# 28 tools under the hood of bot that made $1M on Polymarket
**作者**: cvxv666
**日期**: 2026-04-21T18:15:42.000Z
**来源**: [https://x.com/antpalkin/status/2046654122892403188](https://x.com/antpalkin/status/2046654122892403188)
---
28 repositories. Six layers. One stack. No gaps.
How it works, why it works, and how to make your first $1,000 on Polymarket with same algo - it’s all in here.
Everything you need to build a profitable trading bot on Polymarket - from the brain to the backtest to the production execution layer.
Our patient:
Wallet link: polymarket.com/@coinman2?r=antopotoshka#RLOep5t
This article is not the story of one wallet. It is a map of the stack. Each tool below closes a specific failure mode that bleeds capital from most prediction-market operators: missing data, missing validation, missing discipline, missing speed.
## Before you read:
→ Bookmark this. You'll need it when you actually start building.
→ If you want these exact strategies working for you tonight just copy coinman2 trades with this TG bot: kreo.app/@cvxv666
→ follow me for more alpha @antpalkin
→ You'll definitely need a Polymarket account: polymarket.com
# Coinman2 - main hero.
$1,003,450 PnL. 3,062 predictions. One wallet.
The wallet is "0x55be7aa03ecfbe37aa5460db791205f7ac9ddca3" public handle @coinman2 on Polymarket. Joined November 2024. Biggest single win on record: $113.8K.
When traders first saw the growth trajectory on-chain, the initial reaction from crypto Twitter was simple: fake. A seven-figure cumulative PnL from a retail-profile wallet belongs in a Telegram scam group, not in a verifiable on-chain dashboard.
But the blockchain doesn't lie. Every trade is public. Every settlement is verifiable.
Developers reverse-engineered the approach within weeks. Then they asked Claude to rebuild it from scratch.
This piece is about the full tool stack that makes this possible - and what each layer actually does.
# What Polymarket is and where the edge lives.
Polymarket is a prediction market. Users trade on outcomes: will Bitcoin be higher in 15 minutes? Will the Fed raise rates? Each contract resolves at $1.00 (correct) or $0.00 (wrong). A contract priced at $0.73 means the market believes there's a 73% chance the "Yes" outcome happens.
The platform's weekly volume exceeded $2 billion in early 2026.
The key category for automated trading is short-duration crypto contracts - 5-minute and 15-minute BTC and ETH up/down questions. They resolve fast, provide immediate feedback, and have a structural vulnerability.
Polymarket updates its prices slower than the underlying asset moves on Binance. In 2024, that lag averaged 12 seconds. By Q1 2026, competition had compressed it to 2.7 seconds.
2.7 seconds is an eternity for a machine.
That gap - between what Binance knows and what Polymarket still shows is where every strategy in this article lives.
# The mechanism, step by step.
A 15-minute BTC contract opens at 50/50. Ten minutes in, Bitcoin drops 0.6% on Binance in 30 seconds. The "real" probability that BTC will be lower at expiry just shifted to roughly 78%. Polymarket still shows 54/46.
That's a 24-point edge on a binary contract. It's not a prediction. The outcome, in a probabilistic sense, has already happened.
A bot monitoring Binance's WebSocket feed with under 50ms latency sees this immediately. It calculates the discrepancy, sizes a position using Kelly Criterion, and executes via Polymarket's CLOB API. Two seconds later, the market corrects. The position closes profitable.
Repeat 200–500 times per day.
That's the coinman2 result. Not magic. Industrial-scale exploitation of a gap that still exists today.
# Why the gap exists at all.
Polymarket is a decentralized prediction market. Prices only update when traders actively post orders - there's no dedicated market-making desk refreshing quotes in real time. In the seconds after a Binance move, the "smart" side of the contract has few sellers willing to trade at stale odds.
The gap has narrowed from 12 seconds to 2.7 seconds as more bots entered. It will continue to narrow. But it hasn't closed. And closing it entirely would require the kind of real-time automated market makers that are, themselves, already the arbitrage bots.
