You Can't Design an Insider-Proof Prediction Market

What everyone is arguing about

The framing in every commentary I have read is some version of: prediction markets are accurate, useful, and growing, but they have an insider-trading problem, and the fix is the right combination of surveillance, position limits, KYC, and a statute. The CFTC's proposed rule leans this way. So does the Senate bill (S.4060, the Prediction Markets Security and Integrity Act). So did the Commission's February advisory, which said it can pursue trading on material non-public information as a fraud under the Commodity Exchange Act.

All of that may be good policy. None of it touches the thing that actually makes these markets fragile. The insider-trading problem in a prediction market is not a compliance gap that better rules will close. It is a property of the payoff. To see why, you have to stop thinking about prediction markets as a new kind of betting and start thinking about them the way you would price any other book you have to make against informed flow.

A prediction market is a step function

Take the canonical contract. It pays $1 if some event happens and $0 if it does not, and it resolves on a date.

That looks innocuous. It is not. Compare it to a share of stock. A stock has a continuous, never-fully-revealed fundamental value. Nobody knows it exactly, the best informed trader has a slightly better noisy estimate than the market, and that estimate gets revised forever. Information arrives in a stream, prices move continuously, and even an insider is trading on a better guess, not the answer.

A prediction-market contract is the opposite. Its value resolves to one of two numbers on the back of a single discrete information event — a court ruling, an election call, an arrest, a strike, an earnings print, a Fed decision. Before the event the public price is a probability; after it the contract is worth exactly $1 or exactly $0. Crucially, for many of these events the outcome is already determined before the public learns it. The military operation is already planned. The decision is already drafted. The deal is already signed. Someone, somewhere, is not estimating the probability — they are holding the answer to a coin that has already been flipped and not yet shown.

That person is not a better-informed trader in the equities sense. They are trading against a price that is wrong with near-certainty, in a known direction, for a known amount. That is the cleanest edge in all of markets, and the market maker on the other side has no way to tell them apart from the noise until the contract resolves.

The adverse-selection math

This is the Glosten-Milgrom problem, and the binary payoff makes it about as stark as it ever gets. Set up the standard sequential-trade model. The market's public belief that the event happens is π. A fraction α of order flow is informed — these traders know the resolution and only ever trade the correct side. The remaining 1 − α is noise: uninformed flow that buys and sells with equal probability.

A competitive market maker earns zero expected profit, so the ask must equal the expected value of the contract conditional on receiving a buy order. A buy order is more likely to come from an informed trader when the true answer is YES, so the very act of being lifted is bad news for the maker:

Evaluate it at a coin-flip market, π = 0.5, and it collapses to something you can hold in your head:

The break-even full spread is just α, in dollars, on a contract that is worth at most one dollar. If a tenth of your flow knows the answer, you have to quote a ten-cent spread on a fifty-cent contract just to break even. Put the numbers in a table:

Now anchor that against real markets. A liquid equity trades on a spread of a basis point or two. A thin small-cap might cost you tens of basis points round trip. A binary event market where one order in ten carries the answer needs two thousand basis points just to not lose money. The adverse-selection floor on these contracts is one to three orders of magnitude above anything in equities, and it is set by the payoff, not by the venue.

The reason is the term you do not see written down: the value range the informed trader is exploiting. In Glosten-Milgrom the adverse-selection spread scales with α · (V_high − V_low). In equities that bracket is small — information is incremental, the insider moves your fair value by a few percent. In a binary event market V_high − V_low is the entire contract, a full dollar. Same informed fraction, the whole notional at risk on every toxic order.

Why equities survive informed trading and event markets don't

Equity markets have lived with insiders for a century without seizing up, so it is worth being precise about why prediction markets are different. The reference model is Kyle (1985). A single informed trader who knows a continuous liquidation value does not dump their whole edge at once. They trade gradually, because trading moves the price — market depth is finite, the price impact coefficient λ is positive — and every unit they buy worsens their average price and leaks their information into the tape. The market maker learns from order flow and updates. The informed trader's profit is bounded, the market clears, depth survives. Informed trading is a tax the market can carry because the edge is a noisy estimate that erodes as it is exploited.

