Central Banks' Information Edge Is Disappearing

The precedent: forty years of information advantage

In 2000, two economists documented an uncomfortable fact for every central bank. Christina and David Romer, in a paper on the American Economic Review, showed that the Federal Reserve's internal forecasts, the so-called Greenbook, systematically outperformed those of private analysts.

The mechanism was clear: resources, access, proprietary data. The central bank knew first. The market followed.

That differential sustained the architecture of monetary policymaking for forty years. Today it is eroding. A language model with web search replicates what once required a department of twenty economists.

This is a regime change. The scarcity of macro information, controlled by policymakers, is becoming a good accessible to anyone who pays for a query.

The evidence: the benchmark of August 31, 2026

On August 31, 2026, six researchers published a study on arXiv that deserves attention. The title: «Can LLMs Take the Pulse of the Economy?^[1]». The question is direct. The answer changes the calculation for those managing capital.

The authors built LiveMacroEval, a contamination-resistant benchmark. The technical problem was obvious: indicators like GDP and CPI are widely reported and memorized during pretraining. Evaluating a model on historical releases falsifies the result.

The solution: hourly nowcasts on sixteen U.S. macroeconomic indicators, produced before each official release. Four state-of-the-art LLM agents, configured with web search.

The comparators were the nowcasts of the Federal Reserve's regional banks, the Bloomberg ECOS professional consensus and an auto-ARIMA baseline. The aggregated result over six months: accuracy broadly comparable to institutional and professional benchmarks. Performance varies considerably across individual indicators.

The mechanism: why it happens now

Three forces converge.

First: real-time web search. An agent queries primary sources, regional communications and high-frequency data the moment they are released. Second: frequency. Institutional nowcasts arrive on a fixed schedule, a model produces hourly estimates and updates them with every new data point.

Third: marginal cost. A department of economists costs millions per year. A query costs cents.

The causal mechanism is linear. The central bank's advantage derived from scarcity: few actors had the resources to aggregate and interpret real-time data. That scarcity is disappearing.

When the tool becomes commodity, the information differential between the policymaker and the market compresses. The 48-72 hour cycle between the Fed's view formation and its pricing in markets shortens. A structural signal.

The objection, and why the direction holds

The objection is legitimate. Performance varies widely across individual indicators, the authors write. A model brilliant on CPI remains weak on noisier series. The aggregate hides dispersion.

I add the second limit. Six months is a narrow window. A tranquil market regime rewards tools different from a liquidity crisis.

The true test comes under stress conditions, when data becomes contradictory and human judgment weighs more. I concede these points. They change the magnitude, leave the direction intact.

The trajectory of model capacity points upward. The cost of access points downward. Two vectors that reinforce each other define a structural trajectory.

My position

This desk's position is explicit. Language models are reaching parity with central banks' institutional nowcasts. This transforms macroeconomic information from a rare resource, controlled by policymakers, into a public good accessible in real time.

The erosion touches the heart of the monetary decision cycle since 1980.

What would change this reading? One precise piece of data: that models remain stably behind the Fed's nowcasts in at least three successive publications of the same benchmark, under conditions of market stress. That evidence would dismantle the thesis. Until then, the direction remains described.

Three implications for capital

One, for family offices and sovereign wealth funds. Tactical timing based on the lead in macro data loses value. The information rent migrates to those who integrate these tools into the process. Horizon: 36 months.

Two, for the chief risk officer. VAR models assume that the price reaction to data follows historical latency. That latency compresses. A scenario of near-instantaneous pricing of macro surprises deserves a place in models. Horizon: 18 months.

Three, for the CFO and investor relations. The macro narrative you bring to market now competes with estimates generated in real time at near-zero cost. Guidance founded on an interpretive advantage becomes fragile. Horizon: 24 months.

The common denominator: the information scarcity that commanded a premium is losing price. Anyone building portfolios on that premium should revisit the assumption.

The forecast

By the end of 2027, at least one G7 central bank will publicly integrate LLM agents into its nowcasting flow, in declared form in an official document or speech.

Confidence: 62%. Horizon: December 31, 2027. Verification: an official publication or statement from a G7 central bank describing the use of language models in producing macro estimates.

The signal that would dismantle the thesis: absence of declared adoption by a G7 central bank on the stated date.

What to watch

First: minutes and working papers from the Federal Reserve's regional banks. A reference to LLMs in the nowcasting process anticipates formal adoption.

Second: the update of LiveMacroEval over a second time window. A confirmation under market stress strengthens the thesis, a failure weakens it.

Third: the cost per query for agents with web search. Continuous decline accelerates the commoditization of macro analysis.

The divergence between model capacity and institutional advantage always resolves. The question is how, and in how long.