AI econometrics gains traceability rather than proven cost savings
MaRCo@BIS combines AI-led econometric workflows with verified models and researcher approval. Bank for International Settlements (BIS) researchers find that it improves traceability, while limited paired tests establish no general cost advantage.
Verified machinery meets human judgement
Marco Lombardi, Taejin Park and Fernando Perez-Cruz present MaRCo@BIS in BIS Working Paper No 1395. The public package contains 24 estimation modules, 12 workflow skills and 30 canonical model specifications.
It proposes specifications, retrieves data, estimates models and logs decisions; researchers confirm consequential choices.
Five exercises cover replication, adaptation, paper ingestion and open research design.
In three structural VAR benchmarks, Kilian (2009) reproduces all nine impulse responses to roughly 10^-12. Uhlig (2005) matches qualitatively, with medians dependent on simulation draws.
Peersman (2026) permits only a partial energy-channel replication because a food-shock instrument is unavailable; persistence differs from the original.
More checks, not cheaper sessions
Two paired comparisons with Claude Code alone reproduced Uhlig's headline finding in all four sessions: output-response intervals included zero.
MaRCo@BIS added verified specifications, residual diagnostics and durable decision logs.
Original-sample sessions each used about 1.6 million price-weighted tokens; updated-data sessions used 3.7 million with the framework against 2.2 million without it.
Network barriers contributed to the difference.
The Apache-2.0 package is a proof of concept, distributed without warranty; the comparisons do not establish general cost savings.
Traceability is not validation
The authors distinguish auditable implementation from economic validity: new analyses still require expert review.
Logged revisions expose specification search but cannot prevent selective reporting across fresh sessions.
They also warn that polished answers invite automation bias, even with approval gates.