9 Real-World Examples of Prediction Markets in Action

Prediction markets are usually discussed as an abstract concept: "wisdom of crowds," information aggregation, efficient markets. That's all true but unhelpful if you want to understand what these things actually do in practice.

Here are 9 real cases — government agencies, Fortune 500 companies, and public markets — where prediction markets were deployed on real questions and produced real results. Some worked brilliantly. Some didn't. All of them tell you something about where this tool works and where it falls short.

From 2011 to 2015, the Intelligence Advanced Research Projects Activity (IARPA) — the intelligence community's research arm — ran a forecasting tournament called ACE (Aggregative Contingent Estimation). The goal was to determine whether crowdsourced forecasting could outperform professional intelligence analysts on geopolitical questions.

Five research teams competed. Each team recruited forecasters, developed aggregation methods, and submitted probability estimates on roughly 500 questions over four years. Questions covered geopolitics, economics, and military events: "Will North Korea conduct a nuclear test before April 2013?" "Will the price of oil exceed $110/barrel by June 2014?" "Will there be a military conflict between China and the Philippines in the South China Sea?"

Philip Tetlock's team at the University of Pennsylvania — the Good Judgment Project — won decisively. Their top forecasters beat the control group (intelligence analysts with access to classified data) by 30% on Brier scores. They beat the other four competing teams by 40-60%.

The implications rattled the intelligence community. IARPA's own research showed that a crowd of informed volunteers, reading publicly available information, could consistently outperform the billion-dollar intelligence apparatus. The finding wasn't that analysts were bad — it was that structured, incentivized crowd forecasting was better at integrating diverse information.

Outcome: IARPA continued investing in crowdsourced forecasting. Good Judgment Inc. was spun off as a commercial company and now sells forecasting services to intelligence agencies, the State Department, and private-sector clients. Several intelligence agencies launched internal prediction markets based on the ACE model.

From 2005 to at least 2010, Google operated one of the most well-documented internal prediction markets in corporate history. Employees traded play-money contracts on questions relevant to Google's business: product launch dates, quarterly revenue targets, competitor actions, and technology milestones.

A 2010 study by Cowgill, Wolfers, and Zitzewitz analyzed data from Google's market covering 1,072 questions and found three key results:

1. The market was well-calibrated. Events priced at 80% probability occurred approximately 80% of the time. This held across question types — technical, business, and miscellaneous.

2. Traders showed small but measurable biases. Google employees were slightly overoptimistic about Google's own products (e.g., launch dates that ended up slipping) and slightly overconfident on technology questions. The bias was small enough that the aggregated market price was still more accurate than individual expert estimates.

3. The market surfaced information that management didn't have. On several occasions, internal market prices on product timelines diverged from official project schedules weeks before the project team formally announced a delay. The market was aggregating private information from engineers who knew the project was behind but hadn't yet escalated the issue.

Outcome: Google used the market data to improve project management and resource allocation. The study became one of the most cited papers in the corporate prediction market literature. Google's exact current status is unclear — they don't discuss internal tools publicly — but several former Googlers have confirmed that prediction-market-style tools remain part of Google's internal toolkit.

In the mid-2000s, Hewlett-Packard ran internal prediction markets to forecast printer sales. The results, published by Chen and Plott (2002), showed that market-generated forecasts were more accurate than HP's official sales forecast — produced by a team of analysts with full access to pipeline data, dealer inventories, and historical trends — in 6 out of 8 quarters tested.

The margin wasn't small. The prediction market's mean absolute percentage error was 23% lower than the official forecast. HP's forecasting team had better data, more experience, and more context. But the market, which aggregated estimates from sales reps, product managers, and supply chain staff, incorporated dispersed ground-level information that the centralized team missed.

Intel ran a similar experiment from 2003 to 2006, using internal markets to forecast product demand, die yield rates, and technology adoption timelines. Intel's results (Hopman, 2007) showed comparable improvements: market forecasts outperformed the official planning forecast in 70% of comparisons.

Why this worked: In both companies, the relevant information was dispersed across hundreds of employees in different roles (sales, engineering, supply chain, customer support). No single team had the complete picture. The market created a mechanism to aggregate all of that distributed knowledge into a single number.

Outcome: Both companies used the results to supplement (not replace) their traditional forecasting processes. Neither fully adopted prediction markets as a primary planning tool, partly due to organizational resistance ("you're telling me a betting market is better than my team?") and partly due to practical challenges around participation rates and market design.

The 2024 presidential election was the event that moved prediction markets from niche financial tools to mainstream public awareness. Polymarket's election dashboard was embedded in Bloomberg terminals, cited by the Wall Street Journal, and checked obsessively by political operatives in both campaigns.

Key moments:

June 27: Biden's debate performance. Within 2 hours of the debate ending, Polymarket's "Will Biden be the Democratic nominee?" contract dropped from $0.84 to $0.65. By the next morning, it was at $0.55. The market processed the debate's implications faster than any poll or pundit panel.

