why are prediction markets so accurate, why do they lend themselves to blockchains, and what do blockchains do for them?
What is a prediction market?
A prediction market is a market for the buying and selling of futures contracts on binary events. The market is structured such that individual contracts are priced between $0 and $1. A participant buys a contract on an outcome A, or its complement, ~A, for, say, 50¢, and receives $1 if outcome A occurs, or $0 if it does not. If we designate the price of a contract on A to be pA, then p~A will be priced at $1-pA, although depending on the mechanics of the market, there may be a slight bid/ask spread to provide revenue for the exchange, or alternatively a fee to buy the contract, to preserve the balance of the contract prices.
This is very similar to how bookmakers construct their books by adjusting the odds offered to match punters’ bets against one another, but there is a crucial difference here reflecting the construction of the market. The contracts exist as securities rather than isolated wagers. This allows for considerably greater liquidity, given that participants need not necessarily wait until the event occurs. This liquidity in turn ought, in theory, to aid price discovery, since participants with relevant insight or information can take advantage of what they deem to be temporary mispricing rather than having to lock in their bets.
Where this gets interesting is in interpreting what the prices mean, and therefore what a mispricing is, and what ‘the expectations of participants’ are. The setup of prediction markets lends itself to interpreting the price as the probability of the outcome. For uncertain events for which a probability density function cannot be analytically proven to exist and which cannot be re-run a million times to derive such a function synthetically (so, most of real life) assigning a probability to a specific outcome can only meaningfully be interpreted as stating what odds you deem to constitute a fair bet. A more common and readily understandable application of assigning probabilities is to imply confidence in deviation from a base rate, but this is really a special case of the general principle mentioned above; the base rate is derived from a historic ‘large enough’ sample for which the mean would be deemed ‘fair’ for a random iteration, and the question is whether we have relevant information leading us to believe that the non-random iteration at hand ought to deviate from the distribution mean, and hence whether ‘fair’ odds are different from those derived from the historic data.
If you think something is 50% likely, then you think even odds are fair. In general, if you think something is p% likely, then you think odds of (1-p)/p : 1 are fair. This calculation can be reversed to turn typical bookmaker odds into probabilities, but a prediction market deliberately uses this presentation to state the prices of bets in terms of probabilities rather than odds. If you think the probability of outcome B is x% likely, and that (1-x)/x : 1 is therefore a fair bet, then a price lower than x means odds better than those you deem to be fair. They are to your advantage and should be taken. Of course, if you did take such a bet, in a prediction market as with a bookie, the result would be to push the odds on offer towards what you deem to be fair. With enough people behaving this way, the price is pushed towards what the entire sample deems to be fair, weighted by each individual’s willingness to shoulder the risk implied by the mispricing.
Sounds interesting, but what’s the point?
The point is that, insofar as prediction markets have been allowed to function properly, they have been far, far superior to either ‘experts’ or to public polls at predicting events. There are two separate points to untangle here: why prediction markets are so accurate, and how they have tended to function in real life. Both points lend themselves to discussion of blockchains in the following section.
Why prediction markets are better than experts is simple, and the principle is widely understood in other contexts. The Wisdom of Crowds, after James Surowiecki’s 2004 book of the same name, is most easily conveyed as the principle that the mean of a large sample of estimates of some variable will likely be closer to the true value than the majority of the individual estimates. The key, and indeed a necessary condition, is that the participants’ estimates are independent and hence reflect a diversity of perspectives. This is the essential disqualifying factor for when crowds become exceptionally unwise, as in a stock market bubble, for example. But if the individual estimates draw on esoteric knowledge and insights, a large enough sample will tend to produce a desirable effect. Prediction markets can only function if this is the case: if everybody believes the same thing, then nobody will take the other side of any contract. But this still doesn’t answer why they are better than polls. Prediction markets can arguably be thought of as weighted polls, where the weightings reflect sincere belief, as captured by willingness to risk loss on the prediction.
