Epistemic Status: Sadly Not Yet Subsidized

Robin Hanson linked to my previous post on prediction markets with the following note:

.@TheZvi reviews what one needs for a working prediction market FUNDED BY TRADERS. If someone else will sponsor/subsidize them, a lot more becomes possible. https://t.co/OgEMahI7ZS

— Robin Hanson (@robinhanson) July 26, 2018


I did briefly mention subsidization, as an option for satisfying the fifth requirement: Sources of Disagreement and Interest, also known as Suckers At The Table. The ultimate sucker is an explicit, intentional one. It can serve that roll quite well, and a sufficiently large subsidy can make up for a lot. Any sufficiently large sucker can do that – give people enough profit to chase, and suddenly not being so well-defined or so quick or probable to resolve, or even being not safe from key insider information, starts to sound like it is worth the risk.

Suppose one wants to create a subsidized prediction market. Your goal presumably is to get a good estimate for the probability distribution of an event, and to do so without paying more than necessary. Secondary goals might include building up interest and a marketplace for this and future prediction markets, and getting a transparently robust result, so others or even the media are more likely to take the outcome seriously. What is the best way to go about doing this?

Before looking at implementation details, I’ll look at the five things a prediction market needs.

I. Well Defined

The most cost-efficient subsidy for a market is to ensure that the market is well defined. Someone has to make sure everyone understands exactly what happens under every scenario, and that someone is you. Careful wording and consideration of corner cases is vital. Taking the time to do this right is a lot more efficient than throwing money at the problem, especially trying to build a system and brand over time.

If you’re going to subsidize a market, step one is to write good careful rules, make sure people understand them, and to commit to making it right for everyone if something goes wrong, if necessary by paying multiple sides as if they had won. This is potentially quite expensive even if it rarely happens, so it’s hard to budget for it, and it feels bad in the moment so often people don’t pull the trigger. Plus, if you do it sometimes, people will argue for it all the time.

But if you’re in it to win it, this is where you start.

II. Quick Resolution

Once you’ve got your definitions settled, your next job is to pay the winners quickly once the event happens. People care about this more than you can possibly imagine. The difference between paying out five seconds after the final play, and five minutes after the final play, is a big game to many. Make them wait an hour and they’ll be furiously complaining on forums. When the outcome is certain, even if it hasn’t actually happened yet, it’s often a great move to pay people in advance. People love it. Of course, occasionally someone does something like pay out on bets on Hillary Clinton two weeks early, in which case you end up paying both sides. But great publicity, and good subsidy!

Another key service is to make sure your system recognizes when a profit has been locked in, or risk has been hedged, and does not needlessly tie up capital.

This one is otherwise tough to work around. If you want to know what happens twenty years from now, nothing is going to make resolving the question happen quickly. You can help a lot by ensuring that the market is liquid. If I buy in at 50% now, then a year from now the market is at 75% and is liquid enough, I can take my profits in one year rather than twenty. That’s still a year, and it’s still unlikely the price will ‘catch up’ fully to my new opinion by that time. It helps a lot, though.

III. Probable Resolution

It is a large feel bad, and a real expense, when capital is tied up and odds look good but then the event doesn’t happen, and funds are returned. It hurts most when you’ve pulled off an arbitrage, and you win money on any result.

If, when this happens, you subsidize people for their time and capital, they’d be much more excited to participate. I think this would have a stronger effect than a similar subsidy to the market itself, once you get enough liquidity to jump start trading. Make sure that if money gets tied up for months or years, that it won’t be for nothing.

IV. Limited Hidden Information

If your goal is to buy the hidden information, you might be all right with others not being interested in your market, as long as the subsidy brings the insiders to the table to mop up the free money. That approach is quite expensive. If the regular traders are still driven away, you’ll end up paying a lot to get the insiders to show their hand, because they can’t make money off anyone else. Even insiders start to worry that others have more and better inside information than they do, which could put them at a disadvantage. So it’s still important to bring in the outsiders.