July 21: Biden withdrawal. Polymarket had Biden below 50% for three weeks before the announcement. Kamala Harris contracts surged from $0.15 to $0.82 within hours of Biden's endorsement.

October: A French trader (identified by the WSJ as "Theo") placed over $30 million in bets on Trump across multiple Polymarket accounts. This raised questions about market manipulation, but the bets ultimately proved correct. The incident sparked a debate: was Theo a manipulator distorting the market, or an informed trader who happened to be right?

November 5: Polymarket had Trump at 62% on election morning. He won with 312 electoral votes. The final market price was more accurate than every major polling average and every major forecasting model.

Outcome: Total 2024 election volume on Polymarket exceeded $3.5 billion. The event established prediction markets as a serious alternative to polls for election forecasting and prompted multiple congressional hearings on prediction market regulation.

In late March 2020, the question on Metaculus was simple: "When will a COVID-19 vaccine receive Emergency Use Authorization in the United States?" At the time, the expert consensus ranged from 12 to 18 months, with many virologists arguing that historical base rates (average vaccine development: 10-15 years) made anything under 18 months dangerously optimistic.

Metaculus community forecasters submitted their first estimates in early April 2020. The median prediction: December 2020. Eight months away.

This was not wishful thinking. Metaculus forecasters decomposed the problem into sub-questions: How fast could mRNA technology produce a candidate? (Answer: weeks — Moderna's candidate was designed in January 2020.) How much could Operation Warp Speed compress trial timelines? (Answer: significantly, by running Phase 2 and Phase 3 concurrently and pre-manufacturing doses before approval.) How quickly could the FDA review data? (Answer: within weeks under EUA authority, versus 10 months for a standard BLA.)

Each sub-question pointed toward late 2020. The community median reflected this decomposed analysis, not a gut feeling or a hope.

Outcome: Pfizer-BioNTech received EUA on December 11, 2020. Metaculus's median forecast, submitted 8 months in advance, was within 11 days of the actual date. This remains one of the most impressive forecasting results in Metaculus's history and is frequently cited in arguments for structured forecasting over expert intuition.

The CME Group's FedWatch tool has been tracking Fed rate probabilities using fed funds futures since the early 2000s. It's the most widely used prediction market for monetary policy, and its accuracy record is remarkable: over the past decade, FedWatch correctly predicted the direction of the Fed's decision at every single FOMC meeting when the implied probability exceeded 80% one week before the announcement.

Kalshi entered this space in 2023 with explicit binary contracts on Fed rate decisions. The interesting finding: Kalshi's prices track CME FedWatch probabilities almost exactly. On 14 of 16 FOMC meetings from July 2023 through January 2026, the two platforms were within 3 percentage points of each other the day before the decision.

This convergence matters because it suggests that prediction market accuracy on financial events holds up across platforms and market structures. CME fed funds futures are traded by institutional fixed-income desks. Kalshi contracts are traded by a mix of retail traders, finance hobbyists, and a growing number of professionals. Different participant pools, same conclusion.

The exceptions: Markets diverge most when the Fed faces a true 50/50 decision. In November 2024, FedWatch had a 25bp cut at 53% and a hold at 47% — essentially a coin flip. Kalshi had the cut at 56%. Both were within the margin of noise. The Fed cut 25bp. Nobody deserves credit for calling a coin flip.

Outcome: Fed rate prediction markets have become so accurate that they've changed how central bankers communicate. Fed officials now track market-implied probabilities and use speeches to adjust expectations. When market pricing diverges too far from the Fed's intended path, a governor will give a speech that nudges expectations. The market and the Fed are, in effect, having a public conversation through prices and speeches.

On June 23, 2016, more than GBP 200 million was traded on Betfair's Brexit market. It was the most-traded political event in the exchange's history. And it got the outcome wrong.

In the final days before the vote, Betfair's implied probability for Remain peaked at 88%. On referendum night, as early results trickled in, Remain briefly hit 94%. The final result: Leave 51.9%, Remain 48.1%.

What happened? Three factors converged:

1. Demographic bias. Betfair traders skewed urban, educated, and internationally connected — exactly the demographic most likely to support Remain. They projected their own social circles onto the broader electorate.

2. Anchoring on late polls. Several late polls showed a Remain swing, and traders anchored heavily on these surveys while discounting earlier polls that had Leave ahead. The market overweighted the most recent data rather than the full polling trajectory.

3. The "status quo bias." Markets tend to overweight the status quo in referendum-style votes. People assume that undecided voters break toward "keep things as they are." In Brexit, undecided voters broke for Leave — the opposite of the market's assumption.