In an event market the edge does not erode, because there is no uncertainty in it to erode. The insider is not refining an estimate of a moving target; they are holding a result. They keep buying until the price reaches the truth — to a dollar, or to zero — and every contract along the way is bought below its certain value. The price impact they create is the market becoming accurate. The area between the stale public price and the true outcome is exactly the insider's profit, and it is exactly the maker's loss. The two are the same quantity.

That equivalence is the whole essay, so let me state it plainly. In a prediction market, the trade that makes the price correct and the trade that drains the liquidity provider are the same order. You cannot have one without the other.

Why no design prices the insider out

People keep proposing to engineer the problem away. It doesn't work, and you can see why by walking the two designs that actually run.

The subsidised AMM (LMSR). Hanson's logarithmic market scoring rule is the academic darling, and it is what people reach for when they want a market maker that is always there. It has a beautiful property: the operator's worst-case loss is bounded by a liquidity parameter b. But "bounded loss to the operator" is not a feature against an insider — it is a budget the insider gets to drain. The deeper you make the market (larger b, tighter prices, more size absorbed), the bigger the cheque you have written to anyone who shows up holding the answer. Subsidised liquidity against certain-informed flow is just a transfer to that flow. You can only shrink the transfer by making the market thin and expensive, which is the thing you built the AMM to avoid.

The order book (Polymarket). Polymarket does not use an AMM; it runs a central limit order book, off-chain matching with on-chain settlement, and pays maker rebates to professional firms to keep two-sided quotes up. That is the right design for where its volume concentrates. But it relocates the cost; it does not remove it. An insider simply lifts the resting offer. The maker who posted it has the same two choices every maker against toxic flow has ever had: quote wide enough to survive being picked off, which kills the depth, or quote tight and subsidise the informed trader out of the noise traders' losses and the rebate. The rebate program is, mechanically, the platform paying makers to keep standing in front of the toxic flow. It works precisely as long as the noise and attention volume is large enough to cover it.

I have made this argument before in two other settings and it is the same argument each time. On-chain options vaults bleed to informed flow during oracle lag because the pool quotes a stale price an arbitrageur can pick off — I called it loss versus rebalancing under oracle latency, and the fix was to make the quote defend itself, not to pretend the lag doesn't exist. A prediction market is the limiting case of the same structure: the "oracle lag" is the gap between when an event is determined and when the public learns it, and the informed counterparty is whoever lives inside that gap. The difference is that in options you can condition the quote on inventory and claw most of it back. Here you cannot, because the mispricing is not noise around a known value — it is the answer.

The CFTC is cutting along the wrong axis

The proposed rule's instinct is to permit contracts that aid "price discovery" — it treats sports as largely fine on those grounds — and to ban contracts "vulnerable to manipulation." That sounds reasonable and it is the wrong cut, for the reason the last section built. In a binary event market, price discovery and toxicity are not separable. The informed order is the one that makes the price accurate. Regulating for price discovery while regulating against the informed flow that produces it is asking for the destination without the road.

This is a sharper version of the question I worked through in the Jane Street piece. There, the genuinely hard line was between aggressive market-making that reveals a true price and manipulation that creates a false one — two different trades that can look similar on the tape. In a prediction market that ambiguity disappears, but not in a way that helps you: the informed trade is unambiguously revealing a true price and unambiguously draining the maker. There is no false price being created. The insider is making the market more accurate. That is what makes the harm so hard to legislate against — you are trying to outlaw the mechanism by which the product does the one thing it is sold for.

