Polymarket Weather Trader

When to Use

Use this skill when the user wants to:

• Scan active weather markets on Polymarket
• Execute temperature forecast arbitrage trades (US via NOAA, international via Open-Meteo)
• View or configure weather trading parameters
• Check open positions or trading performance
• Switch between Simmer virtual ($SIM) and Polymarket real (USDC) trading

Architecture

weather-trader/
├── weather_trader.py       # Main entry point (2732 lines)
├── ev_calculator.py        # EV + Kelly Criterion + longtail bias
├── bayesian_update.py      # Position probability management
├── maker_taker_arbiter.py  # Dynamic Maker/Taker mode switching
├── trade_performance.py    # Logging, circuit breaker, health score
├── performance_report.py   # Performance reporting CLI
├── smart_money_signal.py   # Whale tracking signal
├── config.json             # Trading parameters
├── test_ev_calculator.py   # Unit tests
├── test_noaa_calibration.py
└── trades.jsonl            # Append-only trade log

Quick Start

cd weather-trader

python weather_trader.py                    # Dry run
python weather_trader.py --live            # Real trades (venue from .env)
python weather_trader.py --positions       # Show open positions
python weather_trader.py --config          # Show current config
python weather_trader.py --set entry_threshold=0.20
python weather_trader.py --live --quiet     # Silent for cron
python weather_trader.py --smart-sizing     # Portfolio-based sizing
python weather_trader.py --no-safeguards   # Disable safeguards (not recommended)
python weather_trader.py --resume          # Clear circuit breaker

Venue and Live Mode

Supported Locations

US cities (NOAA + MADIS, 16 cities): NYC, Chicago, Seattle, Atlanta, Dallas, Miami, Houston, Phoenix, Philadelphia, Boston, Denver, Las Vegas, San Francisco, Los Angeles, San Diego, Washington DC, Nashville

International cities (Open-Meteo ensemble, 20+ cities): Tel Aviv, Munich, London, Tokyo, Seoul, Ankara, Lucknow, Wellington, Taipei, Shanghai, Beijing, Hong Kong, Singapore, Sydney, Melbourne, Paris, Berlin, Madrid, Rome, Amsterdam, Toronto, Vancouver, Shenzhen, Guangzhou, Chengdu, Mumbai, Delhi, Wuhan

Use config.json or --set locations=NYC,London,Tokyo to configure.

Strategy Logic

Signal = NOAA/Open-Meteo forecast probability vs Polymarket market price
EV     = p × (1-price)/price - (1-p) × 1
Trade  = take if EV > 0 and price < entry_threshold
Size   = Kelly Criterion (Quarter Kelly = kelly_multiplier × full Kelly)

1. Fetch active weather markets from Simmer/Polymarket API
2. Parse market name → location + temperature bucket
3. Fetch weather forecast (NOAA for US, Open-Meteo ensemble for international)
4. Compute forecast probability → derive YES price estimate
5. Check EV gate, safeguards, slippage, time decay
6. Execute with Kelly sizing; apply longtail EV penalty for <20¢ contracts
7. Track open positions; apply Bayesian update if market moves against position
8. Exit when price > exit_threshold or position P&L drops below threshold

Core Modules

ev_calculator.py — EV + Kelly Framework

EV = (p × (1-price)/price) - ((1-p) × 1)
Quarter Kelly = full_Kelly × kelly_multiplier

• calculate_ev(price, win_prob) → EV per dollar
• should_take_trade(price, win_prob, min_ev=0.0) → bool gate
• kelly_fraction(price, win_prob) → full Kelly (0–1)
• quarter_kelly_position(price, win_prob, bankroll, max_pos) → USD amount
• is_longtail_contract(price, cutoff=0.20) → flag for <20¢ contracts
• longtail_ev_adjustment(price, raw_ev) → applies 20% penalty to longtail EV

bayesian_update.py — Position Probability Management

• bayesian_update(prior_prob, likelihood_ratio) → posterior probability
• compute_likelihood_ratio(new_price, old_price) → odds ratio
• should_update_probability(position_pnl_pct, threshold=-0.10) → triggers on -10% PnL
• should_close_position(updated_prob, initial_prob, threshold=-0.15) → close if prob dropped >15%
• compute_new_kelly_size(initial_prob, updated_prob, initial_kelly_size) → rescales position

maker_taker_arbiter.py — Strategy Mode Switching

Configuration (config.json)

Performance & Monitoring

python trade_performance.py --health-score   # 0-100 health score
python trade_performance.py --auto-tune     # Suggest param adjustments
python trade_performance.py --changelog      # View param change history
python trade_performance.py --resume         # Force clear circuit breaker

• Structured logging: Every trade → trades.jsonl (JSONL, append-only)
• Circuit breaker: 3 consecutive real losses → 6-hour pause
• Health scoring: 4-component weighted score (0-100) with half-life decay
• Changelog: Immutable audit trail of parameter changes

Risk Controls

• TRADING_VENUE=sim (virtual $SIM) or polymarket (real USDC)
• --live flag required for real API calls
• Circuit breaker auto-pause
• Max position per trade
• Slippage checks (skip if >15%)
• Time decay guard (skip if <2h to resolution)
• Min liquidity filter

Changelog

v2.1 (2026-04-08) — MADIS Bias Correction + Trade Log

• MADIS airport-observation bias correction: Compare MADIS current observation (airport实测) vs forecast temperature. If forecast is systematically warm or cold, adjust probability accordingly (max ±6% shift). This addresses the core issue identified in @runes_leo's weather trading experience: forecasts can have city-level directional biases.
• Trade log for city-level bias learning: Every executed trade now writes to ~/.hermes/weather-trade-log.jsonl with: city, target_date, forecast_temp, bucket, price, noaa_probability (pre-correction), bias_adjustment, MADIS bias message, and signal data. Over time this builds a per-city forecast bias statistics database.

Known Constraints

Polymarket MIN_SHARES_PER_ORDER = 5

At $0.10/bucket → $0.50 minimum (OK) At $0.20/bucket → $1.00 minimum (OK) At $0.30/bucket → $1.50 minimum (blocks $1 trades) At $0.50/bucket → $2.50 minimum (blocks $1 trades)

For small budgets ($1/position): only buckets ≤ $0.20 are tradeable.

Troubleshooting