An autonomous quantitative trading system that identifies and exploits pricing inefficiencies in Polymarket weather derivative markets. The bot ingests multi-model meteorological ensemble forecasts, computes a probability distribution over temperature outcomes, and bets when its model price diverges significantly from the CLOB market price.

Direction breakdown:

Top 5 trades:

City breakdown:

Note: Paper trading uses real-time CLOB mid prices for entry simulation. Actual live fills would incur bid-ask spread costs (~0.5–2pp per trade) not reflected here.

The system has been deployed live on Polygon via the Polymarket CLOB API. As of Apr 5, 2026, 18 real positions are on-chain across 8 cities (Atlanta, Buenos Aires, Chicago, Miami, Sao Paulo, Seattle, Munich, San Francisco), entered between Apr 1–3, 2026. Live resolution tracking is active and awaiting market settlement.

polymarket-weather-bot/
├── main.py                  # CLI entry point — scan, exit-scan, resolve, status
├── daemon.py                # Scheduler: runs scans on NWP model update cadence
├── config.py                # All tunable parameters (edges, Kelly, cities, filters)
├── db.py                    # SQLite layer (paper_trades.db / live_trades.db)
│
├── data/
│   ├── openmeteo.py         # Multi-model forecast fetcher (GFS, ECMWF, ICON, GEM, MF)
│   ├── noaa.py              # ASOS station historical observations
│   ├── polymarket.py        # CLOB price + order book queries
│   ├── climatology.py       # Historical temperature distribution
│   └── ...
│
├── signals/
│   ├── ensemble.py          # Combines 5 NWP models → mean, std, score
│   ├── edge_calculator.py   # Core alpha engine: model prob → Kelly bet
│   ├── nowcaster.py         # Blends live observations into forecast
│   ├── bias_corrector.py    # Per-city/model bias correction
│   └── ...
│
├── broker/
│   ├── paper_broker.py      # Simulated execution with all risk checks
│   ├── live_broker.py       # Real CLOB order submission via py-clob-client
│   └── position_manager.py  # Resolution engine + P&L settlement
│
├── metrics/
│   ├── calibration.py       # Shrinkage factor from resolved trade history
│   ├── sharpe.py            # Rolling risk-adjusted return tracking
│   └── reporting.py         # Dashboard data aggregation
│
└── web_dashboard.py         # Flask dashboard with live P&L + open positions

1. Forecast ingestion — Fetches the 5 latest NWP runs for the target ICAO station from Open-Meteo: GFS, ECMWF, ICON, GEM, Météo-France.

2. Ensemble statistics — Computes mean, standard deviation, and spread score. High std (models disagree) increases position uncertainty. Low std for NO bets is a skip signal — tight model agreement means temperature is likely heading for a specific bucket.

3. Nowcasting — If the station has a real-time ASOS observation for the current day, it blends into the forecast using a time-weighted regime (observation weight rises from 0% at midnight to 80% by market close).

4. Bucket probability — Integrates a Student-t distribution (ν=4, fat tails) over the temperature bucket [lo, hi] to compute P(actual_high ∈ bucket).

5. Edge calculation — edge = model_prob − market_mid_price. Only signals where |edge| > adaptive_min_edge are tradeable.

6. Filters applied before entry:

   • Minimum edge threshold (dynamic, scales with lead time)
   • NO entry price gate: 0.20 – 0.75 (empirically optimal range)
   • Ensemble std gate: skip NO bets when std < 0.8°C (model consensus = bad for NO)
   • Order book depth check: skip if top-5 CLOB depth < $150 USDC
   • Timing filter: skip if market price has been converging toward model value
   • Global deployment cap: max 40% of portfolio in open positions

7. Kelly sizing — Fractional Kelly (25%) with tier scaling and a hard cap of $15/trade.

pip install -r requirements.txt
# Key deps: py-clob-client, flask, requests, python-dotenv, scipy

cp .env.example .env
# Fill in POLYMARKET_PRIVATE_KEY, POLYMARKET_PROXY_ADDRESS, and API credentials
# Generate CLOB API creds: python scripts/gen_clob_creds.py

# Paper trading scan (safe, no real money)
python main.py --scan --paper

# Live scan (real orders)
python main.py --scan --live

# Check for exits / stop-losses
python main.py --exit-scan --live

# Resolve settled markets and update P&L
python main.py --resolve --live

# Web dashboard (port 5001)
python web_dashboard.py

# Autonomous daemon (scheduled scans + exits)
python daemon.py

The polybot.service and polybot-dashboard.service files are ready-to-use systemd unit files for running the daemon and dashboard as persistent services on a GCP VM.

sudo cp polybot.service /etc/systemd/system/
sudo systemctl enable polybot
sudo systemctl start polybot

This is an experimental research project. Prediction market trading involves real financial risk. Past paper trading performance does not guarantee live results. The average edge per trade is small and individual positions can result in total loss.