Autonomous trading agent for Polymarket prediction markets. Discovers markets, researches evidence, forecasts probabilities with a multi-model AI ensemble, and executes trades with strict risk controls.
Paper trading by default. Three safety gates must be unlocked for live orders.
Need help or an updated version?
Telegram: t.me/dexoryn · Discord:dexoryn_
Most Polymarket tools either mirror another wallet or require you to manually research, price, and size every market. This agent runs the full loop autonomously — discovery, evidence gathering, probability forecasting, risk checks, and execution — so you can focus on configuration and oversight instead of clicking through hundreds of markets.
- Independent forecasts — a 3-model LLM ensemble (GPT-4o, Claude, Gemini) estimates probabilities without anchoring to the current market price
- Autonomous research — category-aware web search, HTML extraction, and domain authority scoring pull evidence from primary sources (
bls.gov,sec.gov,fec.gov, and more) - Self-improving calibration — Platt scaling, Brier score tracking, and automatic retraining after resolved markets tighten forecasts over time
- 15+ pre-trade checks — drawdown heat, Kelly sizing, liquidity/spread filters, category exposure caps, and more; one failure blocks the trade
- Triple dry-run gate — paper trading by default; live orders require three independent unlocks
- Whale intelligence — tracks smart-money wallets and adjusts edge when whales agree or disagree with the model
- 9-tab live dashboard — engine status, positions, forecasts, risk, smart money, and performance metrics on port 2345
- Immutable audit trail — every decision logged with SHA-256 integrity checksums
- Multi-channel alerts — Telegram, Discord, and Slack for trades, drawdown warnings, and errors
| Capability | This AI Agent | Typical alternatives |
|---|---|---|
| Autonomous research & forecasting | Yes — full pipeline | Copy bots only mirror wallets |
| Multi-model ensemble | 3 LLMs with adaptive weighting | Single model or manual guesses |
| Calibration feedback loop | Retrains from resolved markets | Static or no calibration |
| Risk management depth | 15+ checks + drawdown heat | Basic limits or none |
| Smart money integration | Whale scanner + edge adjustment | Wallet copy only |
| Paper trading default | Triple safety gate | Often live-only or no guardrails |
| Live dashboard & audit trail | 9-tab dashboard + SHA-256 logs | CLI-only or opaque |
- Scans active markets from the Polymarket Gamma API with volume, liquidity, and spread filters
- 11-category classifier (MACRO, ELECTION, CORPORATE, LEGAL, TECHNOLOGY, SCIENCE, CRYPTO, REGULATION, GEOPOLITICS, SPORTS, ENTERTAINMENT) using 100+ regex rules — no LLM cost
- Each market gets a researchability score (0–100) that controls how much research budget it receives
- Pre-research quality filter blocks junk markets before any expensive API calls (~90% cost savings)
- Configurable cooldowns prevent re-scanning the same market too frequently
- Query builder generates site-restricted searches per category —
site:bls.govfor macro,site:sec.govfor corporate,site:fec.govfor elections - Includes contrarian queries to avoid confirmation bias
- 3 pluggable search backends — SerpAPI, Bing, Tavily — with automatic fallback if one fails
- Full HTML extraction via BeautifulSoup, not just search snippets
- Domain authority scoring — primary sources (1.0) > secondary (0.6) > unknown (0.3)
- Auto-filters low-quality domains (Wikipedia, Reddit, Medium, Twitter, TikTok)
- Source caching with configurable TTL (default 1 hour)
- Ensemble of 3 frontier LLMs running in parallel:
- GPT-4o (40% weight) — primary forecaster
- Claude 3.5 Sonnet (35% weight) — second opinion
- Gemini 1.5 Pro (25% weight) — third opinion
- 3 aggregation methods — trimmed mean, median, or weighted average
- Models forecast independently — explicitly told not to anchor to the current market price
- Graceful degradation — if a model fails, the ensemble continues with the remaining models
- Adaptive weighting — tracks per-model Brier scores by category and reweights over time
- Platt scaling — logistic compression pulling extreme probabilities toward 0.50
- Historical calibration — learns from past forecast vs. outcome pairs via logistic regression
- Evidence quality penalty — weak evidence pulls the forecast toward 0.50
- Contradiction penalty — conflicting sources increase uncertainty
- Ensemble spread penalty — when models disagree by more than 10%, adds uncertainty
- Calibration feedback loop — retrains automatically after 30+ resolved markets
- Brier score tracking — monitors forecast accuracy over time
Every trade must pass all of these — a single failure blocks the trade:
- Kill switch — manual emergency halt
- Drawdown auto-kill at 20% max drawdown
- 4-level drawdown heat system:
- Normal (< 10%) → full sizing
- Warning (≥ 10%) → half sizing
- Critical (≥ 15%) → quarter sizing
- Max (≥ 20%) → all trading halted
- Max stake per market ($50 default)
- Daily loss limit ($500 default)
- Max open positions (25)
- Minimum net edge after fees (4%)
- Minimum liquidity ($2,000)
- Maximum spread (6%)
- Evidence quality threshold (0.55)
- Confidence filter (MEDIUM minimum)
- Implied probability floor (5%)
- Portfolio category exposure cap (35% per category)
- Timeline endgame check (48h near resolution)
- Arbitrage detection — scans for mispriced complementary and multi-outcome markets
