Learning and monetary policy

This thesis studies implications of different learning mechanisms in various monetary environments. In Chapter 2, adaptive step-size algorithm (Kushner, Yin 2003) is used to model time-varying learning and is studied in the environment of Marcet, Nicolini (2003). The resulting model gives qualitatively similar results to MN and performs quantitatively somewhat better based on the criterion of mean squared error. This model generates increasing gain during hyperinflations that matches findings in Cagan (1956), Khan (1977). An agent behaves cautiously when faced with sudden changes in policy, and is able to recognize a change in regime after acquiring sufficient information. Chapter 3 analyzes the effects of social learning in New Keynesian model described in Woodford (2003). The question is whether the economy will converge to a rational expectations equilibrium under this more realistic learning dynamics. A key result from the literature in this version of the model is that the Taylor Principle governs both the uniqueness and the expectational stability of the rational expectations equilibrium when all agents learn homogeneously using recursive algorithms. The finding is that the Taylor Principle is not necessary for convergence in a social learning context. This paper also contributes to the use of genetic algorithm learning in stochastic environments. Chapter 4 studies cheap talk announcement in an agent-based dynamic extension of Kydland-Prescott model. The government choose inflation announcement and actual inflation and updates its decisions using a model of individual, evolutionary learning (Arifovic, Ledyard 2004). Private agents use naïve and more sophisticated inflati on forecasts and switch between them based on their payoffs. Agents and government can coordinate on Pareto-superior outcomes with positive fraction of naive agents. However, the economy does not stay there. It exhibits recurrent fluctuations in announced and actual inflation as government repeatedly builds up and exploits the proportion of believers. Outcomes with higher fraction of naive forecasters have higher average welfare of agents and government. When cost of sophisticated forecast goes up, the proportion of naive believers goes up. When nonbelievers update slower, naive believers are more likely to disappear. Therefore, quick and accurate sophisticated forecasters ensure positive number of naive agents