More simply, we would say, willingness to bet. This has two mutually enhancing effects. The first was articulated most fully by Hayek in addressing the socialist calculation problem by describing markets as aggregation mechanisms for widely dispersed knowledge (see The Use of Knowledge in Society, and, Competition as a Discovery Procedure) The essence is that the possibility of profit entices those with unusual knowledge to engage in the market, hence moving prices in the direction that best reflects reality, despite the fact that no single person or even large subgroup of people have or could possibly have the relevant knowledge in its entirety. The counterpoint to the possibility of profit is the possibility of loss, which is equally valuable; as profit rewards and pulls in those with knowledge, loss punishes and pushes out those without knowledge. This is particularly valuable in the case of predictions as we want to weed out the pernicious cases of participants who have no idea what they are talking about submitting estimates for some reason other than contributing to the correct prediction — professional prestige, for example, or misguided institutional architectures that pressure participants to make meaningless predictions they would rather not. A bet is a tax on bullshit, goes the saying originating with economist Alex Tabarrok. Put up or shut up, goes another, attributed to many millions of people whose social groups tend to mouth off about sports with the intent to impress one another rather than to be correct.
This understanding of prediction markets leads back to an earlier point, that they have been allowed very limited scope to function in modern society due to widespread mores against gambling in the West. To the best of my knowledge, there have been three large scale attempts to run prediction markets: The Iowa Electronic Markets is a market run by the University of Iowa under an exemption from the Commodity Futures Trading Commission as it is ‘run for research and education purposes’. Bets are limited to $500 and primarily offered on political events, so it is not as robust as we might hope, despite having an impressive record against polls and experts. PredictIt functions similarly, and is exempt from CFTC sanctions on the condition that (basically) it doesn’t scale properly. Probably the most successful private effort was InTrade, which was run out of Dublin and effectively shut down by the CFTC (Buzzfeed documents the decline here) The final example, the Policy Analysis Market, run by DARPA out of the US Defence Department, was a proposal to use prediction markets to forecast terror attacks, which was shut down amid public outcry. I will mention as well that BetFair, substantially more reputable than InTrade ever was, could perhaps be considered a prediction market, but I would argue is really just a web-native bookmaker. The difference is largely academic in some regards, but consists primarily in the inability to use the bet as a security, which I argue is crucial for the functions I will propose below.
What do blockchains do for prediction markets?
Assuming that we now believe that prediction markets are very good at doing what they say on the tin, I propose that there are three reasons to be excited about how blockchains and prediction markets may come to interact. The first two I will discuss relate to how blockchains may enable better functioning prediction markets than previously existed, while the third goes the other way — I believe prediction markets may provide a key enabling feature for other, more exotic sounding functions of blockchains.
The first, and I believe arguably the most obvious and often only reason to believe in any particular public blockchain project, is that of censorship resistance. As mentioned above, most attempts at organising prediction markets skirt regulations on gambling. Given that any number of law enforcement agencies around the world could shut down an attempt to scale a prediction market on a whim, it is pointless for any corporation to devote capital to such an enterprise. If built on a public blockchain (and designed robustly to avoid extraneous attack vectors) this concern disappears immediately as it is unrealistic that the network can be compromised, barring a worldwide clampdown on any and all Internet activity, which seems rather out the question. I believe the most important effect will not be that such a corporation can now exist for profit — I think this is probably the most common misunderstanding of the potential of public blockchains. Insofar as existing for-profit network functions will be ‘replaced’, I think it is far more likely that what replaces them will essentially be automated utilities and not ‘companies’. Whatever new companies come into existence will rely on the fact that this function will be resistant to censorship, rather than providing it themselves.