It's an arms race with no obvious end state.
# Polymarket API infrastructure.
Before any AI, you need the raw infrastructure. Polymarket exposes four surfaces:
The official Python client, "py-clob-client" (pip install py-clob-client), wraps all of this. Three lines to fetch an order book. Five to place a signed limit order on Polygon (chain ID 137, USDC settlement).
Key repos to start with:
- Polymarket/agents (https://github.com/Polymarket/agents) - Official AI agent framework. Gamma API, CLOB API, and LangChain already wired together. Built-in Chroma vector database for ingesting news sources.
- polyterm (github.com/NYTEMODEONLY/polyterm) - Terminal dashboard: whale tracker, arbitrage scanner, insider detection, cross-platform comparison vs Kalshi, market risk grading A–F.
# The stack - layer by layer.
## LAYER 1 - BRAIN
AI reasoning. The coinman2 bot ran on Anthropic's Claude. In March 2026, a controlled experiment ran Claude against the OpenClaw framework - same starting capital ($1,000), same market conditions, 48 hours.
Claude: +1,322% return. OpenClaw: fully liquidated.
Claude reasoning trace - from market context to Kelly-sized order, every decision auditable after the fact.
Researchers traced the gap to one thing: risk management quality. Claude's generated code included more conservative default parameters, more defensive edge cases, and cleaner error handling. OpenClaw overlevered into a losing sequence and couldn't stop.
Claude (Anthropic) - Primary strategist. Reasons about market questions, estimates probability vs current price, identifies edge size.
Qwen3-Coder (github.com/QwenLM/Qwen3-Coder) - Open source coding LLM. Watches live performance, detects crowded strategies, rewrites modules autonomously.
G0DM0D3 (github.com/elder-plinius/G0DM0D3) - Uncensored AI interface. Engages with uncomfortable market theses without refusals.
Claude Squad (github.com/smtg-ai/claude-squad) - Runs multiple Claude instances in parallel. One watches politics, one crypto, one sports - all feeding the same engine.
## LAYER 2 - ORCHESTRATION
Making agents do things. A reasoning engine with no execution layer is just an opinion generator.
Bull vs Bear debate resolved by Risk Manager veto - the framework that keeps confident-but-wrong from becoming confident-and-bankrupt.
Agency Agents (github.com/msitarzewski/agency-agents) - Role-based debate: Bull Agent vs Bear Agent vs Risk Manager veto. Consensus determines trade and size.
ClaudeAgent OneClick (github.com/cvxv666/ClaudeAgentOneClick) - One click Claude agent deploy. Persistent 24/7 market watcher in minutes.
MiroThinker (github.com/MiroMindAI/MiroThinker) - Mandatory chain-of-thought layer. Bot must justify every position before entry.
Superpowers (github.com/obra/superpowers) - Extends agents with web access, file ops, arbitrary API calls. Fresh data every decision cycle.
TradingAgents (github.com/TauricResearch/TradingAgents) - Multi-agent framework: fundamental analyst + technical analyst + sentiment analyst → aggregated signal.
## LAYER 3 - DATA & MARKET SIGNALS
The eyes. The bot is only as good as what it can see. This layer used to be four tools. Adding macro data, indicator math, and a real-time charting engine brings it closer to a Bloomberg terminal than a script.
The edge visualized: Binance-implied probability (blue) vs Polymarket shown (red). The gap between lines is the entire business.
OpenBB (github.com/OpenBB-finance/OpenBB) - Open source Bloomberg. 100+ data sources unified: stocks, macro, crypto, options flow, news. The backbone.
Dexter (github.com/virattt/dexter) - Autonomous deep research. SEC filings, earnings transcripts, analyst reports - two hours of analyst work in seconds.
MCP Server (github.com/financial-datasets/mcp-server) - Financial datasets via MCP protocol. Typed, validated data straight into Claude's context window.
