For business leaders, investors, and policymakers, some of the most critical information about the economy arrives after the fact. Official measures of GDP, inflation, employment, and other economic activity can take weeks or months to be released. But the Fed, businesses, and investors often must make decisions before those numbers arrive because some decisions can’t wait.
Economists call the process of estimating an economic indicator before its official release “nowcasting,” and new research from Georgia Tech asks a simple question: Can AI take the pulse of the economy in real time?
To answer this question, researchers developed LiveMacroEval, a live benchmarking app that tracks AI agents making hourly predictions of 16 major U.S. economic indicators. The agents’ results are compared with Federal Reserve nowcasts, professional forecasts, and traditional statistical models.
Putting AI Nowcasting to the Test
The researchers tested indicators across four areas of the economy: production, consumer spending and prices, jobs, and housing. The research team included Xinyue Zhao, Ruiyi Zhang, Liqin Ye, Rui Cao, Pengtao Xie, and Dr. Sudheer Chava, Regents Professor and Alton M. Costley Chair, and Faculty Director of the Financial Services Innovation Lab, MS-QCF, and Center for Finance and Technology.
Every hour, the AI agents were asked for their best estimate of all 16 indicators. The agents could search the open Web for current information, including government data releases, Federal Reserve statements, news stories, and other publicly available information. However, they did not have paid access to sources such as Bloomberg, so they could not simply peek at the professional consensus.
The live approach addresses a major challenge in evaluating AI forecasting: memorization. Economic data is among the most widely reported information in the world, meaning an AI model may have encountered historical economic numbers during training.
“You can’t test AI agents’ ability to predict the economy by backtesting on historical data,” said Xinyue Zhao, a Ph.D. candidate at Georgia Tech and a researcher. “You can never tell whether its performance comes from memorization or from real inference.”
Instead, LiveMacroEval makes predictions before the official numbers are known, allowing researchers to evaluate whether AI agents can use newly available information to form and update forecasts in real time.
“This research highlights both the promise and the complexity of using AI to understand the economy in real time,” said Dr. Sudheer Chava, Alton M. Costley Chair and professor of Finance at Georgia Tech. “The ability of AI agents to continuously process and synthesize large amounts of new economic information could give economists and business leaders a valuable new tool for assessing changing economic conditions.”
The models tested included GPT-5, GPT-6 Astra, Claude Sonnet 4.5, Claude Sonnet 5, Qwen3, and DeepSeek V4.
Keeping Pace with the Economy
Today, a human nowcaster has finite time and attention. But before deciding whether to update a forecast, they need to track government releases, Federal Reserve speeches, oil prices, earnings calls, and other economic news. It’s a lot of work.
AI agents are continuously processing new information. On April 8, 2026, for example, a government report showed gasoline inventories falling faster than expected, while Federal Reserve meeting minutes warned that inflation could run hot. Within hours, GPT-5 raised its March inflation estimate from approximately 2.7% to approximately 3%.
The official number came in at 3.3%.
“This shows one of the biggest advantages of AI agents: They react the moment news breaks,” said Zhao.
A Key Finding – Success Depends on More than the Model
One of the study's most interesting findings was that AI design can matter as much as the model used. GPT-5 performed on par with the professional consensus, according to the researchers’ main measure of forecasting accuracy, accounting for how strongly each economic indicator tends to impact the stock market.
Other agents performed below the consensus, although most still outperformed a traditional statistical forecasting model. The strongest AI agents also outperformed professional forecasters in production and housing, where fresh information becomes available between expert surveys.
The finding: How AI is designed can matter as much as which model is used. When researchers tested Claude Sonnet 4.5 working alone, it scored –0.08. A team-based setup, in which one AI researched the question, another independently checked the work, and a coordinator combined the findings, it scored +0.02.
The team-based approach was the only setup that outperformed the professional consensus.
Where AI Struggled
AI agents were less successful when an economic indicator suddenly reversed direction. When conditions change and an indicator suddenly reverses, AI tends to play it safe and follow the recent trend. Individual models also showed different blind spots. For example, Claude Sonnet 4.5’s biggest misses came from bold, overconfident predictions. GPT-5 showed a tendency to underestimate housing indicators. Those limitations point to an important role for human economists.
“AI goes through every data release, news story, and Fed statement around the clock, and flags when its estimate moves and why,” Zhao explained. “Economists understand how the economy is structured and have the judgment to spot regime changes.”
The Future of Economic Forecasting – AI as a Cooperative Tool
The researchers see AI as a complement to human expertise rather than a replacement for economists. AI agents can continuously absorb new information and update expectations, while humans can provide the economic knowledge and judgment needed to interpret those changes.
The hourly record of AI predictions can also become a valuable research tool for researchers. By tracking how forecasts change in response to specific news, researchers may be able to identify new signals that influence expectations and incorporate them into traditional economic models.
For business leaders, this research shows how AI can be tested on the future and not just the past. Genuine forecasting ability has to be tested on information the AI has not already seen. And choosing the highest performing model might not be enough, because how AI agents are organized to research, verify, and synthesize information can impact their performance.
The researchers’ next step is to explore how much more sophisticated AI team designs can improve forecasting, and which features make an AI a better forecaster. They also want to test whether AI agents can move beyond macroeconomic indicators to more complex financial data, including stock prices, credit markets, and corporate earnings.
LiveMacroEval offers an early look at a future in which AI continuously monitors a changing economy while human experts bring the judgment needed to understand what those changes mean and when to act on them.
LiveMacroEval is part of a broader effort at Georgia Tech’s Financial Services Innovation Lab to understand how large language models can be used responsibly in finance.
“LLMs are becoming more capable of working with financial information, but capability alone isn’t enough,” said Dr. Chava. “In the Lab, we’re running related projects to understand what these models can do well, where they fall short, and what guardrails are needed before they’re relied on for high-stakes financial decisions. The goal is to give the industry evidence, not hype, about how to use this technology effectively.”