A TypeScript bot that trades Polymarket daily-high temperature markets ("Will the highest temperature in Seoul be 29°C on June 19?") using a probabilistic forecast instead of a point guess.

It does what sharp weather traders do by hand: build a distribution of the day's high temperature, convert it into a probability for every °C bucket the market exposes, compare that to the live order-book price, and bet — selectively and sized — only where the model and the market meaningfully disagree.

ensemble forecast ──► distribution over °C buckets ──► P(bucket)
                                                          │
live order book ──► market price per bucket ─────────────┤
                                                          ▼
                                          edge = P(bucket) − price
                                          │  (gate on min edge / price / liquidity)
                                          ▼
                          fractional-Kelly stake ──► risk caps ──► order

npm install
cp .env.example .env        # then edit .env

Edit .env. Everything has a safe default and TRADING_MODE=paper, so you can run read-only commands with no wallet at all.

npm run forecast -- hong_kong          # print the temperature distribution
npm run forecast -- nyc 2026-06-20     # for a specific date

npm run scan                           # analyze all markets, show edges + intended
                                       # trades — never places orders

npm run run-once                       # one pass that executes (paper or live)
npm run watch -- 10                    # repeat run every 10 minutes

npm run auth                           # derive CLOB API creds to cache in .env
npm run portfolio                      # positions, exposure, realized PnL

scan is always read-only and needs no credentials — start there.

1. Fund a dedicated Polygon wallet with USDC (not your main wallet).
2. Set PRIVATE_KEY, FUNDER_ADDRESS (your Polymarket profile address) and SIGNATURE_TYPE (0 EOA / 1 Magic-email / 2 Safe) in .env.
3. Run npm run auth once and paste the printed creds back into .env.
4. Set BANKROLL to no more than your real balance and keep KELLY_FRACTION low (0.25 = quarter-Kelly).
5. Flip TRADING_MODE=live. The bot refuses to start live without the keys.

The forecast quality is only as good as its alignment with the official resolution station. Two knobs in src/weather/cities.ts matter most:

• rounding — does the station report a rounded (26 = 25.5–26.5) or floored integer? Wrong choice shifts every probability.
• biasC — a small additive correction if the model's 2 m temperature runs systematically hot/cold vs the station's microclimate.

When you run scan you will sometimes see very large model-vs-market edges. Treat those as a prompt to verify your calibration, not as free money — the market often knows the station's quirks better than an uncalibrated model does. Use archiveDailyHighs() in src/weather/openmeteo.ts to backtest model output against historical official highs and tune biasC before risking capital.

This software is for research/educational use. Trading involves real financial risk; you are responsible for your own funds and decisions.

• The @polymarket/clob-client dependency is archived but still the most battle-tested way to trade the live CLOB; it's isolated behind src/polymarket/clob.ts so migrating to Polymarket's new unified SDK later touches only that one file.
• No paid services required: Open-Meteo and NWS are both free and keyless.

Reference