Crucix (github.com/calesthio/Crucix) - On-chain aggregator. Whale wallet movements on Polygon can front-run the visible order book by minutes.
fredapi (github.com/mortada/fredapi) - Every macroeconomic dataset the Federal Reserve publishes, free, behind one Python wrapper. CPI, unemployment, yield curves - pipe directly into Claude's context to anchor any macro-related market.
Binance Collector (github.com/txbabaxyz/mlmodelpoly) - Predicts market direction and computes fair value for short-duration BTC/ETH contracts. The price-discovery half of every latency-arbitrage trade.
Polymarket Assistant Tool (github.com/FiatFiorino/polymarket-assistant-tool) - Indicator engine that surfaces directional bias on live markets. Translates raw orderbook + price action into a usable signal.
lightweight-charts (github.com/tradingview/lightweight-charts) - TradingView's own charting library. 14k stars, 45KB, free. The fastest path to a production-grade dashboard for monitoring positions and signals in real time.
## LAYER 4 - MARKET INTELLIGENCE
What others have already built. You don't have to build everything from scratch. An entire ecosystem of Polymarket-specific intelligence and pre-built bots already exists.
Whale alerts, leaderboards, signal confluence - the institutional-grade view retail never used to have.
Polyscope (thepolyscope.com) - Scans 2,000+ markets. Whale alerts ($5k/$10k/$20k), probability shifts, volume spikes - all to Telegram in real time.
Polywhaler (polywhaler.com) - $10k+ whale trade tracker. Real-time alerts, insider detection, AI-generated signals.
WHALES tracker (Apify) - Cross-references Gamma + Data + CLOB. Outputs Smart Money Consensus, Health Score, Conviction Score as structured JSON.
HyperBuildX bot (github.com/HyperBuildX/Polymarket-Trading-Bot) - Rust-based, sub-100ms latency. AI ranks copy-trade signals when multiple whales pile into the same market.
polymarketanalytics.com - Open trader analytics. Top wallet P&L, biggest wins, Kalshi cross-comparison.
polyrec (github.com/txbabaxyz/polyrec) - Real-time terminal dashboard combining Chainlink oracle, Binance feed, and full orderbook depth. 70+ indicators, automatic CSV logging, built-in strategy backtester. The closest thing to an institutional cockpit you can get free.
Polymarket-Trading-Bot (github.com/dylanpersonguy/Polymarket-Trading-Bot) | 53k lines of TypeScript. Seven prebuilt strategies: arbitrage, momentum, market making, AI forecast, whale copy-trade, convergence, and more. A complete framework - pick a strategy, plug in keys, run. |
## LAYER 5 - BACKTEST & SIMULATION
Prove it before you run it. This is the layer most retail bots skip - and the reason most retail bots blow up.
Reasoning models can sound convincing about strategies that have never made money. Whales can be followed into trades that already moved. A signal that worked last month may have been arbitraged out by yesterday. The only defense is running every idea against historical data and simulated execution before a single dollar of real capital touches the CLOB.
Every strategy earns its way to production by surviving this report first: 17 months of live record, fees and slippage included, drawdowns visible.
prediction-market-backtesting (github.com/evan-kolberg/prediction-market-backtesting) - Backtests trading strategies against real historical Polymarket and Kalshi data. The first sanity check on any new idea before it sees production.
polybot (github.com/ent0n29/polybot) - Full execution and market-data infrastructure with paper trading. Kafka, ClickHouse, Grafana - the analytics pipeline an institutional desk would build. Reverse-engineered from production strategies.
## The complete signal flow.
World event to execution: under 10 seconds in the ideal world.
# Human traders vs bots - the data.
The performance gap between human traders and automated bots using comparable latency arbitrage approaches is documented. Bots generated approximately $206,000 during a tracked period. Humans using the same logic generated roughly $100,000.
2× gap. Same market. Same strategy. Same time window.
