Issues In Information Market Design

by Robin Hanson

Information markets (a.k.a. prediction markets and idea futures) are markets whose primary purpose is to aggregate information, instead of to hedge risk or entertain. This article reviews many issues in the design of such markets.


A common task is how to induce people to collect information on chosen topics, and to reveal what they know to each other, in order to produce consensus estimates on those topics. If people were always open and honest with themselves and others, and believed that others were this way as well, the ideal institution for this task would probably be simple conversation. Unfortunately, this ideal rarely holds.

Information markets are institutions for accomplishing this task via trading in speculative markets, at least on topics where truth can be determined later. Relative to simple conversation, it is harder to say any one thing by market trading, and there are fewer things one can say. When people do not know whom to believe, however, the fact that a trader must “put his money where his mouth is” means that the things people do say in information markets can be more easily believed. Those who know little are encouraged to be quiet, and those who think they know more than they do learn this fact by losing.

The expense and expressiveness limitations of markets relative to ordinary conversation suggest these design issues: how can we lower the costs of trading, how can we expand the set of things that one can say by trading, and which questions should we ask of traders? There are also many other potential problems to be considered.

Every information market asks some question. A simple market where one kind of asset is traded for another kind of asset asks: “What is the relative value of these assets?” For appropriately defined assets, this can be equivalent to “What is the probability of this event?” or “What is the expected value of this parameter?” Markets can also ask for the conditional probabilities and conditional expectations.

The cost of asking questions via markets seems to depend more on the number of questions asked than on which questions are asked. This suggests that we ask the questions that can most influence our most important decisions. And this suggests that we estimate important outcomes conditional on our big decisions. If what we really wanted to know is whether Saudi oil supplies are more likely to be interrupted if we move U.S. troops out of Saudi Arabia than if we do not, we might just directly ask this question.

One potential problem with asking such questions is that market prices can fail to reflect what traders know when traders suspect that decision makers will know even more. This problem is avoided if decision makers and their advisors can trade, if the decision time is known to traders, and if we only rely on prices just before this decision time.

Another potential problem is that people may lie to gain favorable decisions. Laboratory experiments on price manipulation have so far not found this to be a problem. A theoretical model can also be constructed, by adding agents who care directly about the market price to standard market microstructure models. Agents with private information about how strongly they want to bias the price act much like liquidity traders with private information on risks they want to hedge. Such traders might add noise to the price, but do not on average bias the price. And because their existence attracts informed traders who expect to profit from trading with them, the prices can actually become more accurate. Compared to banning a group of traders from trading, price accuracy is higher if trades by a group are just labeled to be as such.

A potential problem with creating incentives to say what is true is that people might cause harm to make what they say true. If you can poison a drug product at random, just after you sell the manufacturer’s stock, and can hide from subsequent trading audits, you might profit in proportion to your wealth and the thickness of this stock market. There are, however, no documented cases of profiting this way, not even in the famous Tylenol poisonings case. And existing markets that seem to offer opportunities for such behavior are far thicker and harder to audit than any imaginable new information markets. Thus such new markets seem to add little to the problem.

There is also the problem of how to induce people to say things at all via markets, rather than otherwise going about their lives. After all, rational agents should not trade solely to bet with each other, even when differently informed, and information markets will not attract many liquidity traders If desired participants talk about the same topics in other contexts, it is plausible to think that they might talk via markets for the same reasons, such as to enhance their reputation. If they are paid to talk by an employer, however, this employer may need to explicitly reward talking in markets.

Rational informed traders who are only interested in making money can be attracted to real money markets by the presence of noise traders and traders who want to bias the price. Instead of subsidizing such noise or bias traders, however, one can attract the same activity without adding price noise by subsidizing automated inventory-based market makers. This subsidy goes on average to those who acquire new information.

In a market whose question is a single estimate, such as a probability or expected value, a trader who makes or accepts an offer to trade in effect says, “I think the estimate is at least (or most) this value.” All other traders in effect say, “For now, I do not disagree with what the current offers say.” A conversation of this form is clearly rather limited, though in theory among rational honest agents it should produce an agreement about the best estimate.

Ordinary conversations about any one estimate are often aided by discussions of related estimates. If two people were trying to estimate how long it takes to drive from Washington, DC to Detroit, MI, they may do better if one of them could say that he thinks it takes eight hours to drive from Washington, DC to Toledo, OH, and the other can say that he thinks that Detroit is an hour from Toledo. If two people were trying to estimate the chance that a man murdered his wife, and one said the murderer was left-handed, while the other said the husband was right-handed, they might together exonerate the husband.

Similarly, market conversations can also be aided by letting traders say more kinds of things, so they can combine pieces of a puzzle into a larger picture. This should help when part of what people know is the relations between different estimates. A direct approach to letting traders say more kinds of things is to create more kinds of assets, and then invite traders to make offers on them. This approach can run into the thin market problem, however. A trader will not bother to offer to trade an asset if he thinks it unlikely to be accepted soon. After all, he can suffer when new public information goes against his offer, and will not gain if it goes the other way. For this reason, trading tends to focus on assets where offers are likely to be accepted quickly.

Call markets, where offers are matched together at standard times, can avoid the problem of offers that wait too long. And combinatorial matching markets, where traders make offers on packages of items, and software searches for sets of offers than can all be satisfied together, can allow trades even when no two traders make offers on the same asset packages. Matching markets can even allow traders to make complex logical dependencies among their offers, such as allowing only one of two alternate offers to be accepted, or making one offer invalid unless another offer is accepted. This allows traders to say many more kinds of things in market conversations.

Even these approaches, however, can only go so far. If one person knew something about a particular estimate, and everyone else knew that they knew nothing related to this estimate, this one person would not reveal what he knew by making an offer to trade with these other people, since none of them would accept such an offer. A subsidized automated market maker on this estimate would accept such an offer, however, and so could induce this person to reveal what he knows.

It might seem too expensive to create such a market maker for every estimate anyone might make, but with a “market scoring rule” the only cost is for adding another base variable, such as “Distance from DC to Detroit” or “Is the husband the murderer?” to expand the space of possible states. Given base variable funding, there is no additional cost to subsidizing trades in all logical and probability combinations of these variables, such as regarding an estimate of the distance to Detroit conditional on the husband being the murderer. And there is always an exact current value for any such estimate. In this situation, traders can then talk freely about all the included variables, having both the ability and the incentive to reveal anything they might know about them via trades.

There still remain computational difficulties in supporting such trades, however, and it is still not clear how to best integrate such market makers into a general combinatorial matching market. Furthermore, it is not clear what sorts of user interfaces can best allow people to browse the current estimates to find ones they want to change, and to help them determine what sorts of package offers they would then want to make.

In summary, the fact that traders must “put their money where their mouth is” when they say things via markets can help us to better share information. To take full advantage of this feature, however, we should ask markets the questions that would most inform our decisions, and we should seek to allow and encourage traders to say as many kinds of things as possible, so that a big picture can emerge from many pieces. Combinatorial matching markets and market scoring rules hold great promise on this front. The trades of those with incentives to bias prices should be distinguished, if possible. And for decision contingent estimates, let decision makers trade and make decision times clear.