The Macro Training Data Only One AI Lab Will Own
Anyone who has tried to fit a serious model to macro data runs into the same wall. There isn't enough history. US quarterly data going back to 1947 gives you about 312 observations. A language model learns from billions of tokens. You can't close that gap by waiting for more quarters.
Siddhartha Chib and Fei Tan have a clever answer in Learning the Macroeconomic Language: stop asking a structural model for forecasts and use it as a data generator instead.
They take the seven Smets-Wouters US series: output, consumption, investment and wage growth, plus hours worked, inflation and the nominal interest rate. The sample is cut into three blocks that never touch. The 50 quarters from 1947:Q3 to 1959:Q4 are used only to estimate a Smets-Wouters style DSGE model with stochastic volatility and fat-tailed (Student-t) shocks. The 231 quarters from 1960:Q1 to 2017:Q3 are the real training data. The last 31 quarters, 2017:Q4 to 2025:Q2, are held out for testing, and that window includes COVID and the 2021-23 inflation surge. Not an easy exam.
From the estimated model they take 10,000 parameter draws and simulate each one forward for 1,000 quarters. That's 10 million synthetic observations, all consistent with the theory, each from a slightly different version of the economy. Each variable is cut into ten percentile bins so the data look like tokens, and seven small transformers, one per variable, read the last four quarters to predict the next. The whole thing is about 52,000 parameters and trains in minutes on a Mac.
The part I like most is the mixing. Pre-training on simulations and then fine-tuning on 231 real quarters would barely move anything, so instead every training batch is 90% synthetic and 10% real. They read that ratio as prior strength, in the Bayesian sense. At zero you have dogmatic theory. At one you overfit a tiny sample.
The revised paper (v3, August 2026, with Zhixun Zhang joining as co-author) puts the approach up against a rolling VAR(4). The theory-heavy transformer matches or beats the VAR on token accuracy for all seven variables, with the biggest jump on the interest rate (0.484 to 0.710). On predictive log likelihood it wins on five of seven, with the VAR keeping an edge on wage growth and hours worked. Flip the mix to 90% real data and the results get worse, which the authors put down to overfitting.
So the network did best when it leaned hardest on theory. A structural model hands a data-starved network coherent, economically disciplined examples, and discipline is what that network needs most. For anyone training an AI to reason about economies, the takeaway is practical. Theory-consistent simulation data is a real training input. The questions that matter are where you get it, how much of the world it covers, and whether you're allowed to train on it.
Taking the idea global
Our Global Macro Economic Simulations dataset is built on the same bet, for a bigger job. Chib and Tan work with one country and seven series. The questions people bring to RoboMacro are conditional and global. What happens to Turkish GDP if oil goes to $200? What does a US-China tariff war do to Korean equities? Those are questions about transmission: how a shock in one place travels through rates, currencies, trade and risk appetite into somewhere else.
History is even thinner here. Very few quarters contain a clean, isolated shock you could learn from. So I built a model to generate them. It covers 30 countries linked through trade, currencies, bond yields, equities, house prices and bank credit. Each simulation takes a named shock, or two at once, and solves the response across all 30 economies.
There are 12 shock types: policy rates, oil, metals, housing, fiscal, productivity, risk premium, the VIX, bank equity, equity prices, sovereign spreads and bilateral tariffs. They can be run singly or combined. A record is one shock in one country: quarter-by-quarter paths for 45 series over 20 quarters, 37 of them country-specific and 8 global, covering GDP, inflation, policy rates, the yield curve, equities, housing, FX, labour, credit and commodities. Every path is theory-consistent, and every record carries a written summary locked to the numbers, plus QA checks.
In the free sample, $200 oil hits Turkey hardest, with GDP 5.81% below baseline at the largest move. $50 oil does the most damage to Saudi Arabia, at -4.05%. Same commodity, opposite direction, completely different losers. That asymmetry is exactly what a model should learn, and one example won't teach it.
That's where scale comes in. For every shock, the model solves dozens of equilibriums simultaneously, one for each linked economy, so the spillovers come out consistent across all 30. The 1,000 and 10,000 simulation packs use the same record format with many more shocks, single and combined, at different sizes. A model that has seen one oil shock has learnt a fact. A model that has worked through thousands of shocks across 30 countries has the material to learn the pattern, so when someone asks about a scenario that never happened, it can reason from transmission instead of borrowing from the nearest news story.
As far as I know, this is one of very few multi-country macro models built from the ground up to generate training data for AI, not just forecasts for people.
Most of the industry says no
There's another difference, and it's commercial rather than academic.
The big economics consultancies build excellent models. They also write contracts that stop you teaching an AI anything with them. Oxford Economics' subscription agreement with the UK Department for Business and Trade, which covers its Global Economic Model and sits on the government's Contracts Finder site, says in clause 3.2.2 that the customer may not use the services, reports, data or software for training algorithms or as AI training data. It goes further. You can't feed any of it into AI tools either, and putting extracts into a chatbot is given as an example of what's banned.
Capital Economics' Terms of Sale (clause 5.5, updated August 2026) prohibit using its data to train, develop, enhance or even prompt an AI system or language model, including fine-tuning someone else's, unless agreed in writing. Moody's Analytics data passed through Kroll's platform carries a similar condition: no use for training or developing AI or machine learning models.
I understand why. Their models are their crown jewels. But it leaves anyone building an AI economist in an odd position. Much of the best structured macro thinking is out there, and the licence says your model can't learn from it.
We've gone the other way. RoboMacro's simulation packs are made to be trained on, and the licence says so. Teaching a model how economies respond to shocks is the use we're selling, not a use we're trying to police. Even the free sample is released under CC BY 4.0, so you can train on it today.
Sold once, to one lab
There's one more thing that sets these packs apart. The full 1,000 and 10,000 simulation packs will be sold exclusively to one AI lab. One buyer, not a queue of licensees. A model that has worked through thousands of shocks across 30 countries has seen something its rivals haven't, and that edge is worth far more when nobody else can buy the same data. So we're only selling it once.
You don't have to take my word for what's in it. The dataset page shows each sample simulation with its training records next to a readable report. The free sample (10 simulations, 300 records) is on Hugging Face, and the same files are on GitHub. Load it, inspect the records, run it through your own pipeline.
If your lab wants an edge in macro reasoning that nobody else can buy, email me at chris@robomacro.com. Once the packs are sold, they're gone.
