Whether a prediction market makes money is decided before it opens

Start with where the money comes from. The fee monetizes retail round-trip flow, and nothing else — informed flow is fee-immune by construction (the example below shows why) and causes all the inventory loss. That settles the shape of the business: order book first for the head markets — flow capture, zero inventory risk. The AMM as your cold-start machine for the long tail, with seed accounted for as paid marketing at a measurable cost. And a true AMM profit center in exactly one niche: high-churn, contested sports and news.

How the money flows

Your pool holds both YES and NO inventory, and every trade executes against it. Noise traders buy and sell back — price-neutral, they just leave fees behind. That is your revenue. Informed traders are the mirror image: they buy the side that wins — taking that side's inventory out of the pool, which leaves you holding the loser — and hold it to resolution; that inventory going to zero is your loss.

The life of a $1,000 market

You seed a coin-flip market with $1,000 at a 1% fee.

Noise traders arrive. Someone buys $100 of YES and sells it back two days later. On the buy, the nominal fee is $1 — but the contract prices the fee at the market value of the side you didn't buy, then converts it to dollars: at a price of 0.5, only $0.50 actually lands. On the sell, you collect the full $1. The round trip earns you $1.50, or 0.75% per $100 traded. Revenue = 0.75 × fee × turnover — that's where the 0.75 comes from.

Informed traders arrive. They keep buying YES and push the price from 0.5 to 0.8. Under the constant product, the pool is left holding 500 YES and 2,000 NO. At resolution: 80% of the time YES wins and the pool holds $500 of YES — half the seed gone; 20% of the time the upset hits and the $2,000 of NO redeems in full — a big win. On average you lose $200, exactly 20%. That is convergence loss: markets where the favorite comes through each lose a little, and the upsets fund the book. Run the same arithmetic on a 0.30 open converging to 0.85 and you lose 53.5% — that's where chapter 2's number comes from.

"List what noise traders love, refuse what informed traders love, cap every correlated cluster, keep half your bankroll out of the pools."

01Catalog & Launch

Count is sports, money is politics — and your P&L is decided by ~10 head events. The catalog is a barbell of roughly 120 markets: a thin head that carries the P&L, and a long tail that exists to fill the board and cold-start liquidity.

Tier	Markets	What	Seed each	Fee
A — heads	8–10	Marquee sports, 2 contested politics (p̄ 0.55–0.8), 1–2 crypto monthlies	$2.5–4k	150–200 bps
B — mid	30–40	Next-tier fixtures, secondary news	$500–800	200 bps
C — tail	70–80	Board filler (median sports event ≈ $7.4k lifetime volume)	$0–150 or book-only	300 bps

Demand mechanics — manufacture the round trip. A round trip is a buy followed by a sell. It doesn't move the price and causes no convergence loss (the $200 the pool lost, on average, in the example above) — it just leaves you two fees. Everything on the demand side serves one goal: users trading back and forth instead of betting once and waiting. Five mechanics:

Open markets on the rhythm of live sports. Prices keep moving during a game, so users add and trim positions repeatedly — every adjustment is a fee.
Make the cash-out button prominent. Without it, a user buys once and waits for resolution — one fee. With an easy way to sell at any time, a one-way bet becomes a round trip that pays a fee on each side.
If your deployment includes a leaderboard, rank it by return percentage, not profit. Ranking by dollars only rewards whales; ranking by percentage gives small accounts a reason to keep trading for position.
Gate community market requests on pre-committed demand. List a requested market only once enough users have clearly said they'll trade it, so you don't open markets nobody plays.