The gap isn't better forecasting. It's execution. Humans make four systematic errors that bots don't:
- Late entries. By the time a human identifies the Polymarket lag, confirms the Binance move, and places the order manually, the window has often already closed.
- Inconsistent sizing. Humans oversize when confident, undersize when uncertain — exactly the inverse of Kelly math. Emotional sizing destroys expected value across thousands of trades.
- Fatigue. A human monitoring 15-minute BTC contracts through an 8-hour session degrades. A bot running for 72 hours makes the same decision at hour 72 that it made at hour 1.
- Drawdown psychology. After a losing sequence, humans either abandon a working strategy or double down. Both destroy capital. A bot with a hard kill switch does neither.
# Why now.
The bots already running have a compounding advantage. The edge exists today. The window is narrowing. The best time to understand this stack was six months ago.
The second best time is right now.
If you read this far:
-> SAVE THIS POST or you’ll lose every important link.
-> coinman2 bot profile on Polymarket: wallet link
-> If it seems too complicated, copying the trades of similar bots is also a profitable strategy. That’s exactly what I do through a TG bot: kreo.app/@cvxv666
-> remember: competition is way weaker than it looks. 99.9% of people will scream "too hard" and call it cap. Only the 0.01% actually try… and eat.
## 相关链接
- [cvxv666](https://x.com/antpalkin)
- [@antpalkin](https://x.com/antpalkin)
- [kreo.app/@cvxv666](https://kreo.app/@cvxv666)
- [@antpalkin](https://x.com/@antpalkin)
- [https://github.com/Polymarket/agents](https://github.com/Polymarket/agents)
- [github.com/NYTEMODEONLY/polyterm](https://github.com/NYTEMODEONLY/polyterm)
- [github.com/QwenLM/Qwen3-Coder](https://github.com/QwenLM/Qwen3-Coder)
- [github.com/elder-plinius/G0DM0D3](https://github.com/elder-plinius/G0DM0D3)
- [github.com/smtg-ai/claude-squad](https://github.com/smtg-ai/claude-squad)
- [github.com/msitarzewski/agency-agents](https://github.com/msitarzewski/agency-agents)
- [github.com/cvxv666/ClaudeAgentOneClick](https://github.com/cvxv666/ClaudeAgentOneClick)
