Optimal monetary policy using reinforcement learning

This paper introduces a reinforcement learning based approach to compute optimal interest rate reaction functions in terms of fulfilling inflation and output gap targets. The method is generally flexible enough to incorporate restrictions like the zero lower bound, nonlinear economy structures or asymmetric preferences. We use quarterly U.S. data from1987:Q3-2007:Q2 to estimate (nonlinear) model transition equations, train optimal policies and perform counterfactual analyses to evaluate them, assuming that the transition equations remain unchanged. All of our resulting policy rules outperform other common rules as well as the actual federal funds rate. Given a neural network representation of the economy, our optimized nonlinear policy rules reduce the central bank's loss by over43 %. A DSGE model comparison exercise further indicates robustness of the optimized rules.