Why Your Prediction Market Is Only as Good as Its Market Maker

Most prediction market operators obsess over the wrong things at launch. They pour resources into UI polish, onboarding flows, and token incentive structures, then go live with thin order books, broken price signals, and traders who show up once and never come back. The culprit is almost always the same: no serious market maker. Not a bad market maker. Not an expensive one. Just the total absence of one, or a placeholder that looks like one until you stress-test it.

This is the central architectural decision that separates a prediction market that becomes a genuine information aggregation engine from one that's just a speculative toy with a nice landing page. Getting this right, and understanding why it matters, is the difference between a platform people cite and a platform people forget.

What a Prediction Market Is Actually For

Before we get into mechanics, it's worth being explicit about the purpose. Prediction markets exist to aggregate dispersed private information into a single public price signal. The price of a YES contract on "Will X happen by December?" trading at $0.63 should mean something: the market collectively believes there's a 63% probability of X happening. That's the product. That's the whole point.

Every design decision on a prediction market platform should be evaluated against this question: does it make the price more or less accurate? Liquidity provision sits at the center of that question, because without it, prices are not aggregated beliefs, they're just the last person's guess, frozen in time until someone decides to trade again.

The forecasting accuracy of a prediction market is only as good as its market design, and that design is only as good as its market mechanism. This is the operational reality that every platform eventually discovers, usually by watching their user retention numbers collapse.

The Spread Is a Tax on Being Right

Start with the most immediate, visible problem: the bid-ask spread. In a prediction market without proper liquidity provision, spreads are wide, sometimes grotesquely wide. A contract where the true probability might be around 60% ends up with YES offered at $0.75 and the best bid at $0.45. What happens?

The informed trader, the person who actually has a useful signal, who has done the research, who should be moving the price toward truth, does the math and walks away. Buying at $0.75 when you believe the fair value is $0.62 is a losing trade. Selling at $0.45 when you believe it's $0.62 is also a losing trade. There is no entry point that's rational. So the person with information, the entire reason the market exists, does nothing.

Meanwhile, the only people trading are the ones who either haven't done the math, are gambling for entertainment, or are sophisticated enough to take the arb when the price is obviously broken. None of these people are contributing to price discovery. They're either burning money or extracting it, but either way, the price isn't moving toward truth.

A professional market maker narrows that spread to something like $0.60/$0.64, and suddenly the informed trader has a rational trade. They buy. The price moves. The market learns something. This is the core mechanic, everything else is downstream of it.

A Taxonomy of Market Making Models

Not all market making is the same, and the choice of model has real consequences for platform health. There are three primary approaches, each with distinct tradeoffs.

Automated Market Makers (AMMs) using LMSR. The Logarithmic Market Scoring Rule, developed by Robin Hanson, is purpose-built for prediction markets. LMSR maintains a cost function over outstanding share quantities and quotes prices as its gradient. Its most valuable property is a provably bounded worst-case loss: the platform operator knows in advance exactly how much the market maker can lose, which makes capitalization and risk management tractable. Prices are always continuous and always sum to one across all outcomes, which means they can be cleanly interpreted as probabilities. The downside is that LMSR is less adaptive under informed flow: its fixed parameterization cannot respond to adverse selection in real time, which is why it underperforms relative to more dynamic mechanisms when the market is dominated by informed traders rather than noise.

Constant Product Market Makers (CPMMs). The market maker maintains an invariant: the product of YES and NO token quantities is always constant. This is simpler, battle-tested in DeFi, and easy for traders to reason about. It works acceptably in two-outcome binary markets. It becomes cumbersome in multi-outcome markets, and it doesn't have LMSR's bounded-loss guarantee, a market maker using a CPMM can, in theory, lose more than intended if the market moves dramatically in one direction.

Inventory-based professional market making. The MM quotes prices based on their inventory position, their estimate of true probability, their model of adverse selection in the current flow, and a spread that compensates for all the risks they're absorbing. Unlike AMMs, this approach is dynamic and discretionary, the MM can pull quotes, widen spreads, or adjust position limits in real time based on what they're seeing in the order flow. Research comparing LMSR to Bayesian market makers (BMM), a more adaptive, information-based approach, shows that BMMs can offer significantly better price stability and lower expected loss when controlling for liquidity, though unlike LMSR they don't guarantee bounded loss. This illustrates the fundamental tradeoff: adaptability to market shocks versus convergence stability during equilibrium.

The practical reality is that the best setups use hybrid approaches: an AMM to ensure baseline continuous liquidity, with a professional MM layered on top to handle the edge cases, tighten spreads during high-volume periods, and actively manage inventory across correlated markets.

The Adverse Selection Problem

Here's what makes market making on prediction markets genuinely hard, in a way that's different from equity market making: your informed traders are sometimes people who actually know the answer.

In equity markets, adverse selection means you might trade against someone who processed public information faster than you. In prediction markets, especially political or event-driven ones, adverse selection can mean trading against someone who has genuine non-public information about the outcome. A market maker on "Will Country X announce a policy change this week?" might be trading against a diplomat who was in the room. A market maker on "Will Company A announce a merger?" might be trading against someone with actual insider knowledge.