Kill criteria

Day 0 = platform launch. Fix these checkpoints before you open, so that pulling liquidity or delisting is a rule you execute, not a debate you have:

Day 30 — activity floor. Tier A sports must do at least 8× turnover per weekly cycle. Any tail market below 1× for two cycles in a row loses its seed: withdraw it and go book-only, or delist the market entirely — it isn't earning fees, it's just locking seed.
Day 60 — flow balance. If more than 70% of a category's flow runs in one direction — nearly every fill in the category is buying the same side — pull the AMM liquidity from every market in that category (the order book stays — users can still trade). One-sided flow means you're not collecting round-trip fees; you're taking the other side of informed bets — and at that point, more volume just means losing faster.
Day 90 — go / no-go. At a 150–200 bps fee, sports must come out profitable whichever side wins. Sports is the structurally safest category; if even sports can't clear the bar, the problem isn't your parameters — the audience isn't there, and more capital only scales the loss.

02Capital & Market-Making

"Opening price is the most-resourced function. A 0.30 open resolving toward 0.85 burns 53.5% of seed."

Routing. Send each market to the venue that fits it. Contested retail markets (p̄ ≤ 0.8) and the whole long tail run on the AMM, where always-on liquidity matters most. Anything decisive (p̄ ≥ 0.9) or insider-prone goes book-only, so informed flow hits other traders' quotes instead of your pool. Head markets run both: a thin AMM keeps the market tradeable at all times, and a quoting bot inside its spread captures most of the flow without inventory risk.
Seeding. Size the seed at 0.5–1× expected daily volume (estimate it from comparable live markets; your first market has no comparable, so start at the floor), with a floor of 10–20× the typical trade size so one order can't shove the price around. Add more only after the market has turned over at least 1× per day for 3 straight days — demonstrated demand, not hoped-for demand. Never grow the pool past 1× daily volume: every extra dollar of seed is an extra dollar exposed to convergence.
Withdrawal. What the pool has already lost is sunk; pulling can't recover it. What pulling does stop is future convergence. Automate the trigger in your own monitoring tooling (the contract has no built-in auto-pull): pull when your forecast of the final price sits more than 15 points from the current price. On a $10k pool at 0.70 with news pointing to 0.95, pulling saves ~$4,800 of future loss and gives up only ~$45 of fees.
Opening price. Anchor every open to de-vigged external odds, and have your own deploy script block the launch if |p₀ − reference| is more than 5 pts apart (this check is yours to build — not a product feature). If no anchor exists, create the market with both channels open but only a minimal sentinel amount in the pool, let the order book trade for 24–48h to discover the price, then align the pool to the discovered price and add the main liquidity (the pool can't be attached later — it has to exist from creation). Always deploy a 10–20% probe tranche before the full amount, and scan the news within 15 minutes of seeding — the cheapest insurance against opening into a story you missed.
Dead time. After the market ends (the event is decided and trading stops) and before the oracle pays out, the money stays locked but earns nothing, so all hurdle math uses duration + oracle window — a 7-day market's profit has to stretch across roughly 9–12 days (the extra 2–5 locked days are a typical range; disputes extend it), cutting annualized return by about 30%. Stagger seeding across the week so redemptions come back daily instead of in one lump, and keep a 15–20% liquid buffer for the gaps.

03Risk Rules

Most operators don't bleed out slowly — they die in one market they never should have seeded. So: six rules, ranked by what actually kills you, not by how often the textbook mentions it.

