Learning the Macroeconomic Language
Paper • 2512.21031 • Published
You need to agree to share your contact information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
This 10-simulation sample is free to use under CC BY 4.0, including for model training. Tell us who you are so we can share updates on the full packs.
Log in or Sign Up to review the conditions and access this dataset content.
Global Macro Economic Simulations and Financial Market Responses
This listing is ten sample simulations only. Two licensed training packs, 1,000 simulations and 10,000 simulations, are for exclusive sale to one AI lab. Same training-record schema, more shocks. They are not in this sample.
Contact Chris: chris@robomacro.com
Ten named global macro simulations. Each one ships the full human-readable report — cover briefing, thirty country chapters with charts, and the unedited Q1–Q20 JSON for all 45 series — plus the same numbers as machine tables. Each human-readable report averages about 11,000 words.
Training records and human-readable reports, side by side: https://robomacro.com/GlobalMacroTrainingDataset/
- Sample page: robomacro.com/GlobalMacroTrainingDataset
- Blog: The Macro Training Data Only One AI Lab Will Own
- Comparison: Same Bet, Different Engine: Chib & Tan's DSGE Transformer Next to Our Multi-Country Sims
- GitHub mirror (same sample): ChrisRoboMacro/Global-Macro-Economic-Simulations-and-Financial-Market-Responses
- Contact: chris@robomacro.com
This Hub listing is the training-record JSONL a model would train on. Do not open .html in the file viewer — that is source code. On robomacro.com each simulation shows the full training records and a typeset report next to them. The report is not the training product.
The model is not included. There is no solver, no weights, and no coupler in this repository. Paths are impulse responses versus a model baseline, not forecasts and not market data.
The point is to improve how an AI reads economics: what happens after an event, and after events in combination. Example: an oil-price spike in Nigeria versus Spain. Or a stronger dollar together with an EM FX collapse, and what that does to Japanese house prices.
| Split | Simulation | Training records | Human-readable report | Largest GDP move |
|---|---|---|---|---|
| train | US policy rate +200bp | Records | Report | US -0.52% |
| train | US policy rate -200bp | Records | Report | US +0.47% |
| train | Oil $200/bbl | Records | Report | TR -5.81% |
| train | Oil $50/bbl | Records | Report | SA -4.05% |
| train | Metals supply -20% | Records | Report | CL +0.58% |
| train | UK housing -25% | Records | Report | UK -3.27% |
| train | VIX 80 | Records | Report | NL -2.26% |
| train | Risk premium +250bp | Records | Report | AR -2.32% |
| eval | US-China tariffs 25% each way | Records | Report | CN -0.51% |
| eval | Oil $180 and US +150bp | Records | Report | TR -4.96% |
Each machine row is one (shock, country): the active treatment, units, Q1–Q20 paths for 45 series (GDP, CPI, policy rates, the yield curve, bond prices, equities, VIX, housing, NEER/USD, labour, credit, energy/metals/food/gas/copper/wheat/gold), locked summary, QA flags.
| Split | Simulations | Country rows | File |
|---|---|---|---|
| train | 8 | 240 | train.jsonl |
| eval | 2 | 60 | eval.jsonl |
Also here
UNITS.md— series names and units (RER,CPI_3yr_pp).catalog.json— the ten active treatments.train_qa.jsonl/eval_qa.jsonl— short questions grounded in the numbers.tool_use_examples.jsonl— three treatment → summary traces.
House commentary, if present, was written on our hardware (GLM-5.3-Flash / DeepSeek v4.1). A customer-designated LLM for language is a separate paid job.
The 1,000-simulation and 10,000-simulation packs are not in this repo. They are for exclusive sale to one AI lab. Contact Chris: chris@robomacro.com
For AI labs
The 1,000-simulation and 10,000-simulation packs are for exclusive sale to one AI lab, under a licence separate from this sample.
What a pack contains, per simulation:
- 30 country records, each with Q1–Q20 impulse-response paths for 45 series: 37 country-specific and 8 global. 25 of them are market series: FX, government yields, bond prices, equities, bank equity, house prices, VIX and commodities.
- Records written as JSON answers: given a shock and a country, the structured answer a model should return (GDP peak and quarter, three-year CPI effect, equity peak, paths at Q4/Q8/Q12).
- QA rows grounded in the numbers, and tool-use rows that teach a model to call for a simulation instead of guessing it.
- A human-readable report of about 11,000 words.
Shocks come from the Global Macro Model v6.4: 12 shock types (policy rates, oil, metals, housing, fiscal, productivity, tariffs, risk premia and more), and they stack, so packs cover combined scenarios, not just single shocks. Evaluation runs under a frozen, SHA-256-hashed eval spec, with a private 20-simulation held-out set (600 country rows) in which shocks located in six reserved countries never appear in anything a model trains on.
Contact Chris: chris@robomacro.com
Related work
Chib, Tan and Zhang, Learning the Macroeconomic Language (arXiv:2512.21031), trained a transformer on DSGE-simulated US data and showed that theory-generated data helps a learner where real data is scarce. Our comparison of the two approaches: Same Bet, Different Engine.
License
This 10-simulation sample is licensed under CC BY 4.0: you may use it for anything, including training and fine-tuning models, with credit to RoboMacro (robomacro.com). See LICENSE.md. The 1,000- and 10,000-simulation packs are not covered and are licensed separately. Paths are deviations from a model baseline and are not investment advice.
- Downloads last month
- 112