28 Dec 2024 Search

Taylor Lorenz writes, in Prediction markets go mainstream,

Prediction markets—platforms where users buy and sell shares based on the probability of future events—are poised to disrupt the media landscape in 2025, transforming not only how news is shared but how it is valued and consumed.

Prediction markets did get some time in the spotlight this year. But the reasons for the long, ongoing prediction market winter are bigger than just prediction markets not being famous. Prediction markets have been around for a long time, and have stubbornly failed to go mainstream.

The first prediction market to get famous was the University of Iowa’s Iowa Electronic Markets which launched in the late 1980s and has been covered in the Wall Street Journal since at least the mid-1990s. They originally used pre-web software and you had to mail in a paper check (update 4 Jan 2025: paper checks are still the only way to fund your account on there). But IEM wasn’t the first. Prof. Robin Hanson, in Hail Jeffrey Wernick, writes about an early prediction market entrepreneur who started his first one in 1981. (A secretary operated the market manually, with orders coming in by fax.) Prediction markets were more famous than Linux or the World Wide Web before Linux or the World Wide Web. Prediction markets have been around since before stop trying to make fetch happen happened.

So the safe prediction would be that 2025 isn’t going to be the year of prediction markets either. But just like the year of Linux on the desktop never happened because the years of Linux in your pocket and in the data center did, the prediction markets that do catch on are going to be different from the markets that prediction market nerds are used to today. Some trends to watch are:

Payment platforms: Lorenz points out, Prediction markets are currently in legal limbo, but I’d bet against a ban, especially given the new administration. Right now in the USA there is a lot of VC money tied up in fintech, and a lot of political pressure from well-connected people to deregulate everything having to do with money. For most people the biggest result will be more scams and more hassles dealing with transactions that are legal and mostly trustworthy today but that will get enshittified in the new regulatory environment. But all those money-ish services will give prediction markets a lot more options for getting money in and out in a way that enables more adoption.

Adding hedging and incentivization: The prediction markets that succeed probably won’t be pure, ideal prediction markets, but will add on some extra market design to attract and retain traders. Nick Whitaker and J. Zachary Mazlish, in Why prediction markets aren’t popular, write that so far, prediction markets don’t appeal to the kinds of people who play other kinds of markets. People enter markets for three reasons. Savers are trying to build wealth, Gamblers play for thrills, and Sharps enter to profit from less well-informed traders. No category out of the three is well-served by existing prediction markets, because a prediction market is zero-sum, so not a way to build wealth long-term, and it’s too slow-moving and not very thrilling compared to other kinds of gambling. And the sharps need a flow of less well informed traders to profit from, but prediction markets don’t have a good way to draw non-sharps into the market.

Whitaker and Mazlish do suggest hedging as a way to get more market participants, but say

We suspect there is simply very little demand for hedging events like whether a certain law gets passed; there is only demand for hedging the market outcomes those events affect, like what price the S&P 500 ends the month at. Hedging market outcomes already implicitly hedges for not just one event but all the events that could impact financial outcomes.

That’s probably true for hedging in a large public prediction market. An existing oil futures market is more generally useful to more traders that a prediction market on all the events that might affect the price of oil. And certain companies’ stocks today are largely prediction markets on future AI breakthroughs and the future legal status of various corporate crimes. But I suspect that it’s different for a private market for events within a company or organization. For example, a market with sales forecasting contracts on individual large customers could provide much more actionable numbers to management than just trading on predicted total sales.

You could, in effect, pay for a prediction market’s information output by subsidizing it, and Whitaker and Mazlish suggest this. A company that runs an internal prediction market can dump money in and get info out. Like paying for an analyst or consulting firm, but in a distributed way where the sources of expertise are self-selecting by making trade/no trade decisions based on what they know or don’t know. But it’s also possible, usually on the smaller side, for a prediction market to become an incentivization market. To me, the difference is that in an incentivization market, a person with ability to affect the results holds a large enough investment in the market that it influences them to do so. The difference is blurry and the same market can be a prediction market for some traders and an incentivization market for others. But by designing incentives for action in, a market operator can make it drift away from a pure prediction market design to one that tends to produce an outcome. related: The private provision of public goods via dominant assurance contracts by Alexander Tabarrok

Proof of concept projects can already address specific information needs: A problem that overlaps with the prediction market incentivization problem in interesting ways is the problem of how to pay for information products and services that can be easily copied. How do we fund open source? is a persistent question. And Bruce Perens, original author of what became the Open Source Definition, wants to move on entirely. The problem of funding open source is hard enough that we mainly hear about it when a high-profile security issue makes the news.

