectation formation can change the stability properties of a model. One common way to alter expectation formation involves agents' use of econometrics to form forecasting equations. Agents update their forecasts based on new information that arises as the economy progresses through time. In this way agents "learn" about the economy.Previous learning literature mostly focuses on agents using a fixed data size or increasing the amount of data they use. My research explores how agents might endogenously change the amount of data they use to update their forecast equations.My first chapter explores how an established endogenous learning algorithm, proposed by Marcet and Nicolini, may influence monetary policy decisions. Under rational expectations (RE) determinacy serves as the main criterion for favoring a model or monetary policy rule. A determinant model need not result in stability under an alternative expectation formation process called learning. Researchers appeal to stability under learning as a criterion for monetary policy rule selection.This chapter provides a cautionary tale for policy makers and reinforces the importance of the role of expectations. Simulations appear stable for a prolonged interval of time but may suddenly deviate from the RE solution. This exotic behavior exhibits significantly higher volatility relative to RE yet over long simulations remains true to the RE equilibrium.In the second chapter I address the effectiveness of endogenous gain learning algorithms in the presence of occasional structural breaks. Marcet and Nicolini's algorithm relies on agents reacting to forecast errors. I propose an alternative, which relies on agents using statistical information.The third chapter uses standard macroeconomic data to find out whether a model that has non-rational expectations can outperform RE. I answer this question affirmatively and explore what learning means to the economy. In addition, I conduct a Monte Carlo exercise to investigate whether a simple learning model does, empirically, imbed an RE model. While theoretically a very small constant gain implies RE, empirically learning creates bias in coefficient estimates.