John Doe
7.0 KiB
🌦️ WeatherBet — Polymarket Weather Trading Bot
Autonomous trading bot that exploits weather forecast errors to find mispriced Polymarket prediction markets — and self-improves over time.
🎯 What It Does
The bot monitors 6 US cities (NYC, Chicago, Miami, Dallas, Seattle, Atlanta) and bets on Polymarket's temperature prediction markets using real ECMWF weather forecasts as its edge. When the forecast predicts a temperature bucket, but the market price implies a different probability, the bot calculates the Expected Value (EV) and places a trade if EV > threshold.
💡 Why It Makes Money
The edge is weather forecast accuracy.
Polymarket traders rely on gut feel and consensus. This bot uses ECMWF — the world's most accurate weather model — to calculate the true probability of each temperature bucket, then compares it to the market price.
True Probability (from ECMWF) vs. Market Price (from Polymarket)
When Market Price < True Probability, the market is underpriced → BUY.
🧮 The Math
Step 1 — True Probability (Gaussian Bucket Model)
def bucket_prob(forecast_temp, t_low, t_high, sigma=2.0°F):
"""
The forecast says 72°F ± 2σ.
What's the probability the actual high falls in the 70-75°F bucket?
"""
from scipy.stats import norm
z_low = (t_low - forecast_temp) / sigma
z_high = (t_high - forecast_temp) / sigma
return norm.cdf(z_high) - norm.cdf(z_low)
Step 2 — Expected Value
def calc_ev(true_prob, market_price):
"""
EV = P(win) × payoff - P(lose) × cost
If EV > 0, the market underprices this outcome.
"""
win = true_prob * (1 / market_price - 1) # profit if we win
lose = (1 - true_prob) * 1 # we lose our stake
return win - lose
Example:
- Forecast: 72°F → 75% chance of 70-75°F bucket
- Market price: $0.30 (implies 30% probability)
EV = 0.75 × (1/0.30 - 1) - 0.25 = +1.25→ Strong BUY
Step 3 — Kelly Criterion (Optimal Bet Size)
def calc_kelly(p, price):
"""
Kelly % = (bp - q) / b
where b = 1/price - 1, p = true_prob, q = 1-p
"""
b = 1.0 / price - 1.0
f = (p * b - (1.0 - p)) / b
return round(min(max(f, 0.0) * KELLY_FRAC, 1.0), 4)
def bet_size(kelly, balance):
return round(min(kelly * balance, MAX_BET), 2)
- Uses 1/4 Kelly (conservative fraction) to survive variance
- Caps bet at
$2.00per trade - Only trades when EV ≥ 10% (adaptive floor, self-improving)
Summary: Why This Strategy Wins
| Component | Detail |
|---|---|
| Edge | ECMWF weather model is more accurate than consensus |
| Signal | Mispriced markets when Market Price < True Probability |
| Sizing | Kelly Criterion — mathematically optimal bet sizing |
| Filter | EV ≥ 10% (adaptive), volume > $500, spread < 3% |
| Execution | Real Polymarket CLOB on Polygon (not simulation) |
| Learning | Self-tuning Kelly fraction + EV floor from trade history |
🧠 Self-Learning System
After each trade, the bot records the outcome and adjusts its strategy:
data/learning/
├── trade_log.json # All trades: city, bucket, cost, outcome, pnl
└── model.json # Learned parameters per city/bucket
Adaptation rules:
- Winrate < 45% → Kelly fraction ×0.8, EV floor +10%
- Winrate > 55% + PnL > $2 → Kelly fraction ×1.1, EV floor −5%
- Per-city winrate tracking adjusts confidence in each market
- Starts conservative (25% Kelly) → converges to optimal as data accumulates
⚙️ Setup
Requirements
- Python 3.13+
- Polygon wallet with USDC.e (on chain 137)
- Polymarket CLOB approval
- Polymarket API credentials
Installation
git clone https://github.com/yourhandle/weatherbot.git
cd weatherbot
python3.13 -m venv venv
source venv/bin/activate
pip install -r requirements.txt
Configuration
Create .env:
PK=your_polygon_private_key
WALLET=your_polygon_address
SIG_TYPE=0
Edit config.json:
{
"balance": 0,
"max_bet": 2.0,
"min_ev": 0.10,
"min_volume": 500,
"scan_interval": 3600,
"telegram_bot_token": "your_token",
"telegram_chat_id": "your_chat_id"
}
Run
# One-shot scan
python bot_v3.py scan
# Continuous trading loop
python bot_v3.py run
# Check status
python bot_v3.py status
📊 Architecture
bot_v3.py
│
├── Weather Data
│ ├── ECMWF API — 10-day temperature forecast (primary signal)
│ └── METAR — current obs for D+0 override
│
├── Signal Evaluation
│ ├── bucket_prob() — Gaussian model → true probability
│ ├── calc_ev() — expected value vs market price
│ ├── calc_kelly() — optimal bet fraction
│ └── Adaptive floor — self-learning EV threshold
│
├── Execution
│ ├── py_clob_client — Polymarket CLOB on Polygon
│ ├── place_buy_order — market order with 10s timeout
│ └── on-chain settlement
│
├── Monitoring
│ ├── Telegram — real-time trade alerts
│ ├── Self-learning — trade_log + model.json
│ └── 60-min loop — continuous scan
│
└── Market Resolution
└── Outcome check — PnL update when market resolves
🔐 Trading Flow
1. Fetch ECMWF forecast for each city (D+0 to D+3)
2. Query Polymarket for temperature bucket markets
3. Calculate true probability (Gaussian model, σ=2°F)
4. Compare to market price → calc EV
5. If EV ≥ adaptive threshold → calculate Kelly bet size
6. Execute market order on Polymarket CLOB (Polygon)
7. Record trade → update self-learning model
8. Send Telegram notification
9. Repeat every 60 minutes
⚠️ Risk Management
| Parameter | Value | Purpose |
|---|---|---|
| Max bet | $2.00 | Cap per-trade exposure |
| Kelly fraction | 25% | Survive variance (1/4 Kelly) |
| Min EV | 10%+ | Only trade positive EV |
| Min volume | $500 | Avoid illiquid markets |
| Max spread | 3% | Avoid high-slippage markets |
| Adaptive floor | 10-20% | Self-tuning from performance |
📦 Tech Stack
- Language: Python 3.13
- Trading: py_clob_client — Polymarket CLOB
- Weather: ECMWF OpenMETAR / Open-Meteo API
- Chain: Polygon (Chain ID 137) — USDC.e stablecoin
- Notifications: Telegram Bot API
- Self-learning: Pure Python JSON persistence (no DB needed)
📝 Disclaimer
This bot trades real markets with real money. Past performance does not guarantee future results. Trade at your own risk. The bot is provided as-is for educational and research purposes.
Built with 🐍 on Polygon — autonomous weather prediction trading.