Adverse selection. Screen every listing with five questions. Is the outcome generated by a process, not decided behind closed doors by fewer than 20 people? Is there no one who plausibly knows the answer more than 24h early? Does a public, continuously-updating signal exist? Is p̄ ≤ 0.9 at a 2% fee? Is projected lifetime turnover at least 2× breakeven? (This is the stricter of the two turnover gates — the adverse-selection screen demands more margin than the basic 1.5× seeding bar.) A single "no" means list it without seeding (book-only), or don't list it at all. Personnel decisions, single-company announcements and niche esports fail this screen by nature — they never get AMM seed.
Correlation. Markets that hang on the same event are one cluster: the same league in the same week, the same election, the same asset. Cap each cluster at 10% of bankroll and each category at 25%, because correlated markets win and lose together — one bad night can hit every market in the cluster at once. When you measure diversification, count clusters, not markets.
Resolution. Use the wording template on every market — source, timestamp, tiebreak and postponement clauses, fixed before listing. Only run markets where the oracle bond plus gas comes to less than 5% of seed; above that, resolution costs eat the economics. Hold a dispute reserve of 2× bonds outstanding. And propose resolution yourself at T+0, so the market settles on your wording rather than someone else's.
Bankroll. When favorites come through, every market loses a little (the 20% from the example box — fees claw part of it back, so the net is a small loss) — the occasional upset is what funds the book. So never deploy more than 50% of bankroll at once, cap any single market at 2%, and size the whole operation to survive three consecutive zero-upset months, which costs about −22% of bankroll. If that drawdown would kill you, the book is too big.
Ops. The deploy script asserts the fee is one of {100, 150, 200, 300} bps, so a typo can't ship a broken rate. Read the on-chain config back before announcing anything — trust the chain, not the deploy log. And never hike fees mid-life. Take a $10k market going 0.5→0.9 at 40× turnover and raise the fee by 100 bps for the final quarter of its life: you collect about $14 more from informed flow — and $750 more from noise traders. The people you want to push away barely pay; the people you want to keep pay 53 times as much.
Abuse. If you run a points or incentive program: cap rewards at 50% of the net fees an address actually paid in the epoch, so nobody can farm out more in incentives than they hand you in fees. And exclude incentive-driven volume from the turnover stats you seed against — wash flow is statistically identical to profitable noise, and seeding against it means seeding against a mirage.

04The Business Model

"Points and token value aren't revenue — they're deferred dilution buying turnover."

Revenue lines, ranked. Four places the money comes from, in order of quality:

① Book taker fees. Charged on every taker fill, with zero inventory risk — this is the line that scales.
② AMM LP margin. Real only in the contested, high-churn band; everywhere else convergence loss eats it.
③ Sponsorships. Real money when it lands, but lumpy — don't budget around it.
④ Points and token value — not revenue. They buy today's turnover and pay for it later in dilution.

Seeding KPIs. Five numbers gate the AMM seed — listing on the frontend is a separate decision:

Projected lifetime turnover ≥ 1.5× breakeven. The basic seeding bar: a market that only just covers its own losses doesn't get seeded. Markets that trip the adverse-selection screen face the stricter 2× gate instead.
Fee capture ≥ 0.6·f·T̂ (T̂ = expected turnover). Realized fees must reach at least 0.6 of the theoretical fee rate times expected turnover. Below that, the volume isn't real or the fee isn't sticking.
Annualized ROI ≥ 30%, including the settlement gap. Count the locked oracle-window days in the denominator — a market that only clears the bar by ignoring dead time doesn't actually clear it.
Seed / daily volume ≤ 1×. If the pool is bigger than a day of volume, you've overseeded: the excess sits exposed to convergence without earning.
Category loss ratio < 0.6. Convergence loss divided by fee revenue, per category. Above that line, the category pays informed traders faster than it collects from noise.

Year-one calls. Three decisions, made up front:

Order book first, AMM as feeder. At the industry's observed median turnover (the 10–30× band) the AMM barely breaks even outside contested sports; inventory-free flow capture is what scales.
2% AMM fee, flat at creation. A 1% AMM fee needs more turnover than most markets ever produce; 2% covers the markets you'll actually run.
Sports for count, politics for money — skip crypto price targets in year one. Crypto targets break even on average at best, and every crypto market hangs off the same chart, so listing more of them adds no diversification.

05Curation & Attention

The fee is capped at 500 bps and the loss is fixed by the terminal price. In the whole revenue chain — revenue = 0.75 × fee × turnover (the example box in the doctrine shows where 0.75 comes from), turnover comes from noise traders, noise traders come from attention — attention is the only variable with no ceiling. That makes curation your first lever.

