AI agents nowcast CPI, GDP and jobs live for six months. GPT-5 tied Wall Street's consensus

Testing whether AI can forecast the economy usually runs into a wall: GDP and CPI figures are everywhere online, so a model tested on past releases may simply remember them. A paper posted to arXiv on August 31, 2026 solves this by testing live. Its authors, led by Xinyue Zhao (Georgia Tech) and Ruiyi Zhang (UC San Diego) with corresponding author Sudheer Chava (Georgia Tech), built LiveMacroEval, a benchmark that scores AI estimates made before each official release.

The setup

• Indicators: 16 headline U.S. releases in four groups. Supply and production (real GDP, industrial production, durable goods orders, ISM manufacturing), demand and inflation (CPI, PCE prices, PPI, real consumer spending, retail sales, ISM services), the labor market (nonfarm payrolls, unemployment rate) and housing (four indicators, including new and existing home sales).
• Agents: GPT-5, Claude Sonnet 4.5, Qwen3-235B and Qwen3-80B, each with web search, given the same prompts.
• Cadence: a fresh nowcast every hour through each pre-release window, which yields hundreds of thousands of estimates.
• Window: target periods from November 2025 through March 2026, covering official releases from December 1, 2025 to May 5, 2026.
• Benchmarks: five Federal Reserve regional-bank nowcasts (Atlanta's GDPNow, New York, St. Louis, Cleveland and Chicago), the Bloomberg ECOS professional consensus, and a simple statistical model (auto-ARIMA).

How it is scored

The study uses two new measures, both tied to markets rather than raw forecast error.

LiveMacro Score. Economic surprises move stocks. The score asks whether an agent's estimate predicts the stock market's reaction at the announcement better than the Bloomberg consensus does. Zero means as good as the consensus, positive means better, and negative means worse; the scale runs from -1 to +1.

LiveBetting Score. The cumulative return from placing hourly $1 bets on Polymarket-style contracts based on each nowcast, across the February and March windows. GPT-5's March real GDP window alone involved 1,423 bets.

The results

Overall: a tie with the consensus. GPT-5 led the four agents with a LiveMacro Score of +0.004, which the authors describe as statistically indistinguishable from the Bloomberg consensus, so a tie (the paper reports 90% bootstrap intervals in a chart). The other three agents sat modestly below the consensus. Claude Sonnet 4.5 and Qwen3-235B still beat the statistical baseline; Qwen3-80B was indistinguishable from it. The equal-weight average of the four agents scored -0.0437, below the consensus. nContext: n=16 indicators, 4 agents · releases Dec 1, 2025 to May 5, 2026 · Zhao et al., arXiv:2608.30110 · 90% bootstrap intervals in Figure 2

The authors call matching the consensus "a non-trivial outcome," since that consensus is known to be very hard to beat, and they describe it as initial, if modest, evidence.

By indicator: large differences. On supply and production and on housing, the agents materially beat the consensus, led by GPT-5 and Qwen3-235B respectively. The authors suggest these areas benefit from component data published at higher frequency that a web-searching agent can pick up.

GDP and jobs: competitive with the Fed. In the betting test on real GDP, all four agents finished with positive cumulative returns and three of them beat the Atlanta Fed's GDPNow; only the New York Fed's nowcast finished ahead. On the unemployment rate, the leading agent finished above the Chicago Fed's nowcast.

CPI: the economists win. On inflation, the Cleveland Fed nowcast and the Bloomberg consensus beat every AI agent. The authors point to the structure of inflation forecasting, built on central-bank communication and a well-anchored inflation regime, where institutional models are tightly tuned.

Were the AIs just copying the consensus?

A fair suspicion: an agent with web search could look up the Bloomberg consensus or GDPNow and repeat it. The authors tested this three ways and found no sign of copying. Across all 16 indicators, the final AI nowcast sat a median 0.42 standard units of surprise away from the Bloomberg consensus, and the gap did not close as the release approached. It was at least half a unit away in 46% of series and at least a full unit away in 25% (series counts not restated in that passage; no interval disclosed). The agents' revisions also did not move in step with the Fed nowcasts. nContext: n=16 indicators, series count not stated · releases Dec 1, 2025 to May 5, 2026 · arXiv:2608.30110 · margin: not disclosed

The limits

This is one six-month window with four agents, and the main result is a tie, not a win. Scores are measured against stock-market reactions and simulated betting, not actual trading. And the strongest results come from specific indicator groups, which makes the aggregate easy to over-read in either direction.

It does fit a pattern we see elsewhere. On ForecastBench, AI systems with tools reach expert level on future questions, while the World Cup forecasting study found frontier agents matching, not beating, the market price. AI agents are now good enough to keep up with the professional consensus. Beating it consistently is another matter.