In case you missed it, there've been some fascinating experiments with self-resolving markets on Manifold lately. The idea of a self-resolving market is that sometimes it's hard or impossible to pin down unambiguous resolution criteria for a prediction. In that case resolving to the consensus of the market participants may make sense. And from a prediction market perspective, the obvious way to define "the consensus of the participants" is the market price. See also Keynesian beauty contests. It's weirdly self-referential but maybe there are conditions in which we expect it to work?
We've seen one spectacular failure -- https://manifold.markets/jack/will-biden-be-president-on-915-reso -- in which an exciting tug-o-war was waged that had little connection to what the market was purportedly predicting.
But I conjecture that that was mostly about the hard(ish) end date. In setting up the experiment we knew that a known end date opens the market up to price manipulation at the last second and we tried to mitigate that with a random close time, but it was a random close time within a small window. It was still possible to temporarily manipulate the price for profit.
So this is an experiment to see how close we get to the truth if the market self-resolves as follows:
The market won't close before the end of October (when the true outcome is known).
The market will then stay open until the market price fully quiesces, meaning it stays the same (to the nearest whole percent) for two business days in a row. The market then resolves to that price.
But if a market participant wants to go on vacation or something, they can just ask in the comments to not let the market close before they return. We want to only resolve to a genuine consensus probability.
If we go all the way [through] November without the market stabilizing then it resolves N/A. That would be a shame but not necessarily bad, in general, in terms of the market price being informative, if it's oscillating around some estimate of the underlying truth.
Point 3 might be silly and unscalable, I'm not sure. It's an experiment. But if it did make sense, one can imagine ways to make it scale to larger markets. The answer might be to stick to a strict definition of quiescence and traders can just use trading bots or limit orders while they're away. (Or we could just say that checking in every other business day isn't that onerous, suck it up.)
Point 4 may open up an unfortunate loophole in that a participant can force the price to oscillate if they don't like the outcome. I think liquidity fees could solve that (which may be important anyway, since it's a very valuable property of markets that liquidity increases with trading volume!) because then it would get more and more expensive to keep the price oscillating. For this experiment, we'll just take the risk that that happens. Since everyone gets their money back in that case, the risk is arguably not so bad.
UPDATE: From the comments, pinning down the criteria for quiescence:
If the market price stays the same or oscillates between consecutive integer percents from noon pacific to noon pacific two business days later (so at least 48 hours, longer over weekends and holidays) then that counts as the market quiescing. I plan to add comments warning us as we approach the end of the window and plan to say yes to requests to temporarily lengthen the window (assuming they're based on real-world reasons like vacations).
PS: The official Schelling Point is 95%. Or oscillating between 94% and 95%.
Related questions
@jack said in a comment below, '50% is clearly the correct quiescence price and Schelling point. Any other price makes it easier to force the market to move.' which sounded right at the time.
But if you could expect a high chance of someone who lost out from the market destroying it, then I'm not sure this is right. Moving the market to 50% after its having gone higher is then equivalent to N/A, and everyone getting nothing. Whereas at least in a game that can be positive sum for the traders (which I think this is?) 95% can be got to without anyone losing out and having an incentive to zero everything. i.e. Wouldn't Jack have done better by participating in pushing the value to 95% and getting a small payout than by pushing it to 50% and thus getting nothing?
I agree things are more complicated if you don't believe anyone else will have the money to oppose you, and that that is more likely to be true at 50%, where the most money would be needed. But in reality, I'm not sure how much you should expect this.
In sum, while this market demonstrated that such markets are a mess and will not necessarily help you find the truth, I remain unsure what the equilibrium is in a market like this composed of perfect economic agents intent on maximizing Ṁ.
@KatjaGrace Yeah, there are definitely some complicated market dynamics here. Sure, picking 90% or 80% or something would give others less incentive to break the quiescence. But it also would have been easier for them to push the price down from there. At 90%, the YES holders still stood to lose some mana, and could easily have broken quiescence by buying NO with the funds they already had available, so I'm pretty confident that wouldn't have worked.
Another note is even 50% isn't the most defensible price, because a lot of other people held YES shares, which means they could in theory push the price down by selling those shares in addition to buying NO, more easily than pushing the price up. That's one reason I realized that 50% is not the ideal quiescence point and why we ended up adjusting the price to 40%; of course the other reason was more potential profits.