The hit formula: timeliness × controversy × celebrity / IP × screenshotability. Publicly reported cases keep validating the same four ingredients:

Listing speed is a product. When a CEO's kiss-cam moment went viral, a market was live within hours; $7M+ traded on whether he'd step down, and the answer landed two days later. The window on a breaking story is measured in hours, not days.
Attention doesn't care about looking good. A market on Zelenskyy's suit did $160M+ in volume; the ruling went against what 40+ outlets had reported, and the dispute itself trended — a disputed ruling is also traffic. Accuracy and heat are two different things.
Slice density is an attention product. One Super Bowl did $1B in volume, and four of the seven biggest markets had nothing to do with the result — halftime-show and ad prop markets. Cutting one mass event into dozens of tradeable slices beats listing dozens of cold events.

Curation means fewer, sharper listings — not more. A public analysis covering 290,000 markets found that 63% of markets that live under a day end with zero liquidity. So the rules are fixed in advance:

Two of the four ingredients, or no AMM seed. A topic that scores zero gets a book-only listing at most — let other traders' quotes take the risk.
Event bundles beat market count. Around one certain mass event, pre-build the main market plus a pack of prop slices. Ship new markets on an hourly clock.
The odds number is a distribution asset. A clean probability gets screenshotted and cited on its own ("according to ..."). Feeding it to media and KOLs in exchange for the citation is the cheapest distribution both incumbents have proven.
Centralize curation, collect suggestions through a form. An editorial desk decides — that buys speed and taste. The suggestion box is a cheap demand radar.
Write the ethics line before you need it. Disaster and hostage topics that get listed and then forced down by public outrage cost far more than a rule written in advance.

Curate first, then screen. This chapter picks what you want to list; the five questions in chapter 3 block what you must not seed. The order matters, because the more controversial a topic, the more likely someone knows the answer early — the attention formula and insider risk share a source. You need both gates.

先想清楚钱从哪来。手续费赚的是散户往返交易的钱，仅此而已——知情交易者（聪明钱）按机制几乎不缴费（原因见下面的例子），库存亏损却全部由他们造成。结论其实很直接：头部市场订单簿优先，吃的是流量，不背库存风险；AMM 用来给长尾冷启动，初始流动性当付费获客记账，成本算得清楚。而 AMM 自己能赚钱的市场，说到底只有一种：高换手、胜负胶着的体育和新闻。

钱是怎么流的

你的池子同时握着 YES 和 NO 两边的库存，谁来交易都是跟池子成交。噪声交易者买了又卖，价格没怎么动，手续费却留下了——这就是你的收入。知情交易者正好是镜像：只买会赢的那边——把会赢那边的库存从池子里买走，剩给你的自然是输的那边——一直拿到判定，那部分库存归零，就是你的亏损。

一个 $1,000 市场的一生

你往一个五五开的市场注入 $1,000，费率 1%。

噪声交易者来了。有人花 $100 买 YES，过两天又卖掉。买入那一笔，名义手续费 $1——但合约的计费规则是按"你没买的那一边"的市价折算后再换成美元入账：价格 0.5 时，实际入账只有 $0.50。卖出那一笔收全额 $1。一来一回你赚 $1.50，相当于每 $100 成交收 0.75%。收入 = 0.75 × 费率 × 换手，0.75 就是这么来的。

知情交易者来了。他们一路买 YES，把价格从 0.5 推到 0.8。恒定乘积之下，池子里的 YES 只剩 500 份、NO 堆到 2,000 份。到了判定的时候：八成概率是 YES 赢，池子只剩下值 $500 的 YES——亏掉一半；两成概率爆冷，池子里 $2,000 的 NO 全额兑换——大赚一笔。平均下来要亏 $200，正好是 20%。这就是收敛损失：热门兑现的市场个个小亏，利润靠爆冷补。同一套算法：0.30 开盘、收敛到 0.85，亏 53.5%——第 2 章那个数字就是这么算的。

"噪声交易者爱玩的市场放心上，知情者占优势的市场再热闹也要拒掉；每个相关性 cluster 都设上限，总资金永远留一半在池外。"

