Aldo Glielmo (Banca d’Italia)

Informal meetings on complexity field

ABMs at central banks: Open source software and machine learning methods for faster and more robust modelling

Abstract: Artificial Intelligence (AI) is transforming science, driving a paradigm shift through the use of algorithms that automatically improve with more data and/or more computational power. Agent-Based Models (ABMs), with their computational nature, are uniquely positioned to lead this transformation within economics and the social sciences. A fundamental prerequisite to any AI deployment is the availability of high-performance, easy-to-use models in open source (OS). In this regard, I will describe recently released packages such as ABCredit and BeforeIT.jl [Glielmo et al., ArXiv 2025], software designed to quickly build state-of-the-art macro ABMs. Once such OS models are available, the use of AI-software agents can relieve modellers from manually defining complex yet accurate behavioural rules, and alleviate the problem of model misspecification. In this regard, I will showcase recent work on incorporating reinforcement learning (RL) -driven agents in traditional ABMs [Brusatin et al., ICAIF 2024; Agrawal et al., MARW-AAAI 2025] and examine the promises and challenges of equipping ABMs with large language model (LLM) agents [Biancotti et al., FinNLP-COOLING 2024].