Outcome: Brexit became the cautionary tale that every prediction market skeptic cites. It's a legitimate failure. But it's also worth noting that Betfair's pre-campaign odds (before the final week) were closer to the result than post-campaign polls. The market failure was concentrated in the final 5-7 days, when a confluence of biases pushed prices away from the true probability. The broader lesson: markets are not infallible, and they fail in specific, predictable ways.

Every awards season, prediction markets become the film industry's unofficial scoreboard. PredictIt, Polymarket, and Metaculus all run Oscar prediction markets, and they consistently outperform expert panels and critic polls.

The data is striking. From 2015 through 2024, prediction market prices for Best Picture correctly identified the winner before the ceremony in 8 out of 10 years. For acting categories, market accuracy was even higher — correctly picking the winner in 36 out of 40 individual acting awards (90%). Entertainment journalists, by comparison, correctly picked Best Picture 7 out of 10 years and acting awards 33 out of 40 (82.5%).

Why are markets good at Oscars? Because the information is dispersed in a specific way. Academy voters are roughly 10,000 people spread across the industry. They talk — to agents, to journalists, to each other. No single journalist knows how the full Academy will vote, but collectively, traders who read entertainment news, follow industry gossip, track guild awards (SAG, DGA, PGA), and monitor social media chatter can assemble a surprisingly accurate picture.

The guild awards are particularly informative. The SAG ensemble award has correctly predicted the Best Picture winner 60% of the time. The PGA's top film award has matched Oscar 72% of the time since 2009. Smart prediction market traders weight these precursor awards heavily, creating a natural information cascade that converges on the likely winner weeks before the ceremony.

Outcome: Oscar prediction markets have become so accurate that some industry observers worry they're reducing the drama of awards season. When the market has one nominee at $0.85 and the rest below $0.10, the "surprise" is gone. Whether that's good or bad depends on whether you watch the Oscars for information or entertainment.

Climate prediction markets are newer than political or financial ones, but they're growing fast. Kalshi now lists contracts on hurricane landfalls, seasonal temperature anomalies, and specific weather events. Metaculus has an extensive climate forecast track with hundreds of questions on temperature projections, sea-level rise, and emissions trajectories.

The early results are interesting. On hurricane season predictions (number of named storms, major hurricanes), Kalshi contract prices have been within the range of NOAA's official forecast in 4 out of 5 seasons tracked. On specific landfall questions ("Will a Category 3+ hurricane hit Florida in 2025?"), Kalshi prices have roughly matched the implied probabilities from NHC seasonal outlooks.

Where prediction markets could add value beyond official forecasts is in economic impact estimation. NOAA tells you whether a hurricane is coming. Markets can tell you the probability-weighted economic impact. A contract like "Will insured losses from 2026 Atlantic hurricanes exceed $50 billion?" aggregates information from meteorologists, insurance analysts, and real estate experts into a single price. No government forecast attempts to answer that question.

The challenge is participation. Weather and climate markets attract far fewer traders than election or financial markets. A typical Kalshi weather contract might have $50,000-$200,000 in volume, versus $5-50 million for a comparable political contract. This limits accuracy — the wisdom of crowds requires a crowd.

Longer term, climate prediction markets could play a role in infrastructure planning and insurance pricing. If a market consistently prices "Will sea level in Miami rise more than 6 inches by 2035?" at $0.70, that's a strong signal for urban planners and mortgage lenders. The information exists in climate models, but markets could make it more accessible and more frequently updated.

Outcome: Still early. Climate prediction markets are a promising application with low current volume but high potential value. The key bottleneck is attracting enough knowledgeable participants — climatologists, insurance actuaries, agricultural economists — to make the prices informative.

Given the evidence — Google, HP, Intel, IARPA — you'd expect every Fortune 500 company and government agency to run internal prediction markets. Most don't. Here's why:

Political risk to managers. If a prediction market says your project will be 3 months late, that's embarrassing for the project manager. If a market says Q3 revenue will miss the internal target, that's uncomfortable for the sales VP. Most organizations don't want a public scoreboard that can contradict official plans. Messenger-shooting is a real risk, and employees know it.

Thin participation. Internal markets need critical mass to produce useful prices. At Google (70,000+ employees), enough people traded to make prices informative. At a 500-person company, you might get 30 active traders — not enough for meaningful information aggregation. And the people with the best information (senior engineers, deal-closers) are often too busy to trade.

Legal and compliance concerns. In regulated industries (banking, healthcare, defense), internal betting markets raise compliance questions. Even play-money markets can create discovery risk in litigation. Legal departments tend to say "no" by default.

Organizational culture. Prediction markets require a culture that values accuracy over consensus and data over authority. Many organizations say they want this. Fewer actually do. When the market contradicts the CEO's stated forecast, the typical response is to shut down the market, not update the forecast.

These aren't technical problems. They're organizational and political problems, which are harder to solve. The companies that successfully use internal prediction markets tend to be ones where data-driven culture already exists (tech companies, quantitative hedge funds) and where senior leadership explicitly protects the market from interference.