01市场目录与冷启动

走量靠体育，走钱靠政治，但真正决定盈亏的，其实是那 ~10 个头部事件。所以目录要摆成杠铃：约 120 个市场里，薄薄一层头部扛起全部盈亏，长尾负责把版面填满、把流动性带起来。

层级	数量	内容	单市场初始流动性	费率
A — 头部	8–10	当周焦点赛事、2 个胶着政治市场（p̄ 0.55–0.8）、1–2 个加密月度市场	$2.5–4k	150–200 bps
B — 中部	30–40	次级赛事、二线新闻	$500–800	200 bps
C — 长尾	70–80	版面填充（体育长尾市场整个生命周期成交中位数仅 ≈ $7.4k）	$0–150 或纯订单簿	300 bps

需求侧：制造往返交易。往返交易就是买入之后再卖出。它不推动价格、不产生收敛损失（就是上面例子里池子平均亏掉的那 $200），留给你的是买卖各一笔手续费。需求侧的所有设计只为一件事：让用户来回交易，而不是押一注等开奖。五个做法：

跟着体育直播的节奏开市场。比赛进行中价格一直在变，用户会反复加仓、减仓，每一次操作都给你留下一笔手续费。
把提前平仓（cash-out）按钮放在显眼位置。用户本来买完只能等判定，全程只缴一次费；有了随时卖出的入口，一次单程下注就变成买、卖各缴一次费的往返交易。
如果你的部署带排行榜，就按收益率排名，别按盈利金额。按金额排只有大户能上榜；按收益率排，小资金的用户也有动力反复交易去刷名次。
社区请求上线的市场，先看到预承诺的需求再接。比如有足够多的用户明确说会来交易，再上线，避免开出来没人玩的市场。

关停标准

以平台开业日为第 0 天。三个检查点开业前就定死——撤出流动性还是彻底下架，到点照规则执行，不靠临场争论：

第 30 天：活跃度下限。A 层体育每周换手必须 ≥ 8×。长尾市场要是连续两个周期都不到 1×，就把 AMM 流动性撤出来——转成纯订单簿，或者干脆彻底下架。毕竟它赚不到手续费，只是把你的初始流动性白白锁在里面。
第 60 天：流向体检。哪个类别的单向流占比超过 70%——也就是这个类别里几乎所有成交都在买同一边——把该类别下每一个市场的 AMM 流动性都撤出（订单簿还在，用户照样能交易）。流量一边倒，说明你收不到往返手续费，只是在给知情下注当对手盘——到这一步，流量越大，亏得越快。
第 90 天：去留判断。在 150–200 bps 的费率档下，无论哪边赢，体育市场都必须是赚的。体育是结构上最安全的类别，连它都打不平的话，问题就不在参数上——说明根本没有受众，再加钱也只是把亏损放大。

02资金与做市

"开盘定价最值得花资源。0.30 开盘、最后收敛到 0.85，烧掉 53.5% 的初始流动性。"