Polymarket saw this play out dramatically in 2026, when a few accounts potentially made over $1.2 million betting on the killing of Iran's Supreme Leader Khamenei, with trades funded hours before the event. The market maker on the other side of those trades didn't have a chance. This is the worst form of adverse selection: a counterparty whose signal is so strong that they can rationally allocate maximum capital across every related market simultaneously.

The implications for MM design are significant. You cannot treat prediction market order flow as random, or even as directionally skewed by skill. You have to model it as potentially adversarial. This means:

• Spread widening on event proximity. As resolution approaches and the underlying event is actually about to occur, the information asymmetry between insiders and the MM spikes. Spreads should widen in response, not tighten.
• Volume-based alerts. Sudden spikes in one-sided volume in a previously quiet market are a signal. A professional MM has systems that flag this and reduce position limits automatically.
• Cross-market correlation monitoring. If you're getting hit on multiple correlated markets simultaneously, someone knows something. That's not a pattern that emerges from noise.
• Resolution-outcome feedback loops. Over time, a serious MM tracks their accuracy: when they got run over, what the flow looked like beforehand, and what that implies for future quote adjustment.

If an MM knows they're being consistently picked off by informed insiders, they have only two rational responses: widen spreads dramatically or leave the book entirely. Either outcome is catastrophic for platform quality. This is why the relationship between platform design, resolution integrity, and MM health should be one system.

Edge Cases That Break Naive Approaches

Resolution ambiguity. This is the single most underappreciated failure mode in prediction market design. A market maker taking on inventory in a contract with vague resolution criteria is exposed to a risk that has nothing to do with the underlying event, namely, that the resolution authority interprets the criteria in an unexpected way. "Will the economy improve in 2026?" is a philosophical question dressed up as a contract. A sophisticated MM will either refuse to quote it, or price in a substantial ambiguity discount that effectively destroys liquidity. The lesson is that resolution criteria must be precise, objective, and resistant to manipulation before a MM will, or should, provide real liquidity.

Mere suspicion of manipulation, even without actual manipulation occurring, is enough to erode trust and lead to the adoption of suboptimal decisions based on market prices. Resolution disputes that become public destroy the credibility of every future market on the platform, not just the one in dispute.

Binary blowups near expiry. As a contract approaches its resolution date, something mathematically predictable happens: the delta explodes. A contract sitting at $0.50 with two days left until resolution can swing to $0.02 or $0.98 on a single news item. The price sensitivity to new information is at its maximum precisely when the MM's inventory is most exposed. Naive LMSR implementations don't account for this, the liquidity parameter b is typically fixed, which means the market maker is providing the same depth at day one and hour one before resolution. A professional MM recalibrates dynamically: position limits shrink, spreads widen, and the MM becomes more aggressive about flattening inventory as expiry approaches. Platforms that let AMMs run without this adjustment will watch their MMs get systematically picked off in the final hours of every high-profile market.

Correlated markets and portfolio exposure. On any reasonably active platform, a market maker is managing a portfolio of positions across dozens or hundreds of simultaneous markets, many of which are correlated. "Will Candidate A win the election?" and "Will Policy X pass?" are not independent. "Will Company A acquire Company B?" and "Will Company B's stock close up this week?" are not independent. A MM who treats each contract in isolation will find themselves with a portfolio that is massively net long or net short a single underlying factor, political outcome, macro data, a corporate event, without realizing it. When that factor resolves, every correlated position moves simultaneously and the losses become multiplicative.

Sophisticated inventory-based market making requires active correlation modeling and portfolio hedging. The MM needs to know their net exposure to each underlying factor at all times, and maintain hard limits on that net exposure regardless of what individual contract quotes would suggest.

The bootstrapping problem. New markets on a platform face a classic cold-start dilemma: traders don't come because there's no liquidity, and liquidity isn't provided because there are no traders. This is a structural failure mode that compounds over time. The first few traders in an illiquid market set a price that is essentially arbitrary. That arbitrary price becomes an anchor. Subsequent traders who arrive without strong priors adjust toward the anchor rather than their own signal. The market converges on the anchor instead of truth, and never recovers. Automated market makers always offer both buy and sell prices even with no participants, which is precisely why they're valuable for bootstrapping. But the capitalization level, the b parameter in LMSR, determines how much inventory the AMM is willing to absorb before prices move. Setting it too low means prices move too fast and early traders have outsized impact. Setting it too high means the operator is subsidizing liquidity at a loss indefinitely. This is a calibration problem that requires ongoing attention, not a one-time parameter choice.

