Approximate Bayesian inference for agent-based models in economics: a case study

Abstract

Abstract Estimation of agent-based models in economics and finance confronts researchers with a number of challenges. Typically, the complex structures of such models do not allow to derive closed-form likelihood functions so that either numerical approximations to the likelihood or moment-based estimators have to be used for parameter inference. However, all these approaches suffer from extremely high computational demands as they typically work with simulations (of the agent-based model) embedded in (Monte Carlo) simulations conducted for the purpose of parameter identification. One approach that is very generally applicable and that has the potential of alleviating the computational burden is Approximate Bayesian Computation (ABC). While popular in other areas of agent-based modelling, it seems not to have been used so far in economics and finance. This paper provides an introduction to this methodology and demonstrates its potential with the example of a well-studied model of speculative dynamics. As it turns out, ABC appears to make more efficient use of moment-based information than frequentist SMM (Simulated Method of Moments), and it can be used for sample sizes of an order far beyond the reach of numerical likelihood methods.