New AI Replays Financial History to Predict How Policy Shakes Markets

Economists have always argued about what would have happened if the government had acted differently. A new research dataset called PAWS tries to answer that with software. The team collected 36 real U.S. financial and economic policy episodes, 12,727 news stories tied to those policies, and 65,291 recorded actions taken by banks, regulators, companies and investors. Every action is linked back to the article that proves it happened, and lined up against how markets actually moved that day. That turns messy history into something a computer can replay, like a flight simulator for economic policy.

Why does that matter outside a university? Because governments make trillion-dollar calls — bailouts, trading bans, interest-rate shifts — with surprisingly little ability to test them first. If an AI can run a policy through a realistic simulation, regulators could spot unintended consequences before real people lose real money. Investors could stress-test their portfolios. Journalists could check whether a claim about a policy actually holds up. The researchers even had AI and human reviewers independently label 2,522 actions and found they agreed 89.4% of the time — a decent sign the data is trustworthy.

To show it works, the team dug into two famous moments: the 2008 ban on short-selling bank stocks (betting that a stock will fall) and 2001 'decimalization,' when U.S. stocks switched from fractions like 1/8 to pennies. In both cases, the system rebuilt a timeline of who did what, in what order, and how prices responded — even when news coverage was thin.

Here's the honest catch. In a test replay, the researchers found something unsettling: an AI can look highly accurate overall while completely missing rare, surprising actions — which are exactly the ones that cause crashes. Nailing the timing and scale of human reactions remains genuinely hard. So treat this as a promising rehearsal tool, not a crystal ball.