From equations to prompts: integrating LLMs into a hybrid macroeconomic ABM

Abstract

Agent-based models (ABMs) typically rely on equations to incorporate agents' decision-making, ranging from ad-hoc rules or simple heuristics to complex, theory- or data-driven approaches. Recent advances in large language models (LLMs) make it possible to represent economic decision-making through information-conditioned decisions rather than explicitly specified behavioural equations. However, whether LLM-driven agents can feasibly replace individual components in an ABM and the degree to which they can replicate observed economic dynamics remains insufficiently understood. In particular, it is unclear whether the integration of LLM-driven agents in a macroeconomic ABM can generate historically plausible macroeconomic trajectories and how their performance depends on the information provided to them and on the scale of the underlying model. This paper addresses these questions by introducing an LLM-driven bank agent into a data-driven macroeconomic model of the Greek economy, where it determines consumer credit and lending rates. Results show that the LLM-driven agent's decisions can reproduce historical consumer credit dynamics relatively well, although their performance in replicating lending rates is weaker and more heterogeneous; broader economic information generally improves performance, while larger models do not systematically outperform smaller ones. Overall, the study provides encouraging evidence that LLM-driven agents can be feasibly integrated into macroeconomic ABMs, assimilating economically relevant information and producing outcomes that closely match observed data, without requiring explicit functional forms for behavioural relationships - pointing towards the possibility of enabling economic simulations to shift from equations to prompts.