- [github.com/MiroMindAI/MiroThinker](https://github.com/MiroMindAI/MiroThinker)
- [github.com/obra/superpowers](https://github.com/obra/superpowers)
- [github.com/TauricResearch/TradingAgents](https://github.com/TauricResearch/TradingAgents)
- [github.com/OpenBB-finance/OpenBB](https://github.com/OpenBB-finance/OpenBB)
- [github.com/virattt/dexter](https://github.com/virattt/dexter)
- [github.com/financial-datasets/mcp-server](https://github.com/financial-datasets/mcp-server)
- [github.com/calesthio/Crucix](https://github.com/calesthio/Crucix)
- [github.com/mortada/fredapi](https://github.com/mortada/fredapi)
- [github.com/txbabaxyz/mlmodelpoly](https://github.com/txbabaxyz/mlmodelpoly)
- [github.com/FiatFiorino/polymarket-assistant-tool](https://github.com/FiatFiorino/polymarket-assistant-tool)
- [github.com/tradingview/lightweight-charts](https://github.com/tradingview/lightweight-charts)
- [thepolyscope.com](https://thepolyscope.com/)
- [polywhaler.com](https://polywhaler.com/)
- [$10k](https://x.com/search?q=%2410k&src=cashtag_click)
- [github.com/HyperBuildX/Polymarket-Trading-Bot](https://github.com/HyperBuildX/Polymarket-Trading-Bot)
- [polymarketanalytics.com](https://polymarketanalytics.com/)
- [github.com/txbabaxyz/polyrec](https://github.com/txbabaxyz/polyrec)
- [github.com/dylanpersonguy/Polymarket-Trading-Bot](https://github.com/dylanpersonguy/Polymarket-Trading-Bot)
- [github.com/evan-kolberg/prediction-market-backtesting](https://github.com/evan-kolberg/prediction-market-backtesting)
- [github.com/ent0n29/polybot](https://github.com/ent0n29/polybot)
- [kreo.app/@cvxv666](https://kreo.app/@cvxv666)
- [Upgrade to Premium](https://x.com/i/premium_sign_up)
- [Paid partnership](https://help.x.com/rules-and-policies/paid-partnerships-policy)
- [2:15 AM · Apr 22, 2026](https://x.com/antpalkin/status/2046654122892403188)
- [1M Views](https://x.com/antpalkin/status/2046654122892403188/analytics)
- [View quotes](https://x.com/antpalkin/status/2046654122892403188/quotes)
---
*导出时间: 2026/5/1 13:59:08*
---
## 中文翻译
# 揭秘百万美元收益 Polymarket 机器人的 28 个底层工具
**作者**: cvxv666
**日期**: 2026-04-21T18:15:42.000Z
**来源**: [https://x.com/antpalkin/status/2046654122892403188](https://x.com/antpalkin/status/2046654122892403188)
---
