Four ways to make a market

Anyone can create a prediction market now. Almost nobody can make one.

Creation is a database row. Making is the harder thing: standing between two people who want opposite sides at different times, and being willing to be wrong for a living. Every structure in this space is an answer to that problem, and every answer costs someone something.

This is the map. It is the background for every comparison in the rest of the library, and it is deliberately written so that the vested mechanism looks like one column among five rather than the point of the exercise.

Order books

How it works. Buyers post bids, sellers post asks, the venue matches them. The venue takes no position.

Who makes the market. A professional market maker, quoting both sides and earning the spread. On real venues they are there because the venue recruited them, and frequently because the venue pays them liquidity incentives.

What it costs. For the venue, incentives and integration. For the market maker, adverse selection — the risk of being picked off by someone who knows more.

What it buys. Everything, when it works. Continuous two-sided quotes, a real price, native exit at any moment, and hedging. There is a reason the deepest prediction markets in the world are books.

Where it stops. It cannot cold-start. An empty book has no price at all — not a bad price, no price. And a professional will not quote regional, niche, or machine-made claims, because the expected spread income does not cover the attention. As the paper puts it, Polymarket's order books are "the deployed existence proof of how far books do reach, and their economics still stop well short of the long tail."

The failure is not gradual. Market ten is fine and market ten thousand has a thirty-point spread or nothing at all.

Constant-function AMMs

How it works. A curve holds reserves of both outcomes and quotes a price off their ratio. Trades move along the curve.

Who makes the market. A liquidity provider, who deposits before anyone trades.

What it costs. On a persistent token pair, an LP's problem is impermanent loss, which is unpleasant but symmetric. On an expiring event contract it is worse and structurally so: as information arrives, the pool mechanically rebalances into the losing side, and at resolution that side is worth zero. The LP absorbs it. Paradigm's pm-AMM (White & Diamandis, 2024) redesigns the curve specifically for expiring claims and reshapes that loss profile — but it still requires a funded liquidity provider per market, which is the constraint that binds.

What it buys. Continuous quotes, exit, hedging, and no matching engine to run. Cheap to deploy, easy to reason about.

Where it stops. Someone has to fund the curve for every single market, and their expected return on that funding is negative.

Scoring-rule market makers

How it works. The LMSR (Hanson, 2003) and its relatives quote from a cost function rather than an order book. Price emerges from the rule; the market maker is an algorithm.

Who makes the market. Whoever funds the liquidity parameter b.

What it costs. The worst-case loss is bounded at b · ln n, which is genuinely elegant — you know your maximum downside before you start. But it is still a subsidy, and it is per market.

What it buys. The best pricing theory in the space, with utility and axiomatic characterizations behind it. Continuous quotes, exit, hedging.

Where it stops. The bound is a ceiling on the loss, not an escape from it. Multiply b · ln n by ten thousand markets and you have a budget, not a mechanism.

The paper is careful here, and it is worth repeating the care: the subsidy need not come from the venue. Manifold runs creator-funded AMM liquidity at scale. That puts the scoring-rule family in exactly the same funding position as the vested seed — per-market capital at risk, supplied by whoever wants the market to exist. The comparison is therefore about incidence, not about whether a subsidy exists.

The classic parimutuel

How it works. Everyone's money goes into a pot. The event resolves. Winners split the pot pro-rata, less takeout.

Who makes the market. Nobody. That is the entire appeal.

What it costs. No market maker, no LP, no operator risk, no curve. The oldest structure in betting, running at scale on racetracks for a century.

What it buys. It works from the first dollar with no counterparty search, which is exactly the property the long tail needs.

Where it stops. Two places.

First, it pays a flat multiple regardless of when you were right. Someone who committed an hour out and someone who committed three seconds before the whistle are paid identically, and the second is paid out of the first. Measured: the last decile of entrants captures 13.20% of the losing pool, and the buzzer strategy has an unconditional expected value of +25.27% pre-fee.

Second, and consequently, venues cope by closing entry before the event — giving up continuous trading, which was one of the things a market was supposed to be. That patch has its own article.

There is also a first-dollar subtlety usually glossed over: a classic pool accepts one-sided money, but if the other side never arrives it hands your stake back at 1× less takeout. That is an escrow with a fee, discovered at settlement.

The fifth column

The vested parimutuel keeps the pool and changes one moment: the losing money is assigned to the opposing positions when it arrives, not at resolution — with a matching constraint that keeps the assignment sound.

Who makes the market. The seeder, who stakes every outcome in a reserved slot and recovers at least that stake in every branch.

What it costs. Fees, carry on locked capital, and resolution risk. Not an expected loss to informed flow.

What it buys. First-dollar cold start, exact dilution protection, no lock window, and O(1) settlement that runs on-chain.

Where it stops. No native exit, no live late price, no hedging, and — with finite κ — an absorbing failure in n-way markets.

The whole map

Read the bottom three rows together, because they are the actual argument. The three structures that give you exit, hedging and a live price all require per-market capital that is spent. The two that need no per-market capital give up all three. The vested pool moves per-market capital from spent to parked, and pays for it with those same three things.

That is a trade, not a dominance claim. Which side of it you want depends entirely on which markets you are trying to run — which is the adoption question.

The five requirements, and two more

The Melee Markets litepaper frames five requirements this design adopts without modification:

• R1 continuous trading
• R2 first-dollar cold start
• R3 no subsidy
• R4 entry-time price integrity
• R5 profitable passive bootstrapping

The paper adds two:

R6. Verifiability. A permissionless market whose pricing rule cannot be inspected has reintroduced the trusted operator through the back door. A mechanism asking participants to accept its solvency on the strength of the designer's own testing is offering a promise rather than a guarantee, however carefully that testing was done.

This is a pointed requirement and the paper is precise about its target. The Melee litepaper states of its own design that "the PMM's production curve family, parameter schedule, and rebalancing implementation are proprietary and are not disclosed." The disagreement is only with that — a non-public curve is not necessarily an unsound one, and the paper makes no claim that it is. The objection is that a permissionless venue's soundness should not require taking the venue's word for it.

R7. A structural origin for the first dollar. Floors make early capital safer; they do not summon it. A venue hosting millions of markets needs a participant class for which discovering, pricing, and seeding a brand-new market is cheap and systematic. That is the agent argument.

Next: Against the classic parimutuel — the comparison where the case is strongest.