分流。每个市场都要送去适合它的场地。胜负胶着的散户市场（p̄ ≤ 0.8）和整条长尾走 AMM——这类市场最需要随时有价。大局已定（p̄ ≥ 0.9）或内幕风险高的只上订单簿，让知情资金去吃别的交易者的挂单，而不是你的池子。头部市场两边都开：薄薄一层 AMM 保证随时可交易，报价 bot 在价差里挂单，吃下大部分流量又不背库存风险。
注入流动性。初始流动性按预期日成交量的 0.5–1 倍配（预期日成交量拿同类已开市场的实际成交当参考；第一个市场没有参考，就从下限起步），下限是单笔典型交易的 10–20 倍——免得一笔单子就把价格打飞。追加要等市场连续 3 天日换手 ≥ 1×，也就是看到真实需求再加，而不是凭期待加。流动性池最多加到 1× 日成交量：多注一块钱，就多一块钱暴露在收敛亏损里。
撤出流动性。已经亏掉的部分是沉没成本，撤出救不回来；它真正能救的，是未来的收敛。在你自己的部署 / 监控脚本里做成自动触发（合约没有内置的自动撤出）：预测终价和现价差超过 15 点就撤。$10k 的池子价格在 0.70、新闻已经指向 0.95，这时撤出省下 ~$4,800 的未来亏损，放弃的只有 ~$45 的手续费。
开盘定价。每个开盘价都锚定去水后的外部赔率，在你自己的部署脚本里加拦截：|p₀ − 参考| 差超过 5 点就不许上线（这道检查要你自己建，不是产品功能）。找不到外部锚，就在创建市场时两个通道都开、但池子里只放最小额的哨兵流动性，让订单簿跑 24–48h 把价格走出来，再把池子价格对齐到订单簿价、补足主流动性（池子没法事后追加创建，必须开盘就带上）。全量注入前永远先放 10–20% 的试探性流动性试水；注入前 15 分钟内把新闻扫一遍——这是防止开盘踩进突发消息最便宜的保险。
死区时间。市场到期（事件结果出来、交易停止）后、预言机赔付前，钱锁着却一分不赚，所以回报门槛一律按实际周期 = 市场时长 + 预言机判定窗口算——7 天市场赚的钱要摊到约 9–12 天上（多锁的 2–5 天是经验值，视预言机有无争议而定），年化少三成。流动性注入在一周内错开，让兑换的钱每天回笼，而不是挤在同一天；手里再留 15–20% 的流动缓冲。

03风控铁律

大多数运营方不是慢慢亏死的，是死在一个本就不该添加流动性的市场上。所以这六条按真实杀伤力排，不按教科书提到的次数排。

逆向选择。每个市场上架前过五个问题。结果是公开流程跑出来的，还是 < 20 个人关起门定的？有没有人能提前 24h 知道答案？有没有公开、持续更新的信号？2% 费率下 p̄ ≤ 0.9 吗？预期换手（生命周期口径）够不够 2× 保本线？（这是两道换手门槛里更严的一道——内幕风险筛查要求的安全边际，高于添加流动性的 1.5× 基本线。）只要有一条是"否"，就只上架不添加流动性（纯订单簿），或者干脆不上架。人事任命、单公司公告、小众电竞，天生过不了这套筛——永远不给 AMM 初始流动性。
相关性。结果系在同一个事件上的市场算一个 cluster：同联赛同一周、同一场选举、同一个标的。单 cluster 不超过总资金 10%，单类别不超过 25%——因为相关的市场会一起赢、一起输，一个坏夜晚就能把整个 cluster 同时打穿。所以算分散度的时候要数 cluster，而不是数市场。
判定。每个市场都要用措辞模板：数据源、时间戳、平局条款、延期条款，全部在上架前写死。预言机保证金加 gas 超过初始流动性 5% 的市场不开——判定成本会把账吃掉。争议准备金备到在外保证金的 2 倍。自己注入流动性的市场，T+0 第一个把判定提议交上去，让市场按你写的条款落地，而不是别人写的。
资金。热门如期兑现的市场个个小亏（例子框里那 20%——手续费抵掉一部分后，净下来是小亏），利润靠偶尔的爆冷补回来。所以：同一时间最多 50% 投入流动性池；单市场不超过总资金 2%；总规模按"连吃三个月零爆冷（约亏掉总资金的 22%）还活着"来配——这个回撤要是会要你的命，说明摊子铺大了。
操作。部署脚本里断言费率 ∈ {100, 150, 200, 300} bps，手滑打错也上不了线。公告之前把链上配置回读一遍——信链，别信部署日志。费率绝不中途上调。一个 0.5→0.9、换手 40 倍的 $10k 市场，最后四分之一段把费率加 100 个基点：从知情资金身上多收约 $14，却向噪声交易者多收了 $750——你想赶走的人几乎没付，你想留住的人付了 53 倍。
刷量。如果你跑积分或激励计划：单地址单周期能拿的奖励，不超过它净缴手续费的 50%，谁也别想薅走的比交上来的多。激励刷出来的量，一律不进流动性决策用的换手统计——对敲量在数据上和真赚钱的噪声流长得一模一样，照着它添加流动性，等于照着海市蜃楼配资金。

