Polymarket Multi-Strategy Trading Bot

Polymarket Multi-Strategy Trading Bot

Automated prediction market trading bot for Polymarket with multi-strategy architecture, AI-powered crowd simulation, and self-evolving parameter optimization.

Performance

Metric	Value
Weather strategy (best)	+$2,323 PnL, 63.4% WR, 325 trades
Golden era (Mar 3-14)	+$1,908, 72% WR, $159/day
Perplexity API savings	$57+ via local Ollama LLM
Strategies	10 built, 3 active (focus > diversification)

Architecture

┌─────────────────────────────────────────────────────────┐
│                    ORCHESTRATOR                          │
│  Market scanning → Signal validation → Risk management   │
├─────────────┬───────────────┬───────────────────────────┤
│  STRATEGIES │   AI LAYER    │      DATA PIPELINE        │
│             │               │                           │
│  Weather    │  4096-Agent   │  REST API (Gamma/CLOB)    │
│  Sniper     │  Crowd Sim    │  WebSocket (real-time)    │
│  Econ Data  │  Deep Research│  SQLite WAL (storage)     │
│  Crowd Sport│  Bayesian     │  Market lifecycle         │
│             │  Updating     │  VWAP / Flow analysis     │
├─────────────┴───────────────┴───────────────────────────┤
│                   RISK MANAGEMENT                        │
│  Quarter Kelly · Triple Barrier · EV Gate · Base Rates   │
├─────────────────────────────────────────────────────────┤
│                 SELF-EVOLVING OPTIMIZER                   │
│  Genetic Programming · Meta-Optimizer · Auto-Compound    │
└─────────────────────────────────────────────────────────┘

Strategies

Active

Strategy	Description	Edge Source
Weather	BUY_NO on temperature range markets	Multi-source forecast divergence (WU + OpenMeteo)
Resolution Sniper	Near risk-free resolution sniping	UMA oracle + local LLM verification
Econ Data Sniper	Trade on BLS data releases within ~1s	Polls BLS page at 100ms, parses headline number
Crowd Sport	50-agent Delphi simulation on sport markets	10 specialist groups, 3-round consensus
Crowd Prediction	4096-agent hierarchical simulation on politics/crypto/geopolitics/entertainment	64 groups → 16 panels → 4 summits → consensus

Key Design Decisions (from 112K wallet study)

• Specialize: Focus on 1-2 categories (+$4,200 avg) instead of 5+ (-$2,100 avg)

• Quarter Kelly: Never risk more than 25% of Kelly optimal

• Price-based exits: Sell on price movement (18-72h hold), don't wait for resolution

• Min edge 8%: Top 1% enter at 8-10% deviation from consensus

• Never negative EV: Strict EV ≥ 0.10 gate on all entries

Core Formulas

# 1. Expected Value — decides every entry
EV = win_prob × payoff_ratio - (1 - win_prob)  # must be ≥ 0.10

# 2. Bayes — chains evidence updates
P(H|E) = P(E|H) · P(H) / P(E)

# 3. Kelly — position sizing
f* = (p · b − q) / b  × 0.25  # quarter Kelly

# 4. Base Rate — the invisible edge
edge = base_rate - market_price  # trade when |edge| > 8%

# 5. KL-Divergence — arbitrage scanner
D_KL(P‖Q) = Σ Pᵢ · ln(Pᵢ / Qᵢ)  # finds mispriced correlated markets

Tech Stack

Trading

• Polymarket CLOB — order execution (Builder Program, gasless)

• Horizon SDK — TWAP/VWAP/Iceberg for large orders

• PMXT — cross-platform scanning (Polymarket + Kalshi)

AI / ML

• DeepSeek V3 — crowd simulations via LiteLLM proxy (~$0.001/call)

• Ollama (qwen2.5:0.5b) — local LLM for resolution verification

• all-minilm — market embeddings for correlation clustering

• TSFresh — automatic feature extraction for meta-labeler

• Prophet — PnL forecasting with weekly seasonality

• ARCH — GARCH/EGARCH/GJR volatility modeling

Data

• Unusual Whales — congress trades, dark pool, insider signals, crypto whales

• FRED — economic data (NFP, unemployment, CPI consensus)

• ESPN — sport statistics for crowd simulations

• SQLite WAL — structured storage with idempotent writes

Risk

• Riskfolio-lib — CVaR/MVO/HRP portfolio optimization

• VectorBt — vectorized backtesting (5-10x faster optimizer)

• PyFolio — tearsheet analytics (Sharpe, Sortino, drawdown)

Self-Evolving System

The bot evolves its own optimization function:

AutoOptimizer v4.0 — Self-Evolving Scoring Genome
┌────────────────────────────────────────────────┐
│ ScoringGenome (15 mutable coefficients)        │
│   wr_base, wr_scale, wr_center, wr_range       │
│   pf_base, pf_scale, pf_center, pf_range       │
│   vol_norm, vol_exp                             │
│   pnl_exp, wr_exp, pf_exp                      │
│   simplicity_floor, drift_tolerance             │
├────────────────────────────────────────────────┤
│ Meta-Optimizer (genetic programming)           │
│   Population: 12 genome variants               │
│   Selection: test PnL (not genome score)       │
│   Generations: 5 per run, 2x/day               │
│   Result: genome that produces best REAL params │
└────────────────────────────────────────────────┘

Setup

# Clone
git clone https://github.com/ludovicocapuano-dev/polymarket-bot.git
cd polymarket-bot

# Configure
cp .env.example .env
# Edit .env with your API keys

# Run (paper trading)
python3 bot.py

# Run (live — requires confirmation)
echo 'CONFERMO' | python3 bot.py --live

Required API Keys

• Polymarket CLOB (private key + API credentials)

• DeepSeek (for crowd simulations)

• Unusual Whales (for signal intelligence)

Optional

• Anthropic (for premium simulations)

• Horizon SDK (for advanced execution)

• FRED (for economic data)

• Zep Cloud (for MiroFish memory)

Cron Jobs

Schedule	Task
4x/day (00:23, 06:23, 12:23, 18:23)	AutoOptimizer + meta-evolve
2x/day (08:00, 20:00)	GitHub repo monitoring (8 repos)
1x/day (04:00)	AutoContext bridge

Monitored Repositories

• karpathy/autoresearch

• hyperspaceai/agi

• greyhaven-ai/autocontext

• 666ghj/MiroFish

• MiroMindAI/MiroThinker

• polymarket/polymarket-cli

• polymarket/py-clob-client

• polymarket/py-builder-signing-sdk

Key Lessons Learned

1. Reserve floor must scale with real capital — caused trading stall when floor > USDC

2. Weather works with golden-era params (5-8% min edge) — don't over-tighten

3. BLS blocks bots without browser User-Agent — need Chrome headers

4. Favorite-longshot bias doesn't work on Polymarket (alpha too low, 0% WR)

5. MiroFish full simulation is too slow for trading — Delphi hierarchical is 1000x faster

6. 112K wallet study: specialize, quarter Kelly, price exits, never negative EV

7. 75% of Polymarket markets resolve NO — structural tailwind for BUY_NO strategies

License

Private repository. All rights reserved.

---

Built with Claude Code