Trading bots generated over $60M in profit on Polymarket in 2025–2026. 77% of that came from the Crypto UP/DOWN market - driven by persistent structural inefficiencies. Here's how to build one.
01 - The opportunity
Why BTC Up/Down markets
The BTC 5-minute Up/Down market on Polymarket is one of the most inefficient segments in prediction markets. The crowd prices directional moves based on emotion - news cycles, social media, gut feel.
Meanwhile, the transition matrix of BTC price states shows something different. When the market is in a committed directional state - the persistence is measurable. The math knows before the crowd does.
That gap between what the math says and what the market prices is the edge. And it's repeatable, scalable, and automatable.
The agent framework we're using is Hermes - open-source, built by NousResearch (backed by Paradigm with $70M). By April 2026, Hermes surpassed Anthropic's Claude Code in total GitHub stars - a clear signal of how fast the developer community adopted it.
- 288 windows/day per asset
- 1 trade every 81 seconds
- Edge window: 5–15% avg gap
- Win rate: 63–72% at p ≥ 0.87
Top Successful bots running right now
Combined: $2,112,019. Three bots. One market segment. Same underlying math.
02 - The edge
How the math works
The model is based on Markov Chain analysis of BTC price states. The core insight: price movement is not random. When the market enters a persistent directional state, the probability of continuation is measurably above 50%.
The entry formula
Δ⁽ʷ⁾ = p̂⁽ʷ⁾ − q⁽ʷ⁾ ≥ ε → ENTER p̂ = model probability · q = market price · ε = 5% minimum gap
r = (1 − q) / q At q = 0.647 → r = +54.5% per trade · At q = 0.441 → r = +126.7% per trade
The bot only enters when p(j\,j\) ≥ 0.87 - the Markov persistence threshold. Below that, no trade. This is why the win rate is consistently above 65% despite no directional prediction.
Kelly f\ = p − (1−p)/b Optimal position sizing per trade · f\ ≈ 0.71 at p = 0.87, b = 0.647
03 - The stack
What you need to build this
The entire setup runs on open-source tools. No coding required. Total cost: under $10/month
$10 min to start → $50 recommended → 2 POL for gas (~$1) → ~30 min setup
04 - Setup
How to set up Hermes in 3 steps
STEP 01
Install Atomic and launch Hermes
Go to atomicbot.ai → download Atomic → choose Hermes agent on the main page. You can run it locally on Mac or choose "Run in Cloud" in the top-right corner - login via Google, same interface. Move app to Applications folder after download.
Atomic offers 100+ integrations, persistent memory, and support for all major AI models (Claude, ChatGPT, Gemini).
STEP 02
Connect model API - use Claude Opus 4.7
In Atomic settings → AI Models → select Anthropic → paste your API key. Choose Claude Opus 4.7 as the model engine - it has the reasoning capacity needed for real-time market analysis and self-improvement loops.
Alternatively: OpenRouter (pay-as-you-go) or OpenAI Codex (free via ChatGPT Pro).
STEP 03
Connect Telegram bot to your agent
Atomic → Skills → Messengers → Telegram → Connect. Create a bot via @BotFather in Telegram → copy token → paste into Atomic. Done in 2 clicks.
From this point your Hermes agent is live and waiting for your trading logic prompt.
05 - Trading logic
Setting up the BTC trading strategy
Instead of building from scratch, use an existing GitHub repo as the base logic - then feed it to Hermes and let Claude Opus adapt it to the latest Polymarket CLOB v2.
Recommended repos
Step 1 - Give Hermes the trading logic prompt
Step 2 - Set up wallet
Step 3 - Environment config
Step 4 - Run dry test first
06 - Self-learning loop
How the agent improves itself
This is what separates Hermes from a static bot. Claude Opus 4.7 reads the execution journal after every session and rewrites the trading rules based on what worked and what didn't.
- Trade executes
Bot enters market at p(j\,j\) ≥ 0.87. Every entry, exit, and P/L is logged to journal.
- Nightly review
Claude Opus reads the full journal. Analyzes which persistence thresholds performed, which windows lost, which entry prices had best EV.
- Strategy update
Opus rewrites the threshold rules, adjusts Kelly sizing, and updates MIN_PROB and MIN_EDGE parameters automatically.
- Next session runs with updated rules
The agent is measurably smarter after 50–100 trades. Let the AI do the heavy lifting.
- Telegram report every morning
Yesterday's trades, updated rules, today's strategy. You review, approve, it runs.
Conclusion
Polymarket trading bots have already taken a large share of the profit from manual traders - and this percentage keeps increasing daily.
With agentic frameworks like Hermes and Atomic, you don't need to be a senior developer to build your own. You need Claude Opus as the brain, a GitHub repo as the starting logic, and time for 50–100 training trades.
The self-learning loop does the rest.
Start small. DRY_RUN=true first. $1–$2 per trade while training. The agent improves with every trade it executes - don't rush the learning phase.
Top example of Bots from article:
https://polymarket.com/@bonereaper?r=joinjoinjoin#tLcpwsE https://polymarket.com/@0xe1d6b51521bd4365769199f392f9818661bd907?r=joinjoinjoin#9TKvd55 https://polymarket.com/@0xb27bc932bf8110d8f78e55da7d5f0497a18b5b82-1772569391020?r=joinjoinjoin#lIVnuAb
Fastest way to find all insights and their next trade before they blow up: https://predictparity.com?code=ricky