04商业模型

"积分和代币价值不是收入，是用递延稀释买换手。"

收入线排序。钱从四个地方来，按质量排：

① 订单簿 taker 费。每笔吃单都收，零库存风险——唯一能上规模的一条。
② AMM 做市利润。只在胜负胶着、高换手那一段是真的；其他区间会被收敛损失吃掉。
③ 赞助。来了确实是真钱，但有一搭没一搭的，别拿它当预算做。
④ 积分 / 代币价值——不是收入。本质是先用代币买来今天的换手，将来再拿稀释去还。

添加流动性 KPI。这五个数字决定的是给不给 AMM 初始流动性——上架到前端是另一个决定：

预期换手（生命周期口径）≥ 1.5× 保本线。这是添加流动性的基本门槛：刚好够保本的市场不值得投入。触发逆向选择筛查的市场，适用的是更严的 2× 那道门。
费用捕获 ≥ 0.6·f·T̂（T̂ = 预期换手）。实际收到的手续费，至少要达到理论值（费率 × 预期换手）的 0.6。低于这个数，要么量是假的，要么费没收上来。
含判定窗口的年化 ≥ 30%。锁在预言机窗口里的天数也要算进分母——靠忽略死区时间才过线的市场，实际上没过线。
初始流动性 / 日成交 ≤ 1×。流动性池比一天的成交量还大，就是注多了：多出来的钱光挨收敛、不产出。
类别亏损率 < 0.6。按类别算收敛损失 ÷ 手续费收入。超过这条线，这个类别给知情交易者送钱的速度，已经快过从噪声那里收钱的速度。

第一年的三个决定。三件事，开局就要定下来：

订单簿优先，AMM 做导流。换手在行业实测的中位档（10–30 倍那一档）时，AMM 在胶着体育之外基本只能打平；能上规模的是零库存的流量生意。
AMM 固定 2%，开盘就定死。1% 的 AMM 费率需要的换手量，大多数市场一辈子都跑不出来；2% 才盖得住你实际会开的市场。
体育走量、政治走钱，第一年别碰币价市场。币价市场期望顶多打平，而且全都挂在同一条行情线上，开再多也是一个 cluster，分散不了风险。

05选题与注意力

费率有 500 bps 的天花板，亏损被终价锁死。整条收入链路——收入 = 0.75 × 费率 × 换手（0.75 的来历见总纲的例子框），换手来自噪声交易者，噪声交易者来自注意力——只有注意力这一环没有上限。所以选题是你的第一杠杆。

爆款公式：时效 × 争议 × 名人 / IP × 可截图。公开报道的案例反复验证这四个要素：

开盘速度本身是产品。某 CEO 的亲吻镜头上热搜后，几小时内就有市场开出来，$7M+ 押他会不会下台，两天后应验。热点的窗口是按小时算的，不是按天。
注意力不挑好坏。Zelenskyy 西装市场做了 $160M+ 的量；判定结果和 40 多家媒体的报道相反，吵上了热搜——裁决争议本身也是流量。准确率和热度是两回事。
切片密度本身是注意力产品。一届超级碗做出 $1B 成交，前七大市场里有四个和比赛胜负无关——中场表演、广告这类 prop 市场。把一个全民事件切成几十个可交易切片，胜过开几十个冷门事件。

Curation 是少而准，不是多开。一份覆盖 29 万个市场的公开分析显示：存续不到 1 天的市场，63% 以零流动性收场。所以规则提前定死：

四要素至少命中两项，才考虑添加 AMM 流动性。零命中的题材最多纯订单簿上架——让别的交易者的挂单去担风险。
事件套餐优于铺量。围绕一个确定会发生的大事件，预制主盘加一组 prop 切片，比铺一堆冷市场有效；上新速度按小时算。
赔率数字是分发资产。一个干净的概率数字，自己就会被截图、被引用（"according to ..."）。免费喂给媒体和 KOL 换一个引用位，是两家头部平台共同验证过的最便宜分发。
选题权中心化，需求走表单。编辑部拍板，换来的是速度和品味；用户建议入口当低成本的需求雷达。
伦理红线事前写死。灾难、人质类题材开了再被舆论倒逼下架，代价远高于一条事前规则。

