BIS releases BISTRO macro oracle

The BIS posted a working paper titled "Introducing BISTRO: a foundational model for unconditional and conditional forecasting of macroeconomic time series" as BIS Working Papers No. 1337 on March 13, 2026, authored by Batuhan Koyuncu, Byeungchun Kwon, Marco Jacopo Lombardi, Fernando Perez‑Cruz and Hyun Song Shin. (bis.org) BIS also ran a shorter write‑up in the March 2026 Quarterly Review describing BISTRO as a general‑purpose time‑series regression oracle built on transformer architectures and fine‑tuned on central‑bank macro data. (bis.org) The paper states BISTRO was fine‑tuned on "thousands" of macroeconomic time series drawn from the BIS data portal and is presented as capable of both unconditional forecasting and conditional (scenario) forecasting tasks. (bis.org) In an empirical case study the authors evaluate BISTRO on the 2021 global inflation surge and report that BISTRO produced forecasts that tracked the persistence of that inflation wave more accurately than standard benchmark models that tended to revert to historical means. (bis.org) BIS published reproducible code, example notebooks and a Google Colab demo in the public GitHub repository bis‑med‑it/bistro, which is released under an Apache‑2.0 license and shows recent commits and example branches such as "bistro‑finetuned." (github.com) The GitHub repository contains a VULNERABILITY_DISCLOSURE.md and an explicit README warning about a dependency with a known critical security vulnerability, accompanied by mitigation guidance before running the notebooks. (github.com) The BIS materials include an online annex that details BISTRO’s attention mechanism as a dynamic weighting scheme and provide the "data behind the graphs" files to reproduce the article’s plots and scenario experiments. (bis.org)