Multi-outcome markets. Most of the discourse around prediction market MM is focused on binary YES/NO markets. But many of the most interesting markets are multi-outcome: "Which of these five candidates will win?" or "What will Q3 GDP growth be?" Multi-outcome markets create a combinatorial complexity problem for market makers. Keeping prices consistent across all outcomes (they must sum to one), managing inventory in each outcome token, and hedging cross-outcome exposure simultaneously is a fundamentally harder problem than binary market making. In practice, most platforms extend LMSR to multi-outcome markets via a softmax cost function, which handles an arbitrary number of outcomes without requiring parallel mechanisms. Research-stage frameworks (e.g., Othman et al.'s liquidity-sensitive LMSR) go further, allowing liquidity providers to submit arbitrary cost functions across the entire price space, but these remain largely theoretical. Few operators think through this before launch. Most deploy a binary AMM, realize it doesn't extend cleanly to multi-outcome markets, and then either avoid those markets entirely (sacrificing the most interesting use cases) or deploy them without proper liquidity (creating exactly the broken price discovery problem we started with). 

What Bad Liquidity Provision Looks Like at Scale

You've seen these platforms. You may have built one. And the symptoms are always the same:

• Markets where the last trade was three days ago, with prices frozen at wherever the AMM initialized.
• Order books with a single resting limit order that some retail user placed six weeks ago and never canceled, sitting ten ticks away from any reasonable price.
• Prices that don't move when directly relevant news breaks, then gap violently when someone finally decides to trade through the stale book.
• "Resolution" events where the market has been sitting at $0.50 for a week, then instantly jumps to $0.95 or $0.02 in the last hour because no one was updating prices in real time.
• High-profile markets that attracted real attention, showed good prices early on, but now have zero depth because the initial LP pulled their funds after the first month.

These are problems that are epistemically harmful. Analysts, researchers, and journalists cite prediction markets when the prices are meaningful and updating in real time. Nobody quotes a market that hasn't traded in a week. Nobody builds policy tools on top of a price signal that gaps by 40 points overnight. The platform's credibility as an information tool, its core value propositions, destroyed by exactly these conditions.

There's also a secondary effect that's harder to see but just as damaging: platform reputation in the market-maker community. Professional MMs talk. If your platform is known for ambiguous resolution, inconsistent rules, or treating MMs as a cost center rather than a core partner, you will have a very hard time attracting serious liquidity providers. And once you have a reputation for that, it compounds: worse MMs mean worse prices mean fewer serious traders mean less fee revenue mean less budget for better MMs.

The Operator's Strategic Framing

Prediction market operators tend to think about market making the way they think about AWS costs, as a line item to minimize. This is exactly wrong. A market maker is not a vendor. They're infrastructure with aligned incentives, or they should be.

The right structure is one where the MM profits from spread capture and calibration quality, the platform gains accurate prices and genuine volume, and traders get a fair, liquid venue. This is not a zero-sum arrangement. A MM who is good at their job makes money precisely because they're providing something valuable: the absorption of uncertainty that everyone else is trying to offload.

The wrong structure is paying a flat fee for someone to run a basic bot that posts stale quotes based on external prices, calls it market making, and invoices you monthly. It looks like liquidity until someone actually tries to trade size, and then the bot either disappears or the spreads blow out to 20 points.

What a serious MM partnership actually provides, beyond quotes:

• Market design feedback. A good MM will immediately identify resolution criteria that are exploitable or ambiguous. They have direct financial incentive to surface these problems before launch, not after.
• Manipulation detection. Because they're monitoring flow continuously, professional MMs often detect unusual trading patterns, potential front-running, coordinated position-taking, wash trading, before the platform operator does.
• New market viability assessment. Not every event makes a good prediction market. A MM with experience across many markets will tell you which categories attract informed traders, which attract noise, and which are structurally illiquid regardless of how you design them.
• Pricing calibration for novel contracts. When you're launching a genuinely new type of market, something with no historical analog, a professional MM can model fair value and uncertainty in ways that an AMM simply cannot. The AMM has no prior. The MM can construct one.

This advisory relationship is genuinely valuable and largely unpriced in most MM arrangements. Operators should be structuring these partnerships to extract this value explicitly, not just paying for quotes.

The Credibility Flywheel

There's a compounding dynamic that's easy to miss when you're focused on individual market quality. Prediction markets that have accurate, liquid, continuously updated prices get cited by serious people, researchers, forecasters, journalists covering the relevant domain. Being cited brings more informed traders to the platform. More informed traders means better prices. Better prices mean more citations. This is the flywheel that every prediction market operator is trying to spin.

The flywheel runs in both directions. Platforms with stale, unreliable prices get cited as examples of what not to trust. That reputation is nearly impossible to reverse because the traders who would repair it, the calibrated forecasters, the domain experts, have long since stopped bothering with the platform.

The market maker is the flywheel's ignition. You can have everything else right, clean UI, good contract selection, strong community, fair resolution, and the flywheel will never start spinning without someone willing to stand in the middle of the book, continuously, and make prices that mean something.

Every other decision you make about your prediction market platform is downstream of this one.

The mechanics described here - LMSR parameterization, correlated portfolio management, expiry-adjusted quote strategies - are implementation details that vary by platform.

‍