I agree things are more complicated if you don't believe anyone else will have the money to oppose you, and that that is more likely to be true at 50%, where the most money would be needed. But in reality, I'm not sure how much you should expect this.
I don't think this is true - picking a price near 50% is motivated specifically by the belief that others have a solid chance of putting together enough funds to contest the quiescence.
In fact, I predicted >50% chance of it on the derivative market https://manifold.markets/Yev/will-the-biden-quiescence-market-re-5fd04f2244bf, and made a tidy profit as a result 🙂
If my estimate of the funds available to the opposition was much lower, then it would be even more profitable to go for a lower price e.g. 20%.
Wouldn't Jack have done better by participating in pushing the value to 95% and getting a small payout than by pushing it to 50% and thus getting nothing?
Targeting 95% would be interesting in a different way - the market was below 95% almost all the time so that would actually mean pushing the price up, and there would be someone else who could profit by pushing it down in opposition. (If I didn't do the NO bet, there were a couple other people who very well might have - and in fact I had been expecting one of them to do it before me.)
I remain unsure what the equilibrium is in a market like this composed of perfect economic agents intent on maximizing Ṁ.
Yeah, I'm not sure either, but from my experience here and on other similar markets, I think the game largely is played outside the market itself, because it revolves around a) controlling large amounts of funds, possibly via building a coalition with other players, and b) deceiving other players about the amount of funds you control to bait them into making unprofitable bets against you.
Also, if we ignore extracting the AMM liquidity from the market, is there any reason to believe there's a Nash equilibrium other than "nobody bets"? For anyone who doesn't believe they can do the above better than the other players, the best move is not to play. And if they don't play, then it's harder for everyone else to profit - so there should be strong adverse selection effects against anyone trying to trade.
Well that was exciting! Huge thanks to @KatjaGrace who did almost all the work in preventing the manipulators from getting the market to resolve at 40%! But also kudos to the manipulators -- mainly @jack -- who proved the point that this variant does not in fact fix the problem in the previous auto-resolve-with-quiescence experiment.
For the question of whether one more patch could work, where the quiescence criteria are hypersensitive so the market only resolves if everyone agrees, @jack has convinced me it doesn't work. It might happen to work for dummy questions like Biden being president but all of this only matters for real questions with real disagreement and that means hypersensitive quiescence criteria just guarantees no quiescence. Disagreement means pulling the price in different directions, after all. So if it's very predictable that the market will resolve N/A then people won't invest resources to make the price be right, whatever they think "right" means. That defeats the point of the prediction market. Removing the incentive for manipulation in an auto-resolving market also removes the incentive to make the market reflect the truth.
I'm not fully convinced of that -- maybe there are real-world conditions in which the chances of manipulation are plenty low -- but, well, this experiment (and especially discussions with @jack about it) increased my understanding of the limitations by a lot.
Thanks everyone!
Thanks all for a very interesting and exciting market!
My main point is that the market resolution needs to have some tie to the actual question in order for the market to function correctly. (See my previous comment https://manifold.markets/dreev/biden-quiescence#xpULTrgpj4AAxnzI1ERO for a list of mechanisms that do that.)
And for anyone who is writing self-resolving markets anyway, I think this experiment does give some interesting insight into at least avoiding last-minute price manipulation - i.e. ensuring that the resolution process takes place over a span of time to give the participants time to react, trade, and influence the price. I still believe methods like randomizing the close time work well here - it has most of the benefits of quiescence, without the biggest drawbacks. I think the biggest problem with the quiescence criteria that have been used so far are that they either can resolve N/A too easily, or they can take way too long to quiescence. There are probably variations on the quiescence criteria that could work better e.g. expanding the quiescence threshold gradually over time. (Again, this only addresses the smaller problem of last-minute manipulation, not the broader problem of manipulation in general, but at least it tends to arrive at a somewhat more stable equilibrium.)
Also check out my updated post https://manifold.markets/group/selfresolving/about with my suggestions on what works and what doesn't.
@OptimizationProcess Agreed, thanks all!