28 个代码仓库。六层架构。一套技术栈。无死角。
它如何运作、为何有效，以及如何利用同样的算法在 Polymarket 赚到你的第一个 1,000 美元——这一切都在这里。
在 Polymarket 上构建盈利交易机器人所需的一切——从“大脑”到回测，再到生产执行层。
我们的研究对象：
钱包链接：polymarket.com/@coinman2?r=antopotoshka#RLOep5t
本文讲的不是一个钱包的故事，而是一张技术栈地图。下面列出的每个工具都旨在解决一种特定的失效模式——正是这些模式让大多数预测市场运营者的资金不断流失：数据缺失、验证缺失、纪律缺失、速度缺失。
## 阅读前必读：
→ 收藏本文。当你真正开始构建时，你会用到它。
→ 如果你想让这些精准的策略今晚就为你工作，直接用这个 TG 机器人跟单 coinman2：kreo.app/@cvxv666
→ 关注我获取更多 Alpha：@antpalkin
→ 你肯定需要一个 Polymarket 账户：polymarket.com
# Coinman2 - 故事主角。
$1,003,450 盈亏（PnL）。3,062 次预测。一个钱包。
该钱包地址为 "0x55be7aa03ecfbe37aa5460db791205f7ac9ddca3"，在 Polymarket 上的公开 ID 是 @coinman2。加入时间为 2024 年 11 月。有记录以来的最大单笔盈利：$113.8K。
当交易者们在链上首次看到这条增长曲线时，加密推特（Crypto Twitter）的第一反应很简单：假的。一个零售级钱包能拥有七位数的累计盈亏，只可能出现在 Telegram 骗局群里，而不是可验证的链上仪表板上。
但区块链不会撒谎。每一笔交易都是公开的。每一次结算都是可验证的。
开发者在几周内就逆向还原了这种方法。然后他们让 Claude（AI）从零开始重建它。
本文介绍的是实现这一切的完整工具栈——以及每一层实际上做了什么。
# Polymarket 是什么，以及优势（Edge）在哪里。
Polymarket 是一个预测市场。用户对结果进行交易：比特币 15 分钟后会上涨吗？美联储会加息吗？每个合约结算价为 $1.00（正确）或 $0.00（错误）。一个价格为 $0.73 的合约意味着市场认为“是”的结果发生的概率为 73%。
2026 年初，该平台的周交易量超过了 20 亿美元。
自动交易的关键类别是短周期的加密合约——5 分钟和 15 分钟的 BTC 和 ETH 涨跌问题。它们结算快、反馈即时，且存在结构性漏洞。
Polymarket 的价格更新速度慢于标的资产在 Binance 上的移动速度。2024 年，这种滞后平均为 12 秒。到 2026 年第一季度，随着竞争加剧，它被压缩到了 2.7 秒。
2.7 秒对于机器来说是永恒。
这个差距——即 Binance 已知的信息与 Polymarket 仍显示的信息之间的差值——就是本文中所有策略的生存空间。
# 机制，逐步拆解。
一个 15 分钟的 BTC 合约开盘时赔率为 50/50。十分钟后，比特币在 30 秒内在 Binance 下跌了 0.6%。BTC 在到期时价格更低的“真实”概率刚刚转变为大约 78%。但 Polymarket 仍然显示 54/46。
这是一个二元合约上 24 个点的优势。这不是预测。从概率意义上讲，结果已经发生了。
一个通过延迟低于 50ms 的 WebSocket 监控 Binance 数据的机器人会立即看到这一点。它计算差异，使用凯利公式确定仓位，并通过 Polymarket 的 CLOB API 执行。两秒钟后，市场修正。仓位平仓获利。
每天重复 200–500 次。
这就是 coinman2 的成果。不是魔法。是对至今仍存在的漏洞的工业化规模利用。
# 为什么这个差距会存在。
Polymarket 是一个去中心化的预测市场。只有当交易者主动挂单时，价格才会更新——没有专门的做市台实时刷新报价。在 Binance 价格变动后的几秒钟内，合约的“聪明”一方很少有卖家愿意以过时的赔率交易。
随着更多机器人的进入，差距已从 12 秒缩小到 2.7 秒。它还会继续缩小。但它不会消失。而要完全消除它，需要一种实时自动做市商，而它们本身就已经是套利机器人了。
这是一场没有明显终局的军备竞赛。
# Polymarket API 基础设施。
在谈论任何 AI 之前，你需要原始的基础设施。Polymarket 提供了四个接口：
官方 Python 客户端 "py-clob-client"（通过 pip install py-clob-client 安装）封装了所有这些功能。三行代码获取订单簿。五行在 Polygon（链 ID 137，USDC 结算）上下达签名限价单。
入手的关键仓库：
- Polymarket/agents (https://github.com/Polymarket/agents) - 官方 AI 代理框架。Gamma API、CLOB API 和 LangChain 已预先连接。内置 Chroma 向量数据库用于引入新闻来源。