However, this raises an interesting problem: before we get too excited about removing all meddling authority from the market, don’t we need an authority in this case, in order to judge what actually happened and hence who won the bets? The short answer is ‘no’, although the reason is a little complex, and, we will see, actually makes it theoretically superior to a centralised source: not only can blockchain-based markets not be censored, the truth can’t be corrupted (or rather, it is very, very difficult to corrupt the truth, as opposed to one central source who just needs to be bribed, or themselves have an interest in the outcome) Augur is one such blockchain prediction market, and I include a link to the white paper here (to be clear, this is not an endorsement of Augur the project, its current functionality, or REP the token — it’s just the most obvious thing to point to currently if readers want to follow up) In short, the ‘truth’ in question is crowdsourced when the event in question has occurred in real life, and those who ‘vote’ on the correct answer are paid out of the participation fee, very possibly the token native to the prediction market but not necessarily, while those who vote on the incorrect answer lose some staked value — the money / token / whatever that previously would have gone to the operator of the market being reserved for this pay-out. The idea here is that the ‘truth’ is an obvious Schelling point to gravitate towards for a great many voters whilst the ‘false’ outcome can only plausibly be pursued in bad faith. That doesn’t mean this won’t happen, and again I am talking in theory here, not about any current market. There are significant issues to be engineered around to successfully scale any attempt. But were this to ever succeed, we see that the widespread belief of the utility of the market is precisely what is leveraged to provide utility in the first place. This is similar to the traditional explanation of the pointlessness of 51% attacks on the bitcoin network — that the action itself would destroy the value of the network that the attack is designed to capture — although in this case directed at a far more specific goal than gaining the ability to alter the transaction record: verifying the truth of prior claims about the real world.
The second reason that blockchains enable more effective prediction markets is far more prosaic: that censorship resistance also implies equality of participation via scalability with no obvious upper bound. Anybody can set up a market without needing to make the case to a central party. This is less nefarious than the first instantiation of ‘censorship resistance’, but is key to the next section regarding smart contracts: in order to write a smart contract, you need to know it will immediately become active, and not need to apply to some central body to verify it. Utilisation of blockchains makes this kind of contract less like betting at a bookie and more like creating a web page. If you have access to a smart contract platform, nothing can stop you.
What do prediction markets do for blockchains?
Another reason to be excited about the combination relates to what is known as an ‘oracle’. To understand the importance of both oracles and prediction markets we first need to grasp the essence of a smart contract. Despite the high-technological undertones, ‘smart contracts’ are really little more than ‘automated contracts’. The idea is that any contract, stipulating in the abstract that, if A does x, then B will do y, need not rest on the interpretation of human arbiters of x and y, particularly if they are fully quantifiable, as many commercial contracts are; when A receives shipment, pay B $z. This is not (really) open to interpretation. And yet there is a layer of time and energy costs sunk in facilitating the contract. Prior to blockchains, this was inevitable, and it would seem a non-sequitur to even bring it up. But now, in theory, both parties can lock into a secure, transparent contract whereby the if-then clause is automatically executed with no intermediary deciding that this should happen.
The problem, as is probably obvious from the example above, is how to determine that A receives shipment actually happened, which would seem to put us back at square one. Taking a step back, it is easy to see that any smart contracts where the variables in the contract are ‘on-chain’, meaning relate to data or metadata existing elsewhere on the same blockchain, can be executed with no external input. Since ‘time’ is part of the metadata of a typical blockchain, we can imagine a fixed term annuity fitting this description as per the following pseudo-code: if receive $x from account A, then while t < t0, pay $y to account A and wait 1 month, which we remember will execute without fault given that the currency being used to pay is the essence of the blockchain itself. The reason the first example is tricky is that there is no obvious ‘on-chain’ way to determine that A received a shipment. This leads to the concept of an ‘oracle’, which is defined more by what it would ideally do than what it is; it reports off-chain (real world) events to the blockchain so that smart contracts can be fulfilled.