One approach is to make the inside information public. Do your own investigations, require disclosure from those participating in the events themselves, work to keep everyone informed. That helps but when what you want to get at is the inside information it only goes so far.

That means that when this is your problem, and you can’t fix it directly through action and disclosure, you’re going to have to spend a lot of money. The key is to give that money to the outsiders as much as possible. They are the ones you need at the table, to get yourself a good market. The insiders can then prey on the outsiders, but that’s much better than preying on you directly.

The counterargument, especially if you don’t need to show liquidity or volume, is that if you buy the information directly there’s less noise, so perhaps you want to design the system to get a small number of highly informed traders and let everyone else get driven away. In cases where the outsiders would be pure noise, where the insiders outright know the answer, and where getting outsiders to be suckers that take a loss isn’t practical, that can be best.

V. Disagreement and Interest

This one’s easy. You are paying a subsidy, so you’re the sucker. Be loud about it so everyone knows you’re the sucker, and then they can fight to cash in. Excellent.

The other half, disagreement, is still important. Many people, whose analysis and participation you want, still benefit from a story that explains why they are being paid to express an opinion, rather than fighting to be slightly more efficient at capturing the subsidy. And of course, if no one disagrees about the answer, then your subsidy was wasted, since you already knew the answer!

In light of those issues, what are the best ways to subsidize the market?

Option 0: Cover Your Basics

Solve the issues noted above. Choose a market people want to participate in to begin with. Ensure there are carefully written rules with no ambiguity, that any problems there are covered. Make sure you’ll get things resolved and paid quickly, that capital won’t be tied up one minute longer than necessary. When possible, disclose all the relevant information, on all levels. If things don’t resolve, it would be great if you could compensate people for their time and capital.

And also, make sure everyone is confident the winners will be paid! Nothing kills a market like worrying you can’t collect if you win. That’s often as or more important even than providing strong, reliable liquidity.

If you can improve your interface, usability, accessibility, user’s tax liability or anything like that, definitely do that. If your market design is poor, such as having the wrong tick size, make sure to fix that. Tick sizes that are too small discourage the providing of liquidity to the market, and are in my experience a bigger and more common mistake than ticks that are too big.

Finally, waive the fees. All of them. Deposit fees, withdraw fees, trading fees, you name it. At most, there should be a fee when taking liquidity that is paid entirely to the trader providing liquidity. People hate paying fees a lot more than they like getting subsidies. They won’t cancel out.

With that out of the way, what are your options for the main subsidy?

Option 1: Be a Market Maker and Provide Liquidity Directly

As the subsidizer of the market, commit to being the market maker with well-defined rules.

The standard principle is, let everyone know that there will always be $X of liquidity available on both sides, and at a fixed cost of Y% price difference between your bid and your offer. So for example, you might agree to offer $1,000 on each side with a difference of 5% at all times, starting with a 48% bid and a 53% offer. You’d then adjust as you did trades.

A simple rule to protect yourself from unlimited downside is if you do a trade for some percent of your liquidity, you adjust your price that percentage of its width. So in this example, if someone took 40% of your offer, you’d adjust by 40% of 5%, which is 2%, and now have a 50% bid and a 55% offer. If you follow such a rule, your maximum loss is what it takes to move the odds to 0% or 100% (and if you let people keep trading until the event is done, you will take that loss). People trading against you in opposite directions can make you money, but can’t cost you money.

For convenience, you can post additional bids and offers so that if someone wants to move the odds a lot, they can see what liquidity they would get from you, and have the option to take it all at once. You’ll lose money every time the fair probability changes, but that’s why they call it a subsidy, an this encourages people to show their information quickly and efficiently.

There are ways to make that smarter, so you can lose less (or make more!) money while offering better liquidity, which will be left as an exercise to the reader. Generally they sacrifice simplicity and transparency in order to make the subsidy ‘more efficient.’ The danger is that if the subsidy is attempting to ensure a sucker is at the table, it does not do that if it stops being the sucker, or it becomes too hard to tell if it is one or not.

