The Policy Analysis Market: the DARPA prediction market...

In late July 2003, a small research program buried inside the Defense Advanced Research Projects Agency became, for about 48 hours, the most notorious idea in Washington. Two senators held a press conference. By the next morning the Pentagon had announced that the program was dead. Within roughly two weeks the retired admiral who ran the office responsible for it had left the agency. The program was the Policy Analysis Market, and its rise and collapse is one of the cleanest examples in the public record of a defensible research method wrecked by an indefensible application.

This entry in the "where the public money goes" series treats the Policy Analysis Market the way the record supports: not as proof that markets are foolish, and not as a simple morality tale about a rogue idea, but as a documented failure with two honest sides. The forecasting method at the core was and is legitimate. The way DARPA proposed to apply it, to a public market that appeared to let people profit from assassinations and terror attacks, was a judgment failure that no analytic upside could survive once it reached daylight.

What it actually was: FutureMAP and PAM

Two program names sit at the center of this story, and they are easy to conflate. The umbrella effort was called FutureMAP. The specific market inside it was the Policy Analysis Market, or PAM.

According to the encyclopedic record of the program, PAM "was a proposed futures exchange developed, beginning in May 2001, by the Information Awareness Office (IAO)" of DARPA, "based on an idea first proposed by Net Exchange, a San Diego, California, research firm," per Wikipedia's Policy Analysis Market entry. The Information Awareness Office is the same DARPA office associated with Total Information Awareness, the mass-data surveillance research effort that was already under heavy political fire. That shared parentage matters, because much of the reaction to PAM was really accumulated reaction to the office that housed it.

The concept was a prediction market, sometimes called a decision market. Traders would buy and sell futures contracts tied to political and economic developments in several Middle Eastern countries. The design premise, per the same record, is a well-established one: a futures contract's price tends to reflect the probability that the underlying event will occur, because participants bid it up or down according to the private information they hold. Aggregate enough of those bets and the market price can become a forecast, one that often outperforms individual expert opinion at predicting events like elections. DARPA was testing whether that mechanism could sharpen intelligence estimates about instability in the region.

That is the legitimate part, and it is important to state it plainly before describing what went wrong. The information-aggregation rationale behind prediction markets is not in dispute. The controversy was never about whether markets can forecast. It was about the subject matter DARPA chose to forecast, and how that choice looked in public.

The contracts critics seized on

When the program surfaced, the specific examples that circulated were the ones that made it politically radioactive. Senate floor statements from July 2003 listed sample contracts resembling bets on catastrophic and violent events. The Congressional Record reproduced by the Federation of American Scientists records senators citing scenarios such as the King of Jordan being overthrown, Yasser Arafat being assassinated, a North Korean missile attack, and a biological attack on Israel. These items were subsequently pulled from the program's website.

Here the record requires care, because this is the single most-disputed characterization in the whole episode. The "betting on individual assassinations and attacks" framing is the critics' framing, and it should be attributed that way rather than stated as settled fact. Program defenders, including economist Robin Hanson, who worked on it, argued that the individual-attack description was inaccurate and that PAM was aimed at broad geopolitical and economic indicators, not wagers on specific terror acts. Hanson has maintained the program was "largely misunderstood," and Net Exchange later proposed relaunching it privately with any violent-event contracts stripped out, per Hanson's own archive of the program at George Mason University, which should be read as a participant's advocacy account.

So the fair statement is this: critics pointed to sample contracts that looked like betting on assassinations and attacks, some of which were real enough to be pulled from the site; defenders say those examples misrepresented a program built around broad instability indicators. Both of those things are in the record, and a neutral account keeps both.

The timeline: a 24-hour death

The collapse was fast, and the dates are worth getting exactly right, because sloppy summaries blur them.

May 2001: Work on PAM begins inside DARPA's Information Awareness Office, based on the Net Exchange concept, per Wikipedia.
July 28, 2003: Senators Ron Wyden (Democrat of Oregon) and Byron Dorgan (Democrat of North Dakota) hold a press conference denouncing the program. Wyden called it "a federal betting parlor on atrocities and terrorism" that was "ridiculous and grotesque." Dorgan called the idea "useless, offensive and unbelievably stupid," per Roll Call's contemporaneous report. Wyden and Dorgan led the charge, and they are the names to attach to the reveal.
July 29, 2003: The morning after the press conference, the Pentagon announced the program was being cancelled, as reported by NBC News via the Associated Press. The encyclopedic record puts it bluntly: "Within less than a day, the Pentagon announced the cancellation of PAM." That same day, additional senators, including Tom Daschle and Dick Durbin, added their own floor statements, which is the record the Federation of American Scientists reproduces.
August 13, 2003: John Poindexter, the retired admiral who headed the Information Awareness Office, left DARPA, per Wikipedia.

The phrase "within a day" refers to the cancellation announcement, not to every administrative step. Registration halts and the takedown of the program's website, policyanalysismarket.org, followed over the next several days. But the decisive gap, the one that gave the episode its "24-hour death" nickname, was the roughly one day between two senators standing at a microphone and the Pentagon pulling the plug.

The money, as reported

The frequently cited figure is that PAM cost about $8 million through 2005. That number should be attributed, not presented as an audited total. It traces to Senator Durbin's characterization on the Senate floor, recorded in the Congressional Record on the FAS site, rather than to a neutral cost review. The same record notes that registration for up to 1,000 traders was set to begin that week, reported as Friday, August 1, 2003, with live trading intended to start later, on January 1, 2004, in program plans reported at the time.