I propose that the behaviour incentivised in prediction markets would make for excellent oracles. Certainly not exclusively — we might find little reason to doubt MLB.com, The Met Office, or any other number of central reporting entities on questions that are both banal and for which an expensive apparatus for gathering the data already exists. But for many cases, I think this makes a lot of sense. To understand why, we need to think about what qualities the ideal oracle would have. There are obvious ones such as accuracy, provability, and consistency, which don’t really tell us much about how an oracle would work, but rather just how it must not. It is easy to see that what we are really asking here is what the ideal qualities of a ‘judge’ are, since all we require is somebody or something to judge that the conditions of a contract have been fulfilled. In fact, professional arbiters are one proposed solution to the oracle problem. I am not arguing against arbiters here, but I suggest that the core question that we will always return to, and which arbiters do not totally solve, is: how do we know it’s not lying? For any proposed solution, we will therefore need to ask, why might they lie? Any hypothetical entity with known members, goals, an identity, etc. — in essence with interests, will almost necessarily have potential reasons to lie, or an ability to be corrupted, in some or other situation. We need a source that is beyond the interests of any self-assembling subgroup of the general population. Perhaps a randomly selected subgroup, or one incorporating the beliefs of everybody, if such a thing even exists.
But we have covered how prediction markets verify truth in (ideally) incorruptible ways. We can be nearly certain that the verified outcomes of prediction markets are not lies. We cannot be completely certain, clearly, but I suggest far more certain than for any centralised authority who can be bribed or can otherwise act in bad faith. Prediction markets leverage their own utility to reward a distributed pool of participants for acting in good faith. In addition, the second benefit of censorship resistance means that anybody can set up the market, most probably the same entity writing the smart contract, while the first benefit guarantees that the contract will not be cancelled or the parties punished once live.
We can go even further than this, in fact. We can utilise not only the trustworthy verification of prediction markets, but also their predictive accuracy. One could write a smart contract that utilises not only the verification of prediction markets, but the prices on the prediction markets for the clauses of contracts in order to grant far more flexibility in judging and responding to probabilities than is currently possible at all. For example, one could write a contract for a clause confirmed in a year that will enter a sub clause if the price of that event in a prediction market goes above 90¢ (which is on-chain). The idea here would be that the initiators of the contract aren’t confident in their understanding of the likelihood of the outcome, but that if the crowd believes it is very likely, then one course of action is probably pointless and another should be pursued instead.
What might come from all of this?
It is a separate, open-ended, and fascinating discussion as to what economic activity is improved upon by the new capabilities of blockchains. I don’t want to get too far into this particular line of thought, but my general outlook is that to ask, ‘where are there network effects?’ is a naïve approach that confuses the causes and effects of networks. Network effects are (shock, horror) effects, whereas the cause is superior utility to customers over alternatives. A better question is, ‘where could there be network effects if a value-based friction to creating them could be overcome?’.
I think the combination of insurance, financial derivatives, and gambling is a good example. Although these products vary dramatically in terms of incentives and social characteristics of participants, and so on, these are all basically the same thing: they are bets. It is not commonly realized or thought about this way, but all forms of betting rely on network effects in that building a book requires (means to some extent) depth and liquidity in a market that is matched off at scale by the insurer/market-maker/bookmaker and oftentimes presented to the customer as simply a ‘product’ concocted by the central entity, rather than the real raw ingredients.
I propose that there is a decent long run probability that blockchain prediction markets consolidate these three functionalities and provide a better service to users. The reason for this is essentially that there are significant value-based frictions to realizing network effects in all three far, far greater than currently exist, and that cannot be realized via a centralized party.