Then again, the dream is to offer a subsidy that doesn’t cost you anything, or even makes you money! Market making can be highly profitable when done skillfully, while also building up a marketplace.

Option 2: Take Liquidity

If you provide liquidity, others will take advantage, but in some ways you make it harder to provide liquidity. If you take liquidity, you make it more profitable to provide it, at the risk of making the market look less liquid.

It also loses money. The more clear you are about what you are up to, the better.

There are a few fun variants of this, if you’re all right with the expense.

One strategy is to take periodically liquidity in both directions. At either fixed or random intervals, examine the order books in the market. If they meet required conditions (e.g. there is at least $X on the bid and offer within Y% of each other) then you hit the bid and lift the offer for $Z.

This costs you money, since your trades net out at a loss. If someone else was both the best bid and best offer, they made money.

That’s the idea. You’re directly subsidizing people to aggressively provide liquidity.

Traders compete to be on the bid and offer to trade with you, the virtual customer, which in turn gives those with an opinion a liquid market to trade against. Sometimes people get far too aggressive providing in such situations, and those trying to capture the subsidy end up losing money because they make bad trades against others, especially if they don’t then hedge.

You can also do this in a more random or unbalanced fashion. If you flip a coin each day and decide whether to be a buyer or a seller, that will cause the price to temporarily become ‘unfair’ to satisfy your demand – you’ll get a bad price. But that creates a trading opportunity for others. It can also make the results hard to interpret, which is a risk.

Option 3: Subsidize Trading / Give Free Money

Often you’ll see crypto exchanges do this as a promotion, offering a prize to whoever trades the most of some coin. By paying for trades, you’re encouraging exactly what you want.

Except that you’re probably not doing that. Remember Goodhart’s Law.

The problem is ‘wash’ trading, where people trade with each other or themselves without taking on positions. This is bad on every level. It misleads everyone about the volume and price, and doesn’t help at all with finding out the answer to the question the market is trying to answer. The last thing you want to do is encourage it!

For that reason, subsidizing trading itself is a dangerous game. But it can be done, if you’re careful with the design.

Many online sites have tried this in the form of the classic ‘deposit bonus’ or even the free play. Anyone can sign up and get Free Money in exchange for engaging in a minimum amount of trading activity. And of course, most of the time, a deposit to match, if the offer is more than a small ‘free play.’ In for-profit markets the goal is to have the required activity make up for the subsidy, then hopefully hook the customer to keep them trading. There are always those looking to game these offerings if you leave them vulnerable.

That can work for you. Getting those same people, who are often quite creative and clever, thinking about how to come out ahead in your system can be a big win if your end goal isn’t profit! So long as you make it sufficiently difficult to do wash trading or sign up for tons of copies of the bonuses, you can give them a puzzle worth maximizing (from their perspective) and effectively rent their labor to see what they think of the situation.

Option 4: Subsidize Market Making

You can also subsidize market making activity, as an alternative to doing the job yourself and butchering it. That’s activity you can’t fake, provided you set the rules carefully. Paying people who provide rather than take liquidity is good, and often paying for real two-sided market making activity is better. As always, make sure you’re not vulnerable to wash trading or other forms of collusion.

Option 5: Advertising

People can’t trade what they aren’t thinking about or don’t know about.

Putting It All Together

Which of these strategies is most efficient and what circumstances change that answer?

It’s expensive to change or clarify your rules and conditions once trading has begun, so invest in doing that first. Other quality of life improvements are great, but take a back seat to establishing good liquidity.

I list Option 0 first because it’s things you definitely should do if you’re taking the operation seriously, but that doesn’t mean you always do all of them first before the direct subsidy. It’s great if you can, but often you need to establish liquidity first.

