Deliberative Prediction Markets — A Reply

Robin suggests that a more robust model of deliberative prediction markets would be useful, and I agree. Experimentation in the field would be even more useful. But my original paper and book section explain the logic of this market approach clearly, with both words and what I acknowledged was a "simple model."

I doubt, in any event, that a more elaborate model will change the basic conclusion: that the deliberative prediction market provides at least some increased incentive to reveal information. I’ll let interested readers look at the original paper for a more developed argument (including math), but it boils down to a very simple point. A prediction market, as Robin and I both note, already provides some incentives to reveal information. But if a trader’s payoff depends on whether the trader actually succeeds at persuading others rather than on whether the trader turns out in the long run to be correct, the trader will have an additional incentive to reveal that information.

Relaxing the assumptions that Robin questions will not undermine this point. First, Robin notes that a trader may get more than one package of information. It’s possible that the deliberative market might lead someone to reveal first one package of information (pointing, say, to a higher value), wait to profit on that, and then reveal another package of information (pointing to a lower value), and cash in on that too.

So, the deliberative market could on occasion delay the release of some information packages. But those information packages might not have been released at all in the standard prediction market. Suppose that a trader in a standard prediction market conducts research that reveals two information packages that cancel each other out. The trader has no incentive to trade at all, let alone to reveal the two packages of information.

The bottom line is that I can’t think of any scenario in which the deliberative prediction market would result in less information revelation, and it might well result in more information revelation. Allowing private information exchange doesn’t undermine this either. Private information exchange can also allow a trader to cash out a position or, in the deliberative market, to make sure that the market does not move too much after the trader’s initial prediction.

Finally, I’m making a straightforward comparison in my paper (see p. 17), assessing the effect of two different possible compensation regimes in the same environment. Robin argues that the assumptions about this environment are unrealistic and that I would reach a different conclusion relaxing the assumptions (though I disagree above). But I just don’t see how I change either the explicit or implicit assumptions underlying the two regimes. Admittedly, I don’t explicitly model the information revelation incentives that already would exist in both markets, but I hold those constant in the two settings.

Robin’s post is valuable and helps point in the direction of useful research, even though Robin does not include a robust formal model to justify his claim that "we can easily get the same info revelation advantages" in the standard prediction market as in the deliberative market. I hope that Robin will agree that the concept of a deliberative market also is valuable and points in the direction of useful research, even though I support it with only three pages of math.

GD Star Rating
loading...
Tagged as:
Trackback URL:
  • http://hanson.gmu.edu Robin Hanson

    Michael, I don’t think you’ve understood my objection. I said that given the assumptions you used all the same info revelation advantages also come in a simple market where people must cash out soon after trading and never trade again. I’m saying it is forcing people with very limited actions to cash out and never trade again that is driving the results – the specific market structure doesn’t matter.

  • Michael Abramowicz

    Robin, I think we are very close to achieving the goal of this blog — overcoming bias — or at least agreeing on the merits of the issue. I saw that aspect of your original post, but thought that you were comparing my proposal to a prediction market without such a rule. In the original article (pp. 113-114, or 13-14 in the acrobat file), I note that requiring traders to liquidate positions would essentially accomplish the underlying goal of encouraging information release. And so I would count such a rule as one means of creating a deliberative market (though I admit I did not define it as such in the original article).

    I do still favor the deliberative market approach that I describe to such a rule (but would not claim that it induces more information revelation than such a rule). The reason that I favor it is that ordinarily under the market scoring rule, a new prediction moves the market prediction, but I would not want liquidation to move the prediction back.

    For example, suppose the market prediction is 50, but trader X has information that leads X to announce a new prediction of 30. Suppose X releases information supporting the new prediction of 30, and that all other traders are persuaded by this information, so no one bothers to announce a new prediction. If X is required to liquidate her position, she could do this under the market scoring rule by again announcing a prediction of 50, but this would mean that she makes zero profit (plus misleadlingly change the market forecast).

    Admittedly, we could require liquidation by an auction, but then we would face the risk that X or a surrogate would buy her own position. Now, I’m not sure I’ll have convinced you that the specific formula that I use for the deliberative market is better than a forced liquidation rule. Moreover, the forced liquidation rule could be as good in a market not operating with an automated market maker implemented by the market scoring rule. And of course we can agree to disagree on the optimal amount of mathematical development needed for such a proposal.

    I hope you’ll at least agree, though, that both rules would yield more (or at least as much) information revelation than we would have in a standard prediction market. That was what I had intended as my central point.

  • http://pancrit.org Chris Hibbert

    I don’t see that the incentives for information revelation are different for the two institutions. In any normal prediction market, you have an incentive both to predict the final determination of the value (if you are a buy-and-hold trader), and to predict the direction of movement in the short term (if you at least sometimes trade in-and-out of a market.) If a trader recognizes a circumstance that will lead to the price moving in a particular direction in the short term, independent of what the final outcome may be (e.g. party nominees for US President get a boost during their party’s convention) there’s an incentive to bet in that direction before the movement happens and take a profit afterwards. If you can release information that will move the price, you can make the same play. (I’ve done that on FX, sometimes successfully, sometimes not.)