And sorry to those of you who bet no on quiescence on the derivative market haha 🙂
Ok, it's the final quiescence period so two possible outcomes:
The limit order wall at 41% holds till Wednesday and market resolves at 41%
All of the Biden-believers collectively have like M$100k and we break through and the market instantly resolves N/A
Again, there's no harm in putting your whole balance on YES at this point, because either we succeed and get the N/A or we fail and any shares bought now cash out at the same price you paid for them.
As I've said before on these types of markets, I want to identify and showcase vulnerable market mechanisms so that people don't rely on them for something that's actually important.
So far I believe these are the most promising methods instead of self-resolving markets:
Resolving to the result of a poll - this is a much safer and more robust method. While they aren't perfect, polls tend to work a lot better at getting a reasonable result and being mostly (but not 100%) robust to market manipulation. Examples: https://manifold.markets/SneakySly/at-the-end-of-2023-will-manifold-us and https://manifold.markets/jack/will-we-believe-sbf-committed-willf
Some random chance of resolving the normal way, otherwise resolve N/A. Imagine a market predicting the result of an expensive experimental trial. This mechanism means you only have to actually run the trial some fraction of the time, and it is incentive-compatible (you can't profit on the market by manipulating it). Example: https://manifold.markets/jack/does-gdpr-require-selfserviceautoma. Downside is that profit incentives for making good predictions are correspondingly lower.
If you really want to resolve-to-mkt, you can at least have some random chance of resolving the normal way, and protections against last-minute price manipulation such as a randomized close time or some quiescence criteria. E.g. https://manifold.markets/Yev/will-biden-be-president-on-october-fb3d01633429 This is still not robust against manipulation but at least it's slightly better than the alternatives we've seen so far.
Also, if you are ok with some author subjectivity, you can say "Resolves to MKT but the author will override if it looks like market manipulation". This has empirically worked ok for many low-stakes markets, but you do have to acknowledge that it becomes very subjective. The line between voting your beliefs and manipulating the market can be very blurry.
@jack What about a stricter definition of quiescence? I keep thinking one more patch to this kind of approach will finally work. Maybe eventually it will?
@AlexRockwell I don't think so. Yev or Jack would have bought a bunch of NO every 48 hours, on the hour. Then they would have sprung the exact same ambush when the deadline neared.
@MichaelWheatley What if (as in the original quiescence criteria that I stupidly agreed to change) the price had to round to the same integer for 48 hours? My probably-still-naive conjecture is that in that case there's no way to force quiescence and so any manipulation attempts would prevent quiescence and so no manipulation is attempted and the price just goes to something reasonable.
@dreev I think in that case, you'd get people speculatively inching the price toward NO in the hopes that their opponents get bored and give up before they do. They do, can prevent quiescence if things don't go their way. So nothing this egregious, but definitely not 99% either.
@dreev I was with you up until "so any manipulation attempts would prevent quiescence". But IMO that just means that anyone who wants to prevent quiescence would attempt to manipulate the market, and so the market would never quiesce.
@dreev Thinking about it some more, it would be so easy to prevent such a market from resolving that it may as well say "Resolves to N/A." I bet you could get pretty good answers there, but by removing any profit incentive, you're forgoing a big part of what makes prediction markets useful. On a difficult question, no-one is incentivized to do any research or surface any new facts.
@dreev Yeah, Yev and I suggested that change right from the start because requiring it to round to the same whole number clearly doesn't work. In any market, someone will have lost mana, and they always do better by forcing n/a resolution to undo their losses. In any realistic market people are disagreeing, and the whole point of the market is to average their information. so rationally you should expect all such markets to fail to quiesce.
Also, more generally, my argument all along has been that resolve-to-mkt is fundamentally subject to price manipulation. The resolution price is determined essentially as the average of the "predictions" of the traders weighted by how much funds they invest. That's also true for prediction markets in general, but in normal markets you make predictions about the question, while in resolve-to-mkt you're just making predictions about what the final market price will be. (My comments in https://manifold.markets/jack/will-biden-be-president-on-915-reso talked about this a lot)
All the quiescence and Poisson mechanisms do is address the smaller problem of last-minute price manipulation - it at least ensures that everyone has some time trade and have price input on the market resolution. It does nothing about the fact that the market is fundamentally asking a different question (about the final market price) than the one you actually want to ask (about real things in the real world).