If ‘no liquidity’ is the pain point and bad experience, there isn’t much that will overcome that. There’s no market. So if you don’t have liquidity yet, providing at least a reasonable amount, or paying someone else to do it, is the best thing you can do. Just throw something out there and see what happens. This makes intuitive sense all around – as an easy intuition pump, if you want to know if something is more likely than not, offering someone a 50/50 bet on it is a great way to get their real opinion.

Once liquidity isn’t a full deal breaker, it’s time to go with Option 0, then return to increasing the subsidy and spreading the word.

What form should the direct subsidy take?

I’d advise to continue to take away bad experiences and barriers first.

The best subsidy is paying to produce reliable, safe and easy to use software, getting ironclad rules in place, being ready to handle deposits, withdraws, evaluation of results and other hassles. Make sure people can find your markets and set up the markets people want to find.

Next best is to avoid fees. People hate fees more than they love subsidies. Yes, you can trick people with deposit bonuses and then charge them a lot on their trades, but the best way to get away with that is bake the fees into the trade prices, so it doesn’t look like a fee.

At a minimum, you shouldn’t be charging fees for deposits or withdraws, or for providing liquidity in the market.

Next up, make trades cost net zero fees. Either charge nothing to provide or to take liquidity, or charge a fee to take liquidity but pay it to those who provide.

After that, my opinions are less confident, but here’s my best guess.

If that’s still not good enough, provide liquidity. Either pay someone else to be a market maker, or provide the service yourself. I like the idea of a ‘dumb’ market maker everyone knows is dumb, and that operates with known rules that hamstring it. If you’re looking to provide a subsidy, this is a great way to do that. A smarter market maker is cheaper, and can provide better liquidity, but is less obviously a target. As the market matures, you’ll want to transition to something smarter. Thin markets want obviously dumb providers.

Once you’ve done a healthy amount of that, then you’ll want to give away Free Money. Give people some cash in exchange for participating in the market at all, or trading a minimum amount. Or give people bonuses on deposited funds so long as they use them to trade, or similar.

You have to watch for abuse. If you can respond to abuse by changing the system, it’s fine to be vulnerable to abuse in theory, and even allow small amounts of it. If you’re going to release a cryptographic protocol you can’t alter, you’ll need to be game theoretically robust, so this won’t be an option, and you’ll have to retreat to taking liquidity.

Taking liquidity seems less likely to motivate the average potential participant, and costs you weirdness points, but does provide a strong incentive for the right type of trader. The best reason I can think of to use such a strategy is that it is robust to abuse. That’s a big game if you can’t respond dynamically to unfriendly players.

At the end of the day, your biggest barriers are that people’s attention is limited, complexity is bad, opportunity cost is high and people don’t do things. I keep meaning to get around to bothering with HyperMind and/or PredictIt, and keep not doing it, and I’m guessing I am far from alone in that. Subsidy can get people excited and make markets work that wouldn’t otherwise get off the ground. What I think they can’t do at reasonable cost is fix fundamental problems. If you don’t have a great product behind the subsidy, it’s going to be orders of magnitude more expensive to motivate participation.

New to LessWrong?

New Comment
8 comments, sorted by Click to highlight new comments since: Today at 8:09 PM

Robin has published at least two papers that seem like necessary background reading on this topic:

Note that the logarithmic market scoring rule comes with a built in method of subsidy (in fact, requires such subsidy), and guarantees that there will always be nonzero liquidity.

Removing fees is of course a valuable and necessary first step. There's never any reason to do something so self-defeating as deliberately destroying liquidity by imposing fees.

I've seen it, and my reaction was "that's very interesting and clever, I'm glad someone figured that out and wrote it up, but man is that way too complex to actually work."

Promoted to curated: It's a good follow-up to the last post, so a lot of the things I said in that curation notice apply here as well. This one was similarly well written, had clear actionables, and I think had some slightly better structure than the last one. I am excited about people implementing these lessons in real-world prediction markets (and/or to adjust their plans based on these lessons).

Why do you believe quick resolution is so important, that there's a huge difference between whether a payment is made day sooner or later?

