Dynamic interactions in economics: from micro-level games to macroeconomic agent-based models

Thesis

We study economic dynamics using two representative classes of models with interacting agents — (i) a microeconomic model of agents repeatedly playing normal-form games and (ii) a large-scale multi-country macroeconomic model. On the microeconomic level, we provide novel results on the likelihood of convergence of best-response dynamics, under which each player selects the pure strategy that maximises their payoff, given the strategies chosen by others. More generally, we show that convergence to equilibrium is rare under a range of well-known learning rules (such as reinforcement learning or fictitious play) when the games’ number of players or number of actions is large. This implies that the assumption that systems converge to an equilibrium may be unrealistic. Therefore, turning to the macroeconomic scale, we employ agent-based models, which, unlike traditional macroeconomic models (e.g. DSGEs), do not rely on this assumption. Specifically, we develop a novel framework for constructing and calibrating large-scale, data-driven, agent-based models of entire economies. Our model incorporates individuals, households, firms, banks, governments, and central banks across 38 countries, encompassing markets for goods, labour, credit, and housing. Each agent is initialised using publicly available microdata by aggregating different data sources of varying levels of granularity, ensuring realistic heterogeneity. We calibrate the model using state-of-the-art simulation-based inference techniques: neural posterior estimation and neural ratio estimation. In both cases, we demonstrate that the calibrated model outperforms a simple benchmark time series model and the current state-of-the-art agent-based model in terms of forecasting accuracy. This improvement is rigorously assessed and confirmed as statistically significant using a Bayes factor approach. Our findings suggest that agent-based models when calibrated with real-world data and advanced inference techniques, offer a compelling alternative to traditional models for understanding and predicting economic dynamics. Our agent-based model is built within a user-friendly software platform we have developed, making it easy to create, calibrate, validate, run, and evaluate alternative models.

Authors: Wiese, S

• DOI: 10.5287/ora-5reg8nv9g
• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English
• Keywords: agent-based modelling; game theory; calibration; forecasting
• Subjects: Microeconomics; Macroeconomics