先选题，后筛查。这一章选出"想上的"，第 3 章的五问挡掉"不能上的"。顺序不能反——越有争议的题材，越可能有人提前知道答案，注意力公式和内幕风险常常同源。两道工序缺一不可。

Market state	Visible on frontend	AMM seeded	Tradeable
Listed (book-only)	✓	✗	✓ order book
Seeded	✓	✓	✓ AMM + order book
Seed pulled (one market)	✓	✗	✓ order book only
Category-wide pull	✓	✗	✓ order book only
Delisted entirely	✗	✗	✗
Market ended	✓	✓ (locked until resolution)	✗
Resolved	✓	✗ (redeemed)	✗ (positions redeem instead)

市场状态	前端可见	AMM 有流动性	可交易
上架（纯订单簿）	✓	✗	✓ 订单簿
已添加流动性	✓	✓	✓ AMM + 订单簿
撤出流动性（单市场）	✓	✗	✓ 仅订单簿
整类撤出流动性	✓	✗	✓ 仅订单簿
彻底下架	✗	✗	✗
市场到期	✓	✓（锁定至判定）	✗
判定完成	✓	✗（已兑换取回）	✗（持仓走兑换）

Whether a prediction market makes money is decided before it opens.

预测市场赚不赚钱，开盘之前就定了。

Start here

The Operator's Doctrine

运营总纲

All lessons

The APMM Calculator

APMM 计算器

Catalog & Launch

市场目录与冷启动

Capital & Market-Making

资金与做市

Risk Rules

风控铁律

The Business Model

商业模型

Curation & Attention

选题与注意力

The Operator's Doctrine

运营总纲

01Catalog & Launch

02Capital & Market-Making

03Risk Rules

04The Business Model

05Curation & Attention

01市场目录与冷启动

02资金与做市

03风控铁律

04商业模型

05选题与注意力

The APMM Calculator

APMM 计算器

Parameters

Results

Path to profit

Cumulative P&L over the trading window

Net P&L vs terminal price

Glossary

术语表

AMM (Automated Market Maker)AMM（自动做市商）

Adverse selection逆向选择

Annualized ROI年化回报

Bankroll总资金

Barbell catalog杠铃式目录

Basis points (bps)基点（bps）

Breakeven turnover保本换手率

Cash-out提前平仓

Cluster (correlation)相关性 cluster

Convergence loss收敛损失

Dead time死区时间

De-vigged odds去水赔率

Fee capture费用捕获

Fee switch费率切换

Head market头部市场

Hurdle rate门槛收益率

Informed flow知情流（聪明钱）

Kill criteria关停标准

Long tail长尾

Noise flow噪声流（噪声交易者）

Opening price (p₀)开盘价（p₀）

Oracle预言机

Order book订单簿

Probe tranche试探性流动性（probe tranche）

Resolution判定

Round trip往返交易

Seed liquidity初始流动性（seed liquidity）

Terminal price (p̄)终价（p̄）

Turnover换手率

Wash trading刷量

Withdrawal trigger撤出流动性触发

AMM (Automated Market Maker)

Adverse selection

Annualized ROI

Bankroll

Barbell catalog

Basis points (bps)

Breakeven turnover

Cash-out

Cluster (correlation)

Convergence loss

Dead time

De-vigged odds

Fee capture

Fee switch

Head market

Hurdle rate

Informed flow

Kill criteria

Long tail

Noise flow

Opening price (p₀)

Oracle

Order book

Probe tranche

Resolution

Round trip

Seed liquidity

Terminal price (p̄)

Turnover

Wash trading

Withdrawal trigger