Experience both as a participant, and as someone who has run prediction markets in the past. This is people's revealed preference. Quick resolution is important to people on a different order of magnitude than you'd expect - people would happily pay much higher fees in order to not have to wait.

Thanks for writing this up. I agree that subsidies are a crucial ingredient. I don't agree completely with everything else you wrote, so I'm going to write a longish reply.

Having enough interest in the market is the best solution, but this isn't always available, and the markets we want to see don't always coincide with the ones that people are most interested in participating in. There are spillover effects, so hosting popular questions on sports, politics and celebreties can bring in enough people that a few of them will also bet on the questions that provide actionable intelligence on the future, like technology and legislation.

I agree that it's important to have well-defined outcomes, so people can rely on the correct outcome, and a transparent interpretation of the question. This is pretty straightforward when you stick to popular, highly repetitive areas like sports. When you want to support questions in a variety of areas, this means you have to write custom descriptions for each question, and there's a lot more room for the vagaries of the real world to throw curveballs at you.

I worry most about this with systems like Augur, where anyone can write a claim. If the trading platform has no way to ensure consistent quality, the traders will learn to be sceptical of all innovative claims. Foresight Exchange (play money, somewhat moribund) got around this by having an active community that participated in writing the claims. They occasionally still got into disputes, and but I think their approach of having named judges with declarations on how they were going to judge made the most sense. Augur's voting system seems like a bad compromise to me: all questionable outcomes turn into beauty contests. The voters all have an incentive to favor the most popular outcome, which seems unlikely to lead to results that are predictable.

I doubt that paying off both sides on tough judgement calls will induce more participation. I agree that that incentives bettors to raise a stink on any sligthly unusual outcome. You don't want to encourage people to put money on bad claims because they'll be paid off regardless.

I think quick payouts matter much more in the highly repetitive markets (especially sports) than in questions that take a long time to settle. In the case of sports, you're competing with sports books for the attention of bettors who often have several questions open at a time, and want to get their money back into play. If they've invested in a quesion that will be open for 6 months or a year, getting the answer right matters much more, particularly if your interest is in getting the market to be active enough that the prices can be used as guides by people who aren't putting their money on the line.

I'm interested in your ideas about paying off early, and "recognizing when a profit has been locked in, or risk has been hedged". Prediction Markets already do this so much better than other betting venues that I'm not sure what could be added. Markets that support combination claims have been designed, but none have become popular. (Hanson's Combinatorial design, mentioned above; Peter McCloskey's USIFEX; Dave Pennock also developed something at Yahoo) In the absence of these, what can be done?

I agree that getting rid of fees matters a lot. The most frequent betters are the ones that pay the most attention to how much the house's cut is.

The problem with being a market maker, as you alluded to, is that it's hard to do it without being a money pump that can be exploited for unlimited gains. Robin's design (which is demonstrated by my Zocalo) shows how the house can do it. If you aren't the house and mediating all trades, it's hard to write a hard and fast rule for entering bids that can't be exploited, or doesn't lose money just because it doesn't participate sufficiently during rapid price movements. If your program says to enter a bid at a certain price, but the market has already moved, then the opportunity is gone.

For long term bets, where the opportunity cost of tying money up in these bets becomes high, I would have thought that the bets should be denominated in US bonds (or other agreed minimal-risk interest rate asset) to minimize this cost.

Even if the bet does not pay out one way or another, the money still accumulates interest.

Other than being incompatible with Augur, are there any theoretical or practical hurdles to using this? It would hopefully reduce the subsidy required to make an attractive market without incurring cost in and of itself.

Paying in some reasonable investment is an improvement, and in a world with different 'right' investments would be a bigger one, but tying up the money that long in a sufficiently safe investment is still pretty expensive.

It would be interesting if one could bet, say, shares of SPY US plus associated dividends, so you were leaking a lot less alpha. Of course, that changes the odds on long bets quite a bit if they are correlated with SPY US, which most of them will be.