For AI labs · exclusive sale to one AI lab
Global macro and financial market training data for one AI lab
Language models still guess at macroeconomics. Ask one what an oil spike does to Nigeria versus Spain, or what a stronger dollar plus an EM FX collapse does to Japanese house prices, and you get a confident paragraph with no numbers behind it. Real macro history can't fix that: there are only a few hundred quarters of it, and almost none of the combined shocks labs care about.
RoboMacro's Global Macro Economic Simulations fix it with theory-consistent synthetic data, written for the models labs actually train.
The 1,000-simulation and 10,000-simulation packs are for exclusive sale to one AI lab. Contact Chris: chris@robomacro.com
What's in every simulation
- 30 countries. One record per country, so every shock shows how it travels across economies.
- 45 series per country record: 37 country-specific and 8 global, Q1–Q20 impulse responses.
- 25 market series: FX, government yields, bond prices, equities, bank equity, house prices, VIX and commodities. Markets and macro move together in the same record.
- Records written as JSON answers. Given a shock and a country, the structured answer a model should return: GDP peak and its quarter, the three-year CPI effect, the equity peak, and paths at Q4, Q8 and Q12.
- QA rows grounded in the numbers, such as which country takes the biggest GDP hit.
- Tool-use rows that teach a model to request a simulation instead of guessing it.
- A human-readable report of about 11,000 words: cover briefing, thirty country chapters with charts, and the full Q1–Q20 JSON for all 45 series.
The engine
Every simulation is solved by the Global Macro Model v6.4. It understands 12 shock types (policy rates, oil, metals, housing, fiscal, productivity, tariffs, risk premia and more), and they stack. Korea hikes 200bp while China and Korea put 20% tariffs on each other: one scenario, solved across all 30 countries.
An eval built to be hard
- A frozen, SHA-256-hashed eval spec: task format, metrics, thresholds and decoding settings locked before any model is scored.
- A private 20-simulation held-out set (600 country rows). Shocks located in six reserved countries never appear in anything a model trains on.
- Models are scored on direction, magnitude error, cross-country ranking, peak timing and QA accuracy, against the base model, few-shot prompting and a sign rule.
Packs
| Pack | Simulations | Country records | Availability |
|---|---|---|---|
| Free sample | 10 (8 train + 2 eval) | 300 | CC BY 4.0, training allowed |
| 1,000-simulation pack | 1,000 | 30,000 | Exclusive sale to one AI lab |
| 10,000-simulation pack | 10,000 | 300,000 | Exclusive sale to one AI lab |
Same schema across every tier. Paid packs are licensed separately from the free sample.
Start with the free sample
The 10-simulation sample is free under CC BY 4.0, including for training and fine-tuning:
- Sample page: training records and reports side by side
- Hugging Face: robomacro/global-macro-economic-simulations (configs:
default,qa,tool_use) - GitHub mirror
- Blog: The Macro Training Data Only One AI Lab Will Own
Why synthetic macro data works
Chib, Tan and Zhang (Learning the Macroeconomic Language, arXiv:2512.21031) trained a small transformer on DSGE-simulated US data and showed that theory-generated data helps a learner where real data is scarce. We take the same idea wider: 30 countries instead of one, 45 series instead of seven, markets as well as macro, combined shocks, and records written for language models. Read the comparison: Same Bet, Different Engine.
Talk to us
One lab gets these packs. If that should be yours, email Chris Apostolou: chris@robomacro.com
RoboMacro Ltd, Cyprus. Simulated paths are impulse responses versus a model baseline, not forecasts and not investment advice.