The Bank for International Settlements (BIS) has published a working paper introducing MaRCo@BIS, the Macroeconometric Research Companion, and released the accompanying framework for central bank economists and other macroeconomic researchers. The proof of concept uses artificial intelligence (AI) agents to propose research designs and conduct econometric analysis, while requiring researchers to confirm or override consequential decisions. Experiments suggest it can reduce mechanical research work, but economic interpretation and policy use still require expert review.
The framework combines a verified dictionary of 30 published model specifications with 24 tested estimation modules and a structured workflow. Agents propose models, retrieve and harmonize data, run estimation and diagnostics, and document their reasoning. Researchers retain control over substantive choices, including variables, identification strategies, samples and data substitutions. Variables and identification cannot be selected solely for statistical fit and must have an economic or literature basis. Decision logs record alternatives and approvals, while reports disclose specification search and show initial results alongside revised estimates when revisions occur.
Benchmarking showed that replication accuracy depended on the model’s identification scheme and data availability. A model with a uniquely determined answer and deposited data reproduced the original numerical results to near machine precision. Agreement was less exact where multiple models satisfied the identifying assumptions or original data were only partly available. An open fiscal policy exercise found that immediate effects were more stable than results at longer horizons, which depended heavily on modelling choices and prior assumptions. The safeguards make these limitations easier to review but cannot prevent selective reporting across separate sessions or eliminate automation bias and inappropriate economic interpretations.