As Luis Villa points out,

If you don’t know what’s in the box, you can’t secure it, so it is your responsibility as builders to know what’s in the box. We need better tools, we need better engagement to enable everybody to do that with less effort and less burden on individual volunteer maintainers and non-profits.

Companies that use open source software need to measure and reduce risks. The problem is that the biggest open source risks are related to hard-to-measure human factors like developer turnover and burnout. Developers of open source software can take actions that help companies understand their risks, but they’re not compensated for doing it. A prediction/incentivization market can both help quantify hidden risks and incentivize changes.

If you have an internal market that functions as both a prediction market and an incentivization market, you can subsidize both the information and the desired result by predicting the events that you don’t want to happen. This is similar to how commodities markets and software bug futures markets can work. Some traders are pure speculators, others take actions that can move the market. Farmers can plan which crops to plant based on predicted or contracted prices, companies can allocate money to fuel futures and/or fuel-saving projects, developers can prioritize tasks.

Synergy with AI projects: An old corporate Intranet rule of thumb [citation needed] is that you need five daily active editors to have a useful company or organization Wiki. I don’t know what the number is for a prediction market, but as Prof. Andrew Gelman points out, prediction markets need “dumb money” to create incentives for well-informed traders to play and win.

Noisy, stupid bots are a minus for most kinds of social software, but a win for markets. If only there were some easy way to crank up a bunch of noisy, stupid bots. Oh, wait, there’s a whole AI boom happening. Good timing, right? And AI projects need ways to test their output quality in a scalable way, just as much as prediction markets need extra trading churn. AI projects and prediction market projects solve each other’s problems.

  • Prediction markets need liquidity and dumb money. Bots can already do those.

  • AI projects need scalable quality checks. Slop is easier to make than to check, so evaluating the quality of AI output keeps growing relative to the declining costs of everything else. You can start up a lot of bots, fund each with a small stake, and shut down the broke ones. The only humans required are the traders who can still beat the bots. and if at some point the humans lose all their money, you know you won AI. Congratulations, and I for one welcome our bot plutocrat overlords.

Bots can also be run behind a filter to only make offers that, if accepted, would further the market operator’s goals in some way. For example, bots can be set up to be biased to over-invest on predicting unfavorable outcomes (like buying the UNFIXED side of bug futures) to add some incentivization.

Fixing governance by learning from early market experiences: Internal prediction markets at companies tend to go through about the same story arc. First, the market launches with some sponsorship and internal advocacy from management. Second, the market puts up some encouraging results. (Even in 2002 a prediction market was producing more accurate sales forecasts than the official ones at HP.) And for its final act, the prediction market ends up perpetrating the unforgivable corporate sin: accurately calling some powerful executive’s baby ugly. So the prediction market ends up going to live with a nice family on a farm. Read the (imho, classic) paper, Corporate Prediction Markets: Evidence from Google, Ford, and Firm X by Bo Cowgill and Eric Zitzewitz, and, in Professor Hanson’s post, why a VC firm could not get prediction markets into portfolio companies. Wernick blames the ego of managers who think their judgment best, hire sycophants, and keep key org info close to their chests.

The main lesson is that the approval and budget for the prediction market itself needs to be handled as many management levels as possible above the managers that the prediction market is likely to bring bad news to. Either limit the scope of issues traded on, or sell the market to a more highly placed decision maker, or both. The prediction market administrator needs to report to someone safely above the level of the decision-makers for the issues being traded on. The really interesting experiment would be a private equity or VC firm that has its own team drop in and install a prediction market at each company it owns. The other approach is bottom-up: start with limiting the market to predicting small outcomes like the status of individual software bugs, and be disciplined about not trading on more consequential issues until the necessary sponsorship is in place.

So, is 2025 the year of prediction markets? Sort of. A bunch of factors are coming together. Payment platform options, the ability to do proof of concept niche projects, and the good fit as a QA tool for AI will make internal market projects more appealing in 2025. And if market operators can learn from history to avoid what tends to happen to bearers of bad news, this could be the year. More: On Prediction Market Sales Engineering