Abstract
This chapter presents a general simulation-based framework for estimation of agent-based models in economics and finance based on kernel methods. After discussing the distinguishing features between empirical estimation and calibration of economic models, the simulated maximum likelihood estimator is validated for utilization in agent-based econometrics.
As the main advantage, the method allows for estimation of nonlinear models for which the analytical representation of the objective function does not exist. We test the properties and performance of the estimator in combination with the seminal Brock and Hommes (J Econ Dyn Control 22:1235-1274, 1998) asset pricing model, where the dynamics are governed by switching of agents between trading strategies based on the discrete choice approach.
We also provide links to how the estimation method can be extended to multivariate macroeconomic optimization problems. Using simulation analysis, we show that the estimator consistently recovers the pseudo-true parameters with high estimation precision.
We further study the impact of agents' memory on the estimation performance and show that while memory generally deteriorates the precision, the main properties of the estimator remain unaffected.