AI-powered trading agent for prediction markets
Finds mispriced odds on Polymarket using ensemble weather forecasts and LLM superforecasting.
How It Works · Weather Strategy · LLM Strategy · Architecture
Mispricer scans Polymarket for low-probability contracts ($0.01–0.15) where the real probability — estimated by data APIs or AI — is significantly higher than what the market implies. It makes many small bets ($1–5), exploiting systematic mispricing.
1. Scan → Find active prediction markets priced $0.01–0.15
2. Classify → Route each market to the best-fit strategy
3. Analyze → Each strategy independently estimates the real probability
4. Filter → Keep only signals with sufficient edge and confidence
5. Execute → Position sizing via fractional Kelly criterion with risk limits
The agent runs automatically on a schedule: check open bets for resolution, scan for new opportunities, analyze with active strategies, and execute trades.
Compares 31-member GFS ensemble forecasts with Polymarket weather contract prices. When the ensemble predicts a significantly higher probability than the market implies, we bet.
- Edge source: Calibrated probabilistic data (31 independent simulations)
- Kelly fraction: 0.25x (quarter-Kelly — higher confidence from verified data)
- Min edge: 8%
Example: Market asks "Will Miami hit 90°F on March 18?" priced at $0.04 (4%). Ensemble forecast: 12 of 31 members ≥ 90°F → 38.7%. Edge: +34.7% → BUY
AI-powered probability estimation for politics, entertainment, science, and other categories using structured superforecasting methodology — base rates, evidence decomposition, calibrated probability.
- Edge source: AI judgment (Claude)
- Kelly fraction: 0.10x (conservative — LLM estimates less calibrated than weather data)
- Min edge: 8% AND confidence ≥ MEDIUM
Example: Market asks "Will Oppenheimer win Best Picture?" priced at $0.12 (12%). LLM analysis: strong precursors, SAG ensemble, $950M box office → 65%. Edge: +53% → BUY
| Component | What It Does |
|---|---|
| Market Scanner | Continuously monitors Polymarket for mispriced contracts |
| Classifier | Routes each market to the most suitable analysis strategy |
| Weather Engine | Ensemble weather forecasts (31 independent simulations) |
| LLM Engine | AI superforecasting with structured reasoning and evidence |
| Risk Manager | Kelly criterion position sizing with per-strategy risk limits |
| Order Executor | Places orders with slippage protection |
| Dashboard | Real-time PnL tracking, strategy performance, admin controls |
| Notifications | Telegram alerts for bets placed and resolved |
Web dashboard with real-time PnL tracking, strategy performance breakdown, bet history, calibration charts, and admin controls.
- Cumulative PnL and bankroll charts with timeframe selectors (7D/30D/90D/All)
- Per-strategy performance table (win rate, ROI, avg edge)
- Calibration plot — forecast probability vs actual win rate
- Multi-language support (EN/RU)
- Role-based access (admin, client, viewer)
- Client self-service: wallet credentials, Telegram notifications, 2FA
| Parameter | Description |
|---|---|
| Kelly fraction | How aggressively to bet — higher for data-verified strategies, lower for AI judgment |
| Min edge | Minimum probability gap between our estimate and market price |
| Bet limits | Min and max bet size ($1–$5) with daily exposure cap |
Each strategy runs independently with its own risk parameters, allowing fine-tuned control over bet sizing and market selection.
- Agent: Python 3.11+
- Dashboard: Next.js, Tailwind CSS, Recharts
- Data: Polymarket CLOB API, Open-Meteo GFS ensemble, Claude API
- Database: SQLite
- Notifications: Telegram (per-client bots)
- Deployment: VPS, cron-scheduled
MIT — based on discountry/polymarket-trading-bot