Focus
We examine how agentic artificial intelligence (AI) could support macroeconometric analysis in central banks. Unlike conventional econometric software, which estimates models specified by the researcher, an AI agent can help formulate the research design itself by selecting variables, identification strategies, samples and model assumptions. We introduce the BIS Macoreconometric Research Companion, MaRCo@BIS for short: a framework intended to assist economists while preserving human responsibility for consequential decisions.
Contribution
MaRCo@BIS combines an AI agent with a verified dictionary of published model specifications, a tested library of econometric methods and a structured workflow. The agent proposes models, retrieves and harmonises data, conducts estimation and diagnostics, and explains its choices. Researchers must confirm or override key decisions, while the system records the rationale, alternatives considered and results of each stage. The framework is benchmarked against published structural vector autoregression (VAR) studies and involves replications, adaptations to new data and economies, paper ingestion and an open research question.
Findings
The experiments suggest that agentic AI can substantially reduce the mechanical effort involved in econometric research and help identify economic and data-related problems at an early stage. It successfully reproduced published findings where the specification and data were sufficiently well defined, while highlighting the greater uncertainty associated with set identification and incomplete data. An open exercise on fiscal policy showed that short-run results were more robust than medium-run effects, which depended heavily on modelling choices and prior assumptions. We conclude that AI should support, rather than replace, expert judgment. Human review remains essential because of the risks of cherry-picking specifications, automation bias, data limitations and economically inappropriate interpretations.
Abstract
Over the past half-century, econometric software has greatly reduced the burden of implementation, while leaving substantive choices about model specification, identification and sample largely to the researcher. Agentic AI can propose specifications and explain their economic rationale, while earlier automated selection methods could simply searched across candidate specifications using statistical criteria. This paper introduces MaRCo@BIS, the MAcroeconometric Research Companion: a framework of plain-text workflow files, a verified dictionary of canonical model specifications and a tested reference library, read by an off-the-shelf agentic coding environment. MaRCo@BIS proposes specifications and provides their rationale, retrieves and harmonises data, estimates, checks results against economic priors and logs every choice for replication, while the researcher confirms or overrides each (consequential) decision. We document the design, benchmark the agent’s replications of three widely-used structural VARs against the originals, and illustrate the framework across five exercises ranging from a verified replication to an open design question. We also discuss the necessity of human supervision, the risks of specification search and automation bias, and the safeguards against these risks. The framework is released as a public good for the central banking community and macroeconomic researchers more broadly.
