Bayesian estimation of a large-scale macroeconomic policy agent-based model

Empirical parameter estimation of large scale agent-based models has long been recognised as computationally challenging. Their bottom-up nature imposes the use of non-parametric or indirect inference methods, which in turn typically requires a significant amount simulated data. However, their high computational requirements makes the application of these estimation methodologies unfeasible in practice. This hurdle can limit the applicability of agent based models for quantitative policy advice, for example in scenario analysis, in cases where the parameter calibration cannot be checked against empirical data. We show how this problem can be overcome by estimating the Dosi et al. (2015) agent-based model. This extends the original ‘Keynes meets Schumpeter’ model of Dosi et al. (2010) allowing for the interaction of fiscal and monetary policy. 18 free parameters of the model are estimated on 10 standard macroeconomic US variables, using annual and quarterly data, and the estimates obtained are shown to improve the fit of the baseline model on the data. A model-selection exercise is carried out, investigating impact of changing the expectation-formation mechanism. Finally, the original policy experiments are replicated using the new empirical estimates, showing that pushing the model into a low-growth regime leads to several key differences relative to the original conclusions. Overall, the exercise establishes the feasibility and relevance of macro-economic empirical parameter estimation and mechanism selection to ABM designers for improving the fit of scenario analyses.