- polyterm (github.com/NYTEMODEONLY/polyterm) - 终端仪表板：巨鲸追踪器、套利扫描器、内幕检测、与 Kalshi 的跨平台对比、市场风险评级 A–F。
# 技术栈 - 逐层解析。
## 第 1 层 - BRAIN（大脑）
AI 推理。coinman2 机器人运行在 Anthropic 的 Claude 上。2026 年 3 月，一项对照实验让 Claude 与 OpenClaw 框架对决——相同的起始资金（$1,000），相同的市场条件，48 小时。
Claude：+1,322% 回报。OpenClaw：完全爆仓。
Claude 推理轨迹——从市场背景到凯利定单，每个决策都可事后审计。
研究人员将差距归结为一件事：风险管理质量。Claude 生成的代码包含了更保守的默认参数、更防御性的边缘情况和更整洁的错误处理。OpenClaw 则过度杠杆化了一系列亏损交易且无法停止。
Claude (Anthropic) - 首席策略师。负责思考市场问题、估算概率与当前价格的差异、识别优势规模。
Qwen3-Coder (github.com/QwenLM/Qwen3-Coder) - 开源代码大模型（LLM）。监控实时表现，检测拥挤策略，自主重写模块。
G0DM0D3 (github.com/elder-plinius/G0DM0D3) - 无审查 AI 界面。能够介入敏感的市场论题而不会遭到拒绝。
Claude Squad (github.com/smtg-ai/claude-squad) - 并行运行多个 Claude 实例。一个看政治，一个看加密货币，一个看体育——都为同一个引擎提供信息。
## 第 2 层 - ORCHESTRATION（编排）
让代理执行任务。没有执行层的推理引擎只是一个观点生成器。
由风险经理否决权解决的多空辩论——这是防止“自信但错误”变成“自信且破产”的框架。
Agency Agents (github.com/msitarzewski/agency-agents) - 基于角色的辩论：多头代理 vs 空头代理 vs 风险经理否决权。共识决定交易及规模。
ClaudeAgent OneClick (github.com/cvxv666/ClaudeAgentOneClick) - 一键部署 Claude 代理。几分钟内就能启动一个 7x24 小时的市场监控者。
MiroThinker (github.com/MiroMindAI/MiroThinker) - 强制性思维链 层。机器人必须在入场前证明每个仓位的合理性。
Superpowers (github.com/obra/superpowers) - 为代理扩展网络访问、文件操作和任意 API 调用功能。每个决策周期都能获取最新数据。
TradingAgents (github.com/TauricResearch/TradingAgents) - 多代理框架：基本面分析师 + 技术分析师 + 情绪分析师 → 聚合信号。
## 第 3 层 - DATA & MARKET SIGNALS（数据与市场信号）
眼睛。机器人只能看到它所能看到的东西。这一层过去需要四个工具。添加宏观数据、指标数学和实时图表引擎后，它更接近彭博终端，而不是简单的脚本。
可视化优势：Binance 隐含概率（蓝线）vs Polymarket 显示价格（红线）。两条线之间的差距就是整个生意。
OpenBB (github.com/OpenBB-finance/OpenBB) - 开源彭博。统一了 100 多个数据源：股票、宏观、加密货币、期权流、新闻。它是骨干。
Dexter (github.com/virattt/dexter) - 自主深度研究。SEC 文件、财报电话会议记录、分析师报告——几秒钟内完成两小时的分析师工作。
MCP Server (github.com/financial-datasets/mcp-server) - 通过 MCP 协议提供的金融数据集。经过类型化和验证的数据直接输入 Claude 的上下文窗口。
Crucix (github.com/calesthio/Crucix) - 链上聚合器。Polygon 上的巨鲸钱包移动可能比可见订单簿提前几分钟发出信号。
fredapi (github.com/mortada/fredapi) - 美联储发布的每个宏观经济数据集，免费，封装在一个 Python 包中。CPI、失业率、收益率曲线——直接导入 Claude 的上下文以锚定任何与宏观相关的市场。
Binance Collector (github.com/txbabaxyz/mlmodelpoly) - 预测市场方向并计算短期 BTC/ETH 合约的公平价值。这是每一次延迟套利交易中“价格发现”的一半。
Polymarket Assistant Tool (github.com/FiatFiorino/polymarket-assistant-tool) - 指标引擎，用于揭示实时市场的方向性偏差。将原始订单簿 + 价格行为转化为可用的信号。
