Corporate Prediction Markets

Yes, some of the world’s biggest companies have run prediction markets internally: Google, Ford, and Hewlett-Packard among them, letting their own employees bet on launch dates, product demand, and project risk. The results were good enough to be genuinely useful, and limited enough to explain why the practice never became standard.

An internal, or corporate, prediction market works like any other market, except the traders are employees and the questions are about the company itself. Will this feature ship on time? How many units will we sell next quarter? Prices become probabilities, and the crowd of insiders often knows things that no single manager does.

What companies forecast internally

The appeal is simple. A large organization holds a huge amount of scattered knowledge, and most of it never reaches the people making decisions. An internal market is a way to pull that knowledge into one number. In practice, firms have used them for three kinds of question.

• Launch and ship dates: will a given feature, product, or office actually be ready by the target date? Engineers close to the work usually know before management does.
• Demand and sales: how many units will sell next quarter, or how will a new product be received? This is the classic use, because sales forecasts are both valuable and hard.
• Project and operational risk: will a milestone be hit, a supplier deliver, or a metric land inside its range? Markets surface the doubts people are reluctant to raise in a meeting.

The Google, Ford, and HP experiments

Google ran internal markets for years, starting in 2005. Employees traded on questions such as product launch dates, office openings, and demand, using a play-money currency they could later exchange for prizes. Ford ran markets to help forecast vehicle sales and gauge interest in features. Both generated a rich data trail that researchers were later able to study.

The definitive analysis is by Bo Cowgill and Eric Zitzewitz, who studied internal markets at Google, Ford, and a large anonymous software firm and published “Corporate Prediction Markets: Evidence from Google, Ford, and Firm X” in the Review of Economic Studies in 2015.

Hewlett-Packard ran some of the earliest experiments of all. A small group of employees, sometimes only a dozen or so, traded on future printer sales, and the market forecast frequently beat the company’s own official forecast. That HP result, from work in the late 1990s and early 2000s, is a large part of why the idea spread in the first place.

What the results actually showed

The headline is encouraging. Across the markets Cowgill and Zitzewitz studied, prices were relatively efficient and improved on the forecasts experts were already making, by up to a 25% reduction in mean squared error. In plain terms, the crowd of employees was measurably more accurate than the existing process, not by a little on a lucky question, but on average across many.

The markets were also reasonably well calibrated: outcomes priced around 70% tended to happen about 70% of the time, the same property that makes public markets trustworthy.

That calibration is the heart of why markets forecast well at all. For the full argument, see why prediction markets are accurate.

The biases they revealed

The same study is honest about the flaws, and the most notable one was an optimism bias at Google. Prices for good outcomes, an early launch or a strong number, ran a little too high. The effect was strongest among newer employees, on questions tied to their own work, and on days when Google’s stock had risen, which is exactly the pattern you would expect if mood and self-interest were leaking into the prices.

Encouragingly, the bias shrank as employees gained trading experience, and the markets became more accurate over time. Optimism is also a manageable flaw: once you know prices lean high, you can discount them. It is a smaller problem than a forecast that is confidently wrong for reasons nobody can see.

Why corporate prediction markets stayed rare

If internal markets beat the experts, why is almost no company running one today? The honest answer is that the conditions that make a market accurate are hard to reproduce inside a firm.

• Weak incentives: internal markets usually run on play money or small prizes, so a busy employee has little reason to research a question and trade it seriously.
• Thin participation: a market needs a critical mass of informed, independent traders. Most teams are too small to supply one, and the same handful of people tend to dominate.
• Secrecy and politics: a market that predicts the flagship will slip is an awkward thing to have on a screen, and aggregating what insiders really think can reveal information management would rather keep quiet.
• No obvious owner: the forecast competes with the plans people are already paid to defend, so it can feel like a threat rather than a tool.

Some of these problems are technical rather than human. Thin markets, in particular, are usually run with an automated market maker so trades can happen even when few people are online, a design that comes from Robin Hanson’s market scoring rules. It solves the plumbing, but it cannot manufacture the informed, motivated traders a good forecast needs.

Taken to its logical conclusion, the idea becomes a way to run an organization, not just forecast for one. That proposal has a name and its own guide: futarchy explained.

The takeaway

Corporate prediction markets are neither a failed gimmick nor a magic oracle. The evidence, from real markets at Google, Ford, and HP, is that they genuinely can beat expert forecasts, with known and manageable biases. They stay rare not because they do not work, but because running one well demands active, honest, independent participation that most companies struggle to provide.

The easiest way to understand what makes these markets tick is to trade in one yourself. Clutch lets you predict real news and sports with in-app credits and keeps score, so you can feel the incentives that make a market accurate. Get the app and try it.