    But your proposal adds extra emphasis to convincing other participants about the relevance of particular information rather than the final judge. If the rewards for convincing traders is high enough, people will spend time and effort developing evidence aimed at the traders, and the value of convincing the final judge is reduced.

    In a standard prediction market, there’s already a value to convincing other traders, as it enables in-and-out traders to make short term gains. But our proposal disenfranchises (at least relatively) the buy-and-hold traders. They would be required to liquidate (or pay or receive bonuses) based on the intermediate price. In a standard market, the buy-and-hold traders who have firm beliefs about what the final judge will decide don’t have to worry about the path that the price takes before close. Their reliance on the final judgment serves to tie the intermediate price more firmly to the final judged outcome.

    To the extent that intermediate rewards are based on consensus, bluffing, or the opinion of the traders rather than the ruling of the judge, everyone is forced to treat the market as a popularity contest rather than a prediction of what the judge willl decide at the end.

  • Michael Abramowicz

    Chris, You’re absolutely right that a disadvantage of the deliberative prediction market is that it “disenfranchises buy-and-hold traders.” The approach essentially says that traders have to persuade one another to profit. This will push more people from the buy-and-hold strategy to a short-term strategy in which they will have more of an incentive to release information because they will be judged on the short-term “popularity” of their position. So, some people might drop out of playing, and that could decrease accuracy, but the sharing of information might increase accuracy. And you’re right that there might be less effort in convincing the “final judge”; this will, of course, be more important for some markets than others, depending on the subjectivity of the ultimate event being predicted.

  • http://hanson.gmu.edu Robin Hanson

    Michael, in my critique I quoted you from your book saying that it is the multi-stage design that gives the info revelation incentive:

    When one prediction market is used to predict the outcome of another, participants who have private information in the first prediction market have an incentive to reveal that information to participants in the second.

    I’m glad you acknowledge the same info revelation incentive comes in a simple market where traders must cash out soon after trading and are prevented from teaming or privately talking. But you haven’t yet acknowledged that you need those same restrictions to get info revelation incentives in your multi-stage design. Without such restrictions, traders who in an ordinary market would wait for a later chance to use their info again, or sell their info privately, could easily follow the same strategies in a multi-stage market.

    Perhaps you think that such strategies would not be similarly profitable given your market design but without the restrictions. However, I don’t think you understand equilibrium trading behavior under your design. Your design creates strong incentives to try to manipulate future markets in order to increase the value of assets obtained in previous markets. It still seems to me that, contrary to the impression you give in your book, it is the restrictions, not the market structure, that is doing all the work of creating stronger incentives for info revelation.

    All this should not distract us from the main point though, which is that there will be few situations where it is feasible to impose such restrictions – usually people can team and exchange info secretly, without the knowledge of those who set up trading rules.

  • Michael Abramowicz

    Robin, Well, we’re pretty close to agreeing anyway. It’s the cash-out requirement (whether with an auction, with the formula I suggest, etc.) that produces incentives for information revelation. I agree that “traders who in an ordinary market would wait for a later chance to use their info again, or sell their info privately, could easily follow the same strategies in a multi-stage market,” but the deliberative market gives them at least some incentive not to follow those strategies.

    This will be true even if traders can team or talk privately. I agree that there are “strong incentives to try to manipulate future markets in order to increase the value of assets obtained in previous markets,” and discuss this in my paper. But this won’t be easy to do, because the exact time of liquidation is uncertain, and the possibility of manipulation gives extra incentives for third parties to try to identify such manipulation efforts (as you’ve shown with conventional prediction markets). If X seems to persist in a prediction not supported by any released information, then Y will have some incentive to trade against X.

  • http://hanson.gmu.edu Robin Hanson

    My main original claim was that your book makes many claims for market structures but does not offer support beyond your intuitions, at least to an economist’s standards. Your book claimed that your multi-stage design offered more incentives for info revelation, but now we see that while we agree that a cash out requirement can have this effect (given certain unrealistic strong assumptions) your claim that your market design gives an additional info revelation benefit is based only on your intuitions, which differ from mine.

  • Michael Abramowicz

    You misunderstand the claim I was making, or at least trying to make. My claim was that approaches that based the trader’s profit on later market forecasts (including multi-stage design or a cash-out rule) would produce more information revelation than approaches in which the payout depended on the eventual event. I made a second claim, elaborated above, that the multi-stage design would be preferable to the cash-out rule for other reasons. You still seem to think I was comparing the multi-stage design to a cash-out rule based on which would produce more information revelation. I wasn’t.

    It seems to me that we’ve gotten to a point where we’ve identified the misunderstanding, and recognized that we have some continued disagreement about whether I was relying on unrealistically strong assumptions. I think we’re probably the only two people still following this thread, so perhaps we should leave it at that. Meanwhile, I’ll look forward to your promised post arguing for Futarchy over Predictocracy.

  • http://hanson.gmu.edu Robin Hanson

    Michael, I was saying the cash out rule can apply to all market structures, and that it is the cash out rule that is doing the work, not the market structure, contrary to your claim in the book, which is that one market structure does better than another.

  • Michael Abramowicz

    Robin, Good, we’re in agreement. It’s true that my book did not explicitly discuss the cash-out rule in its brief summary of the deliberative prediction market. As the original article explains, that rule can also do the info revelation work, though it may have other flaws relative to the approach I did describe.