And notice what the rule actually keys on. "Vulnerable to manipulation" gets attached to a topic — politics bad, sports fine. But vulnerability is not a property of the topic. It is a property of the information structure: does anyone hold private knowledge of a discrete outcome? A football match with no inside information is clean regardless of how you write the contract. A market on "will the CEO be fired on Friday" is structurally toxic no matter how carefully you draft it, because someone in that building already knows. Regulating the subject line instead of the information asymmetry will permit toxic markets that happen to look benign and ban benign ones that happen to look dangerous.

The trilemma

It helps to name the trade-off, because once you see it you stop looking for the fix that isn't there. A prediction market wants three things at once:

• Accuracy — the price tracks the true probability, including on events where some people know more than the crowd.
• Liquidity — tight spreads and real depth, so size can trade.
• Integrity — insiders cannot systematically extract from everyone else.

You can have any two. Not three.

Accuracy + liquidity means you are paying the insiders. The price is accurate because informed flow corrects it, and it is deep because makers eat the loss, funded by noise traders and rebates. This is Polymarket on geopolitics. The uncanny accuracy everyone celebrates is partly the footprint of people who knew, and the depth is the noise traders' subsidy in disguise.

Accuracy + integrity means you stop the informed flow from trading size — hard position limits, identity, surveillance, slow settlement windows. You can keep the price honest and keep insiders from cleaning up, but only by throttling exactly the participants who carry information, and the market goes thin. This is the regulated direction of travel, and thin is the price of it.

Liquidity + integrity is only achievable where there was no private information to begin with — sports with public statistics, the weather, a macro print everyone reads at the same instant. Deep, fair, and accurate, but trivially so: there was nothing to discover. This is the safe harbour, and it is also the boring one.

So the three things prediction-market advocates claim in the same breath — eerily accurate, deep and liquid on everything, and fair — cannot all hold on the markets that actually matter. The markets that are deep and accurate (broad elections, macro) are the ones where information is widely distributed and the trilemma doesn't bind. The markets that are exciting (a single arrest, a ruling, a takeover, a strike) are structurally thin and toxic, and there is no AMM, no rebate schedule, and no oracle clever enough to change that.

Where this leaves the category

Liquidity will keep concentrating where information is symmetric, because that is the only place a maker can survive. The headline single-event markets will stay shallow and will keep getting picked off, subsidised by attention flow until the attention moves on — which is the same dynamic I wrote about as the bid-ask spread of attention: a market kept alive by the very flow that makes it lossy to provide into.

Every live policy response — the CFTC rule, the Senate bill, the February advisory, India's block — is a surveillance response, and surveillance acts on α. It raises the cost and the risk of being the insider, so fewer people dare, and the toxic fraction shrinks. That is worth doing. But it operates on who is willing to trade, never on what the trade is worth. The payoff is still a step function, the edge is still the whole dollar, the maker still cannot tell the insider from the noise until resolution. There is no insider-proof prediction market. There is only an insider-deterred one, and deterrence is a dial, not a fix.

What I'm not claiming

I am not claiming prediction markets are useless. On genuinely public-information events they aggregate dispersed belief better than almost anything we have, precisely because there is little private information to extract — the trilemma is slack and you get accuracy and liquidity for free. That is a real and valuable product.

I am not claiming most of the flow is insiders. The overwhelming majority is noise, sentiment, and entertainment, and that is the whole point — it is the subsidy that keeps the lights on. The argument is about the marginal toxic order, not the average one.

I am not claiming surveillance is pointless. Shrinking α is the only lever that exists, and pulling it is sensible. I am saying it is a lever on participation, not a repair of the mechanism, and it should be sold as such.

And the binary contract is a stylisation. Real markets are scalar and multi-outcome and messier than $1-or-$0. But any contract that resolves on a discrete information event inherits the step in its payoff, and with the step comes the same adverse-selection floor. The shape is what matters, and the shape does not change. Wherever a single discrete outcome determines the payoff and somebody can know it early, the market is structurally toxic — and the better it is at discovering the price, the more completely it pays the person who already knew.