Bubbles in the lab: experimental asset markets explained

An experimental asset market is a laboratory market in which traders buy and sell an asset whose dividends follow announced rules, so its fundamental value can be computed. Vernon Smith, Gerry Suchanek, and Arlington Williams built the standard version, reported in the September 1988 Econometrica. Fourteen of their twenty-two experiments produced price bubbles followed by crashes relative to the asset's intrinsic dividend value. Experience reduced the chance of a bubble but did not eliminate it. A 2009 paper says the design dominated experimental research on financial markets for the next two decades.

The declining fundamental value design

The 1988 abstract describes the environment. All traders receive the same dividend, drawn from a known probability distribution, at the end of each of 15 or 30 trading periods. Kirchler, Huber and Stöckl, in a 2009 Innsbruck working paper, summarize the consequence: the value is the expected value of a finite stream of dividends, so it falls deterministically as periods pass.

A 2003 working paper by Dufwenberg, Lindqvist and Moore shows the arithmetic in its own version. That paper's asset lives ten periods and pays 0 or 20 cents each period with equal probability, an expected 10 cents. For a risk-neutral trader, the value with k periods left is k times 10 cents, falling to 10 cents in the last period. The authors note that in the 1988 settings the experimenters controlled the dividend process and the asset's life and made both public, so values could be derived by backward induction.

The same 2009 paper points to two side effects. The value falls to zero within a set number of periods, and dividend yields are extreme, starting at 10 percent in the first period of a 10-period setting and reaching 33, 50, and 100 percent in the last three.

What happened to prices

In the bubble markets, prices rose above intrinsic dividend value and then crashed. Smith, Suchanek and Williams regressed the change in mean price on the previous period's excess bids, the number of bids minus the number of offers. They treated excess bids as a measure of demand driven by homegrown expectations of capital gains. When excess bids went to zero, prices converged to rational expectations, the point where arbitrage stops paying. They also interpret the bubble as a form of temporary myopia from which traders learn.

Smith's 2002 Nobel prize lecture draws the lesson that common information about value is not enough to produce common expectations, which traders reach through repetition. Smith says this reading fits asset trading experiments with undergraduates, small businesspersons, corporation managers, and over-the-counter traders. The only exceptions with inexperienced subjects Smith knew of came with advanced graduate students. The lecture puts it in one line: experimental asset markets bubble and crash "on the long path of experience to equilibrium."

The effect of experience

The 1988 abstract reports that experience reduces but does not eliminate the probability of a bubble. Four of twenty-six experiments, all with experienced subjects, appeared to converge early to rational expectations, though price fluctuations of a few cents remained. The abstract counts twenty-two experiments in one sentence and twenty-six in the other.

Dufwenberg, Lindqvist and Moore asked what happens when experience is mixed. In the working-paper design, run at Virginia Tech in October 2001, six subjects traded together for three rounds of ten two-minute periods. In a fourth round, two or four of them were replaced by newcomers, leaving two thirds or one third of the market experienced. Each treatment ran five sessions. At a 2002 conference, most of about thirty participants guessed that both mixed markets would bubble. Bubble-crash patterns were not common in either. The December 2005 article states the result: even with a minority of experienced traders, bubbles were substantially abated. The working paper adds that experienced traders earned more than newcomers.

Confusion, and machine traders

Kirchler and colleagues compared the declining-value design with markets whose value followed a random walk. Overvaluation averaged 50.5 percent of fundamental value in the declining-value markets and 5.4 percent in the random-walk markets. Asked what the value would be one period later, more than 60 percent of traders in the declining-value markets expected it to stay constant or rise, and only 35 percent answered correctly. The authors suggest that at least some of the mispricing reflects confusion about the value path, not only speculation.

Henning and colleagues, including Colin Camerer, ran language-model agents in a related design with a constant value of 14 over 30 periods. They tested six commercial models, with 20 copies of one model per market, against nineteen human markets. The models generally priced near fundamental value, with only a muted tendency to bubble, while the humans consistently produced bubbles. For the models tested, an experienced second session left results essentially unchanged. The authors advise caution in replacing human subjects with off-the-shelf agents.

What to check

• Does the fundamental value decline, as in the 1988 design, or stay constant, and was traders' understanding of its path tested?
• Were traders inexperienced, experienced, or mixed, and in what share, given the 2005 finding that a minority of experienced traders abated bubbles?
• How is a bubble counted, and over how many experiments, given that one 1988 abstract cites both 22 and 26?
• Do measures such as overvaluation as a percent of fundamental value come with the number of markets behind them?
• For language-model agents, does pricing near value reflect an understanding of the dividend path, or a fixed habit that experience leaves unchanged?

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. 1Bubbles, Crashes, and Endogenous Expectations in Experimental Spot Asset Markets · Smith, Suchanek and Williams, Econometrica 56(5), September 1988 (Econometric Society abstract page)
2. 2Constructivist and Ecological Rationality in Economics · Vernon L. Smith, Prize Lecture, December 8, 2002
3. 3Bubbles and Experience: An Experiment · Dufwenberg, Lindqvist and Moore, American Economic Review 95(5), December 2005
4. 4Bubbles and Experience: An Experiment on Speculation · Dufwenberg, Lindqvist and Moore, Research Institute of Industrial Economics Working Paper No. 588, January 2003
5. 5Bubble or no Bubble - The Impact of Market Model on the Formation of Price Bubbles in Experimental Asset Markets · Kirchler, Huber and Stöckl, University of Innsbruck Working Paper 2009-26, November 2009
6. 6LLM Agents Do Not Replicate Human Market Traders: Evidence From Experimental Finance · arXiv (Henning, Ojha, Spoon, Han and Camerer), version 3, October 2025

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.