- Fractional Kelly position sizing with 7 multipliers (confidence, drawdown, timeline, volatility, regime, category, liquidity)
- Auto strategy selection:
- Simple — single limit order for small trades
- TWAP — splits large orders into 5 time-weighted slices
- Iceberg — shows only 20% of true order size
- Adaptive — adjusts pricing based on orderbook depth
- Triple dry-run safety gate:
dry_runflag on each order objectexecution.dry_runin config.yamlENABLE_LIVE_TRADINGenvironment variable- All three must allow it for a real order to go through
- Fill tracker — monitors fill rate, slippage, and time-to-fill per strategy
- 6 exit strategies — dynamic stop-loss, trailing stop, hold-to-resolution, time-based exit, edge reversal, kill switch forced exit
- Wallet scanner tracks top Polymarket traders seeded from the leaderboard
- Auto-discovers top 50 wallets by profit and top 50 by volume
- Delta detection — spots new entries, exits, size increases, and decreases
- Conviction scoring — combines whale count × dollar size into a signal
- Edge integration — whales agree with model → +8% edge boost; disagree → -2% penalty
- 7-phase liquid scanner pipeline:
- Seeds wallets from leaderboard → fetches markets → scans global trades → per-market whale scan → ranks addresses → deep wallet analysis → scores and saves to database
- API pool rotates requests across multiple endpoints with round-robin, least-loaded, or weighted-random strategies
- Order flow imbalance across 60min, 4hr, and 24hr windows
- VWAP divergence — signals entry when price is below volume-weighted average
- Whale order detection — flags individual trades above $2,000
- Trade acceleration — detects unusual activity surges (>2× baseline)
- Book depth ratio — measures bid vs. ask pressure
- Smart entry calculator — combines all signals to recommend optimal entry price
9-tab Flask dashboard with glassmorphism dark theme on port 2345:
- Overview — engine status, cycle count, P&L, equity curve
- Trading Engine — start/stop controls, cycle history, pipeline visualization
- Positions — open positions with live P&L, closed trade history
- Forecasts — evidence breakdown, model vs. market probability, reasoning
- Risk & Drawdown — drawdown gauge, heat level, Kelly multiplier, exposure breakdown
- Smart Money — tracked wallets, conviction signals, whale activity feed
- Liquid Scanner — 7-phase pipeline status, discovered candidates, API pool health
- Performance — win rate, ROI, Sharpe, Sortino, Calmar, category breakdown, model accuracy
- Settings — environment status, config viewer, kill switch toggle
Protected by DASHBOARD_API_KEY. Auto-refreshing with live status indicators.
- structlog JSON logging with automatic sensitive data redaction
- Multi-channel alerts — Telegram, Discord, Slack (with cooldowns to prevent spam)
- Alert triggers — trades, drawdown warnings, kill switch activations, errors, daily summaries
- Sentry integration — optional error tracking with data scrubbing
- API cost tracking — per-call cost estimation for LLM and search usage
- JSON run reports — exportable reports saved to
reports/
- SQLite with WAL mode for concurrent reads and writes
- 10 automatic schema migrations
- Immutable audit trail — every decision recorded with SHA-256 integrity checksums
- TTL cache — search results (1hr), orderbook (30s), LLM responses (30min), market list (5min)
- Automated backups with rotation (max 10), triggered via
make backup
git clone https://github.com/dylanpersonguy/polymarket-ai-trading-bot.git
cd polymarket-ai-trading-bot
python3 -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"
cp .env.example .env
Add your API keys to .env — at minimum OPENAI_API_KEY and SERPAPI_KEY.
make dashboard
Open http://localhost:2345.
Help: Telegram @dexoryn · Discord dexoryn_
cp .env.example .env
docker compose up -d
bot scan --limit 20 # discover markets
bot research --market <ID> # research a market
bot forecast --market <ID> # full pipeline: research, forecast, risk, size
bot paper-trade --market <ID> # simulated trade
bot trade --market <ID> # live trade (needs ENABLE_LIVE_TRADING=true)
bot engine start # continuous trading loop
bot engine status # engine health
bot dashboard # launch dashboard
bot portfolio # portfolio risk report
bot drawdown # drawdown status
bot arbitrage # scan for arbitrage
bot alerts # alert history
All config lives in config.yaml and .env.
Required keys — OPENAI_API_KEY, SERPAPI_KEY
Optional keys — ANTHROPIC_API_KEY, GOOGLE_API_KEY, BING_API_KEY, TAVILY_API_KEY, DASHBOARD_API_KEY
Live trading — set ENABLE_LIVE_TRADING=true and add POLYMARKET_API_KEY, POLYMARKET_API_SECRET, POLYMARKET_API_PASSPHRASE, POLYMARKET_PRIVATE_KEY
- Dry run by default — three independent gates (order flag, config flag, env var) must all allow live trading
- 4-level drawdown heat system — progressively cuts position sizes, halts at 20% drawdown
- No secrets in the codebase — everything via
.env - Docker runs as non-root user
make test
make lint
make format
Dexoryn Labs — Polymarket AI trading automation
- Telegram: @dexoryn (fastest response)
- Discord:
dexoryn_ - WeChat: DexorynWe — scan the QR code below
MIT
Trading on Polymarket involves significant risk of loss. Dexoryn is not responsible for losses from use of this software. Wallet security, configuration, and capital risk are your responsibility.
Only trade with funds you can afford to lose.
If this project helps you, consider starring the repo or opening an issue/PR. Questions: Telegram @dexoryn · Discord dexoryn_.
Built for the prediction market community. Not financial advice.