lightweight-charts (github.com/tradingview/lightweight-charts) - TradingView 自有的图表库。14k 星，45KB，免费。这是构建用于监控仓位和信号的顶级仪表板的最快路径。
## 第 4 层 - MARKET INTELLIGENCE（市场情报）
别人已经构建好的东西。你不必从零开始构建一切。整个 Polymarket 特定的情报和预构建机器人生态系统已经存在。
巨鲸预警、排行榜、信号共振——这是散户以前从未拥有过的机构级视野。
Polyscope (thepolyscope.com) - 扫描 2,000 多个市场。巨鲸预警（$5k/$10k/$20k）、概率偏移、成交量激增——所有信息实时推送到 Telegram。
Polywhaler (polywhaler.com) - $10k+ 巨鲸交易追踪器。实时预警、内幕检测、AI 生成的信号。
WHALES tracker (Apify) - 交叉引用 Gamma + Data + CLOB。输出 Smart Money Consensus（聪明钱共识）、Health Score（健康评分）、Conviction Score（信念评分）作为结构化 JSON。
HyperBuildX bot (github.com/HyperBuildX/Polymarket-Trading-Bot) - 基于 Rust，延迟低于 100ms。当多个巨鲸涌入同一市场时，AI 会对跟单信号进行排名。
polymarketanalytics.com - 开放式交易员分析。顶级钱包盈亏、最大盈利、与 Kalshi 的交叉对比。
polyrec (github.com/txbabaxyz/polyrec) - 实时终端仪表板，结合了 Chainlink 预言机、Binance 数据源和完整订单簿深度。70 多个指标，自动 CSV 记录，内置策略回测器。这是你能免费获得的最接近机构驾驶舱的东西。
Polymarket-Trading-Bot (github.com/dylanpersonguy/Polymarket-Trading-Bot) - 5.3 万行 TypeScript。七种预建策略：套利、动量、做市、AI 预测、巨鲸跟单、收敛等。一个完整的框架——选择策略，插入密钥，运行。
## 第 5 层 - BACKTEST & SIMULATION（回测与模拟）
运行前先证明。这是大多数散户机器人跳过的一层——也是大多数散户机器人爆炸的原因。
推理模型可能会对从未赚过钱的策略说得头头是道。巨鲸可能会被跟单进入已经启动的交易。上个月有效的信号可能昨天就已经被套利消失了。唯一的防御措施是在一美元真金白银接触 CLOB 之前，针对历史数据和模拟执行运行每一个想法。
每个策略都必须先通过这份报告才能投入生产：17 个月的实盘记录，包含费用和滑点，回撤清晰可见。
prediction-market-backtesting (github.com/evan-kolberg/prediction-market-backtesting) - 针对真实的 Polymarket 和 Kalshi 历史数据回测交易策略。任何新想法投入生产前的第一道理智检查。
polybot (github.com/ent0n29/polybot) - 完整的执行和市场数据基础设施，包含模拟交易。Kafka、ClickHouse、Grafana——机构交易台会构建的分析管道。从生产策略中逆向工程而来。
# 完整的信号流。
从世界事件到执行：在理想情况下不到 10 秒。
# 人类交易员 vs 机器人 - 数据。
人类交易员与使用类似延迟套利方法的自动机器人之间的表现差距是有据可查的。在一个追踪期内，机器人产生了约 206,000 美元的收益。使用相同逻辑的人类只产生了约 100,000 美元。
2 倍的差距。同样的市场。同样的策略。同样的时间窗口。
差距不在于更好的预测。而在于执行。人类会犯四种机器人不会犯的系统性错误：
- 入场太晚。当人类识别出 Polymarket 的滞后，确认 Binance 的变动，并手动下单时，窗口往往已经关闭。
- 仓位不一致。人类在自信时过度下注，在不确定时下注不足——这正是凯利数学的反面。情绪化仓位在数千次交易中摧毁了期望值。
- 疲劳。一个通过 8 小时会议监控 15 分钟 BTC 合约的人类，其表现会下降。一个运行了 72 小时的机器人在第 72 小时做出的决定与第 1 小时完全一样。
- 回撤心理学。在一系列亏损后，人类要么放弃有效的策略，要么加倍下注。两者都会摧毁资金。一个带有硬性终止开关的机器人两者都不会做。
# 为什么是现在。
已经在运行的机器人具有复合优势。这个优势今天依然存在。窗口正在缩小。了解这个技术栈的最佳时机是六个月前。
第二好的时机就是现在。
如果你读到了这里：
-> 保存这篇文章，否则你会丢失所有重要链接。
-> coinman2 机器人的 Polymarket 资料：钱包链接
-> 如果这看起来太复杂，复制类似机器人的交易也是一种盈利策略。这正是我通过 TG 机器人所做的：kreo.app/@cvxv666
-> 记住：竞争比看起来要弱得多。99.9% 的人会尖叫“太难了”并称之为扯淡。只有 0.01% 的人真正去尝试……并以此获利。