Bayesian inference for economic agent-based models

Interest in agent-based models of financial markets and the wider economy has increased consistently over the last few decades, in no small part due to their ability to reproduce a number of empirically-observed stylised facts that are not easily recovered by more traditional modelling approaches. Nevertheless, the economic agent-based modelling paradigm faces mounting criticism, focused particularly on the rigour of current validation, calibration, and model selection practices, which still, in many cases, remain qualitative and stylised fact-driven. While the literature on quantitative and data-driven approaches has seen significant expansion in recent years, many studies have focused on the introduction of new methods that are neither benchmarked against existing alternatives nor rigorously tested on the large-scale and computationally-expensive variants that now dominate the literature, instead choosing to focus on highly-simplified models that are no longer representative of the current state of the art. This has ultimately led to some doubt regarding the applicability of current approaches to the kinds of problems of interest to most contemporary modellers and policymakers. In response, this thesis attempts to make a number of meaningful contributions, both by providing a more complete and unified perspective on the existing literature and by additionally advancing it in a number of key areas. To elaborate, in Chapters 1 and 2, we review existing contributions and compare a number of prominent agent-based model estimation methods, both established and new, through the use of an experimental protocol that allows for the determination of the respective strengths and weaknesses of each approach and the quality of the resultant parameter estimates. Overall, we find that a simple, likelihood-based approach to Bayesian estimation consistently outperforms several members of the far more popular class of simulated minimum distance methods and results in reasonable parameter estimates in many contexts, with a degradation in performance observed only when considering a large-scale model and attempting to calibrate a substantial number of its parameters. This ultimately leads us to argue for a paradigm shift, suggesting that an increased emphasis on Bayesian methods is much needed and that attempts should be made to improve the performance of existing techniques. Following on from the above recommendations, Chapter 3 proposes a neural network-based approach to Bayesian estimation that addresses a number of the fundamental weaknesses of the best performing methodology identified in Chapter 2. When benchmarking our proposed approach, we find that it consistently results in more accurate parameter estimates in a variety of settings, including the estimation of financial heterogeneous agent models, the identification of changes in dynamics occurring in models incorporating structural breaks, and applications to a large-scale, state-of-the-art model of the UK housing market. Chapter 4 further builds upon these ideas and demonstrates how the proposed Bayesian estimation methodology may be extended to facilitate the application of contemporary Bayesian model selection tools to a variety of inter- and intraday financial agent-based models and to benchmark such models against traditional asset return modelling approaches. Finally, Chapter 5, which concludes the thesis, demonstrates how recent advances in Gaussian process emulation may be applied to the approaches developed in Chapters 3 and 4 to significantly reduce the overall computational costs and thus yield a computationally-efficient and widely-applicable framework for Bayesian inference in simulation models.