It is only currently possible to buy a future on an event if the event is known to be of such wide interest that a deep and liquid market can be virtually guaranteed to exist. This is because the bookmaker must set the initial odds and risk loss as a covering counterparty should they be wildly off the mark in the initial period before liquidity coming in starts to move the odds. If you can virtually guarantee eventual liquidity then this is less of an issue. So, sports and politics and the like are fairly safe to make markets in. (n.b. I’ll say ‘buy a future’ to cover all three options rather than ‘place a bet’. The latter is probably more easily understood but is unfortunately deemed unsavoury when the purpose is something other than entertainment)
There is one exception to this general rule: highly esoteric events can give rise to markets for futures but usually only if the buyer or seller of the future is very well capitalized. This presents two ways to overcome the risk to the bookmaker, broadly speaking: stake collateral or pay a fee. So, this might be a gambling addict pledging his car to a gangster to get in on a dog fight, or it might be a retainer paid by a wacko hedge fund to hold a credit default swap with an investment bank. We needn’t get too much further into this as it is clearly getting a little divorced from anybody’s real or relevant experiences — and I promise I’m only aware of this stuff because I’ve seen it in movies — but the general point is that the undercapitalized can only buy futures in markets that are strongly suspected in advance to be highly deep and liquid.
Prediction markets based on blockchains remove the counterparty risk to the bookmaker because there is no bookmaker in the first place: this role is automated. What this means is that there are no restrictions on what kinds of futures markets can exist: the relatively undercapitalized can make a market in anything at all. On the one hand this means that insurance, derivatives, and gambling can be done slightly more efficiently, but this brings us back to the point above about the network effects: this is not enough. It would have to be much, much better rather than just a little bit, to have any disruptive potential on this account.
Where there is enormous disruptive potential is in allowing futures contracts on anything at all. Who would want a future? Maybe an Uber driver realizes they make 3 times as much on days it rains and wants to hedge against it being too sunny for too long. Maybe a surf instructor has exactly the opposite problem and happily sells the future to the Uber driver buyer. Maybe you work in real estate in New York as well as having a mortgage on your house there, and so in a sense are massively leveraged to the New York real estate market. If something happens to the industry, your career will tank in perfect correlation to your home equity. Maybe you want to hedge that away, or maybe you explicitly want exposure to London real estate instead, for whatever reason. Maybe you are an investor in Tanzania and want to cross-hedge your exposure to Indonesian and Peruvian investments. A bank will do this for you, but for an enormous fee that is likely to outweigh any portfolio benefit.
What I am driving at here is any set of cash flows that are insecure or unpredictable for any reason whatsoever can now be hedged. The fact that people might be explicitly gambling rather than hedging simply adds depth and liquidity to make the whole thing even smoother. Clearly this could not happen before because Betfair will not sell you a weekly rolling betting line on the weather in Long Beach, nor will JPMorgan concoct you a derivative on the London property market. They might if you offer to pay them lots and lots of money, or stake lots and lots you already have. But if you don’t have lots and lots then that’s the end of it. In fact, even this argument misses the point slightly. In all these hypotheticals, you have to go and check if such a market exists, and ask that one be created if it doesn’t. In a blockchain prediction market, you can simply create markets that don’t exist without anybody considering their counterparty risk, assessing liquidity, or whatever else they think is important before they grant permission. A blockchain prediction market allows for the exchange of futures contracts by providing a relatively enormous platform for participants to find one another rather than having to go through one of a select few central exchanges, which are necessarily very large to begin with due to the subtle network effects of that business. A nice way to think about this is perhaps to compare terrestrial television with YouTube. As many people watch YouTube as watched terrestrial television 20 years ago, adjusting for population growth, but via YouTube they can all interact with each other with minimal friction. Very few people made content for terrestrial TV but anybody can make it for YouTube. It would be difficult to argue that creativity has not flourished as a result. Like YouTube is for video,* prediction markets are truly open platforms for futures.
*by ‘is’ I mean ‘should be’, but I didn’t want to disrupt the flow …
Concluding thoughts and further reading
I have tried here to sketch a case for the potential utility of prediction markets to blockchains, and blockchains to prediction markets, with as little technical detail as possible. Those wanting to dive deeper are encouraged to read the Hayek essays mentioned on market organisation, or the work of Robin Hanson on prediction markets specifically, popularised here, and fully organised here. For background on the statistical meaning of prediction markets, see the National Bureau of Economic Research papers here and here (there are many more on the same topic but I found these to be the most helpful).
follow me on Twitter @allenf32