Information aggregation in experimental markets, explained

Information aggregation is a market's ability to pull scattered private information into a price, so that traders who know little can read from the price what others know. Charles Plott and Shyam Sunder tested it in laboratory security markets. In their August 1982 Journal of Political Economy paper, insiders whose identity was hidden knew the state, and prices revealed it. In their September 1988 Econometrica paper, no trader knew the state, but pooled clues would identify it. Prices then revealed the state in some market designs and not in others. The difference is the main lesson for anyone testing traders, human or AI, on private news.

Two models of what a price reveals

Plott and Sunder set two models against each other. In a fully revealing rational-expectations (RE) equilibrium, prices reveal the state, because traders' expectations must be consistent with the prices they see. In a prior-information (PI) equilibrium, traders do not learn from price. Each acts only on what that trader already knows, and price follows from supply and demand.

The 1982 markets were oral double auctions among students at the University of Chicago Graduate School of Business. Each period, a one-period security paid a dividend that depended on a randomly drawn state and differed by trader type. Most markets had 12 traders of three types, each starting a period with two certificates and 10,000 francs of working capital. In most periods, two of the four traders of each type received a clue card naming the state, and the rest got blank cards, so insiders could not be identified.

Insiders: prices revealed the state

In market 1, insiders received only a noisy sample of 10 draws, and the results gave little support to RE. Since the authors' own expectations leaned toward PI, they gave insiders perfect information in market 2 to give RE its best chance. With that and later replications, prices adjusted immediately to near the RE price. Insiders' profits were virtually indistinguishable from those of noninsiders, and efficiency converged to near 100 percent. Market 5 raised the number of states from two to three, and RE still predicted well. In two periods of market 2, every trader got a clue so traders could learn which prices went with which state, but later markets showed this training was unnecessary.

The authors conclude that, given time and replications, these markets behaved substantially as RE models predict. They did not identify which variables carried the information, but suspected that unaccepted bids and offers, in addition to price, were of primary importance. They also state that they did not evaluate the models as applied to U.S. securities markets.

Diverse clues: aggregation depended on design

The 1988 study ran eleven markets with three states. If the state was x, half the traders were told "not y" and half "not z." No one knew the state, but the clues together identified it. Results split by series. In Series A, a single security paid different dividends to different traders, and RE predicted prices poorly. Series A efficiency averaged only 47 percent, and was lower when traders were more experienced. In Series B, traders had a complete set of state-contingent claims, and efficiency was around 90 percent. In Series C, a single security paid all traders alike, and RE was the best model.

Two cautions follow from the authors' conclusions. Series A markets passed fair-game tests, since trading rules did no better than buy and hold, yet they were not efficient. And some knowledge of other traders' preferences may be necessary for prices to aggregate.

Later tests of the result

Forsythe and Lundholm, in the March 1990 Econometrica, found that trading experience and common knowledge of dividends were jointly sufficient for an RE equilibrium, but neither was sufficient alone. Corgnet, Deck, DeSantis, Hampton and Kimbrough reported a direct replication in January 2020 that failed to reproduce the 1988 support. A new treatment structured like a prediction market did find strong support for RE. Page and Siemroth, pooling many market experiments in which each trader's information was known, give a broader estimate in the 2021 Review of Financial Studies. Public information was almost completely reflected in prices, but less than 50 percent of private information was. In the standard terms, prices were consistent with semistrong efficiency, which concerns public information, while possibly very far from strong-form efficiency, which includes private information.

AI traders with private signals

Spyros Galanis gave AI agents private signals and had them trade in prediction markets under four information structures of rising difficulty. Eight models, including GPT-4o and Claude Haiku 3.5 and 4.5, formed twelve teams of three traders, and the first wave, in January 2026, ran 1,772 markets. The median market priced the correct outcome at almost 1 in the Easy and Medium structures, 0.73 in Hard, and 0.5 in Very Hard. Galanis notes that most human subjects in related experiments reason up to two steps.

Cheap talk, market length, and strategic prompting did not improve accuracy, nor did feedback on past performance. In 99 percent of markets, agents on average hoarded information, posting public messages shorter than and detached from their private reasoning. A third wave of 576 markets in April 2026, with GPT-5.4, Claude Opus 4.6, and Gemini 3.1 Pro, aggregated better in the three easier structures but not the hardest, where confidently wrong markets gave way to markets hedging near 0.5.

What to check

• Did any trader know the state, as insiders did in 1982, or did aggregation require pooling clues, as in 1988?
• Which instruments were traded: one security, or a complete set of state-contingent claims?
• Did traders know others' dividends or preferences, which the 1988 authors suspect may be necessary?
• Is accuracy scored against the known state, or only by fair-game tests that inefficient markets passed?
• How hard was the information structure, and was the result rerun with newer models?

Azcuna's CALIBER puts AI agents in practice markets with play money and a news feed whose truth is known, and measures from their trades whether they are gullible, follow the crowd, overpay, or can be steered. It was proposed to DARPA under SBIR Phase I topic DPA26BZ06-DV026, and the proposal was submitted on October 21, 2026.

Sources

1. 1Efficiency of Experimental Security Markets with Insider Information: An Application of Rational-Expectations Models · Plott and Sunder, Journal of Political Economy 90(4), August 1982 (CaltechAUTHORS)
2. 2Rational Expectations and the Aggregation of Diverse Information in Laboratory Security Markets · Plott and Sunder, Econometrica 56(5), September 1988 (CaltechAUTHORS)
3. 3Information Aggregation in an Experimental Market · Forsythe and Lundholm, Econometrica 58(2), March 1990 (RePEc record and abstract)
4. 4Reconsidering Rational Expectations and the Aggregation of Diverse Information in Laboratory Security Markets · Corgnet, Deck, DeSantis, Hampton and Kimbrough, Chapman University ESI Working Paper 20-03, January 2020
5. 5How Much Information Is Incorporated into Financial Asset Prices? Experimental Evidence · Page and Siemroth, Review of Financial Studies 34(9), 2021 (RePEc record and abstract)
6. 6Information Aggregation with AI Agents · arXiv (Spyros Galanis), version 4, September 2026

Where this shows up in our work

• CALIBER, the market testbed for AI agents measures from an AI agent's trades whether it is gullible, herds, overpays, or can be steered.
• LEDGER, auditable Army acquisition decisions turns a trade study into a package where every number cites its source and AI drafts but never decides.