Every one of those figures belongs in the "reported or planned" column, not the "executed" column, because the program was killed before it launched. The trading start date comes from program plans described at the time. The dollar figure and the trader count come from senators' statements. Treating them as senator-sourced and plan-stage, rather than as settled fact, is the honest way to carry them.

This also reframes what the efficiency critique of PAM is really about. The critique is not that $8 million is a large sum by the standards of federal research. It is not. The critique is that the entire investment was guaranteed to be thrown away the moment the concept met public scrutiny, because the packaging ignored the political and moral optics completely. A research bet can fail on the science and still teach something. PAM failed on presentation before it ever produced a forecast.

Poindexter, and who not to blame

It is tempting to make John Poindexter the villain of this story. The record does not support that, and the fact base for this piece flags it directly as a landmine.

Poindexter's departure on August 13, 2003 was real, and the PAM fallout contributed to it. But it was driven by the broader backlash against the Information Awareness Office, and against Total Information Awareness in particular, not by PAM alone. Wikipedia notes he left "insisting his controversial anti-terrorist program was largely misunderstood," a defense he applied to his office's work generally. Program defenders further argue that Poindexter had limited direct involvement in PAM's specific design. So the accurate framing is that a small, ethically clumsy market was the final flashpoint in a much larger controversy over an office already viewed by many in Congress as a surveillance overreach. PAM was the spark; the fuel had been accumulating for months.

The ROI case and the honest caveats

Because this is a failure rather than a success, the balanced closing is not "the return on a high-risk bet." It is the both-sides reading the record demands: a legitimate method attached to an indefensible use.

The case for the underlying method

The forecasting tool at PAM's core is genuinely defensible, and it did not disappear when the program did.

Prediction markets aggregate dispersed information. The core claim, that a market price can combine the private knowledge of many participants into a probability estimate that often beats individual experts, is well established in academic work and in practice.
The method is used widely and legitimately. Corporate internal forecasting markets, election markets, and academic research on market-based forecasting all rely on the same mechanism. None of them are considered scandalous.
The intelligence community itself examined the idea. The CIA's in-house journal, Studies in Intelligence, later published a piece titled "Using Prediction Markets to Enhance US Intelligence Capabilities," available as a CIA PDF from 2006. That the CIA's own analysts studied prediction markets seriously, after the PAM blowup, underlines that the method was never the problem.

The honest caveats

None of that vindicates PAM as it was proposed.

The application was indefensible in public. Whatever the internal design intent, a government market that appeared to let people profit from assassinations, coups, and terror attacks was untenable the moment it was described in plain language. The method may be legitimate; this use of it was not.
The optics were fatal and predictable. Two senators and one press conference were enough to end it, which is the clearest possible evidence that the concept had no defensible public framing to fall back on.
The disputed characterization cuts both ways. Defenders are right that the "betting on individual attacks" description was contested and that some cited contracts were about broad indicators. But the program still shipped, on a public site, examples inflammatory enough that they had to be pulled. Even granting the defenders' best case, the presentation was the failure.

The lesson the record supports is narrow and specific: right tool, indefensible use. It is not "prediction markets are stupid," and anyone drawing that conclusion is overreading a story that is really about judgment, optics, and the difference between a method and its subject matter.

Fact-check notes and sources

PAM was a DARPA program inside the Information Awareness Office, under the FutureMAP umbrella, begun in May 2001 from a Net Exchange concept. Per Wikipedia, Policy Analysis Market, which states PAM was "part of the FutureMAP project" and "a proposed futures exchange developed, beginning in May 2001, by the Information Awareness Office (IAO)" based on "an idea first proposed by Net Exchange." FutureMAP is the umbrella; PAM is the specific market.
The design premise was that market prices aggregate dispersed information into forecasts. Documented as the standard prediction-market rationale in the same Wikipedia entry. The method is not in dispute; the application was.
Critics cited sample contracts resembling bets on assassinations, coups, and attacks (Arafat, the King of Jordan, a North Korean missile attack, a biological attack on Israel), later pulled. From Senate floor statements in the Congressional Record reproduced by the Federation of American Scientists. Presented here as the critics' characterization; program defenders dispute the individual-attack framing.
Publicly denounced July 28, 2003 by Senators Ron Wyden (D-OR) and Byron Dorgan (D-ND), with the "federal betting parlor" and "useless, offensive and unbelievably stupid" quotes. Per Roll Call. Daschle and Durbin joined on July 29; Wyden and Dorgan led.
The Pentagon cancelled it the next day, July 29, 2003, the "24-hour death." Per NBC News / Associated Press and Wikipedia, which states the cancellation came "within less than a day."
John Poindexter left DARPA on August 13, 2003; the PAM fallout contributed but the broader IAO / Total Information Awareness backlash drove it. Per Wikipedia. Defenders say he had limited direct involvement in PAM's design.
Reported cost of about $8 million through 2005, up to 1,000 traders, registration set for around August 1, 2003, trading planned for January 1, 2004. The $8 million figure and trader count come from senators' floor statements (Durbin and Daschle) in the FAS Congressional Record; the January 2004 trading date is a reported program plan. Attributed as reported and planned, since the program was killed before launch, not as audited or executed figures.
Prediction markets are a legitimate forecasting tool the CIA's own journal examined. Per the CIA's Studies in Intelligence piece "Using Prediction Markets to Enhance US Intelligence Capabilities," available as a CIA PDF. This supports the method in general, not PAM's terrorism-market application.
Program-defender account. Robin Hanson's George Mason University archive preserves the "largely misunderstood" defense and Net Exchange's private-relaunch proposal. Treat as a participant's advocacy account.

The Policy Analysis Market: the DARPA prediction market cancelled within a day