BISTRO: A General-Purpose Oracle for Macroeconomic Forecasting

BISTRO (BIS Time-series Regression Oracle) is a general-purpose time series model for macroeconomic forecasting. It is based on the transformer architecture used in Large Language Models (LLMs) and is fine-tuned on thousands of macroeconomic time series from the BIS data portal.

This repository contains the code and example notebooks for using BISTRO.

Security Vulnerability

Important: This project contains a dependency with a known critical security vulnerability. Please read our Vulnerability Disclosure for details on the associated risks and mitigation steps before using this code.

Overview

Many traditional forecasting approaches require a different model for each variable. BISTRO provides a low-cost and flexible tool for baseline forecasts and conditional scenarios.

Intro video

Key Capabilities

• Unconditional forecasting: Produces baseline forecasts for key aggregates (e.g., inflation).
• Conditional scenarios: Lets you fix future paths of variables (e.g., oil prices, exchange rates) to explore alternative scenarios.
• Nonlinear patterns: Captures nonlinear relationships that standard linear models may miss.
• Fine-tuned: Trained on 4,925 time series across 63 economies.

Try it now (no installation required)

You can run everything directly in your browser using Google Colab. It is free and requires no setup.

Notebook	Description	Link
Forecast Single Series	Simple forecast for one variable (e.g., inflation)
Forecast Unconditional	Forecast with extra variables (covariates)

How to use Google Colab

1. Click a link above to open the notebook.
2. If asked, sign in with your Google account.
3. In the top menu, click Runtime → Run all.
4. Scroll down to see the results.

Python version (recommended)

• Recommended option: Use Google Colab, which comes with a pre-configured Python environment.
• If you run locally (optional): Use Python 3.11 (or a close equivalent). This is typically the most compatible choice for current ML libraries.
  (Exact package versions are listed in requirements.txt.)

Using your own data

To forecast your own data:

1. Create a CSV file like the examples (usually a Date column and a Value column).
2. In Colab, click the folder icon on the left.
3. Drag and drop your CSV file.
4. Update the filename in the notebook (e.g., change 'bis_cpi_us_yoy_m.csv' to 'my_data.csv').

Running locally (optional)

Most users do not need this. Use this only if you want to run the notebooks on your own computer.

Model weights and Git LFS (optional but recommended)

Some large files (e.g., model weights) may be stored using Git LFS. If Git LFS is not installed, git clone may not download these files correctly.

Install Git LFS first (recommended):

1. Install Git LFS: https://git-lfs.com/
2. Then run:

   git lfs install
   git clone https://github.com/bis-med-it/bistro.git

Project structure

• data/: Sample CSV data (e.g., US CPI, policy rates).
• script/: Notebooks and scripts for running the model.

Citation

If you use this model or code, please reference the paper:

BISTRO: A General-Purpose Time Series Model for Macroeconomic Forecasting
Batuhan Koyuncu, Byeungchun Kwon, Marco Lombardi, Hyun Song Shin, Fernando Perez-Cruz
BIS Quarterly Review, March 2026