Polymarket Weather Trader

Trade temperature markets on Polymarket using NOAA forecast data.

🚨 Framework, not a production trading system. Read DISCLAIMER.md before connecting to a wallet with real funds.

Template skill. Defaults to dry-run mode (no real money). The --live flag is a deliberate single-command opt-in for real-money execution. Configure tunables (entry/exit thresholds, locations, etc.) via env vars listed below.

Safety rails (read first)

This skill executes real-money trades on Polymarket only when the --live flag is passed AND the human's wallet is linked to their Simmer account. Trading is bounded by default:

• Dry-run is the default. python weather_trader.py (no flag) shows opportunities but executes no trades. The --live flag is required for real-money execution. There is no "auto-graduate" path.
• $SIM paper sandbox option. Set TRADING_VENUE=sim to trade Simmer's $SIM virtual currency at real prices — useful for validating the strategy without USDC exposure.
• Real-money trading requires explicit human verification. A wallet must be linked at simmer.markets/dashboard before any real trade lands. Without a linked wallet the SDK rejects real-money order construction.
• Per-trade cap. SIMMER_WEATHER_MAX_POSITION_USD defaults to $2.00 per trade. Configurable via env var, capped at the user's dashboard-set platform per-trade limit.
• Daily caps. Platform-level daily caps apply (max trades/day, max USD/day). Set at simmer.markets/dashboard → SDK settings.
• Auto stop-loss is ON by default. Server-side risk monitor watches every buy. Threshold is configurable per user at simmer.markets/dashboard → Settings → Auto Risk Monitor. It cannot protect against gap-resolution, though: weather temperature buckets jump straight to about 0 at resolution rather than decaying through your stop, so a percentage stop has no price to trigger on and no liquidity to exit into. Size for the full loss, not for the stop. See DISCLAIMER.md.
• Strategy-side safeguards. Beyond platform risk monitors, this skill checks flip-flop, slippage (SIMMER_WEATHER_SLIPPAGE_MAX, default 15%), time-decay (SIMMER_WEATHER_MIN_HOURS_TO_RESOLVE, default 2h), and resolved-market status before every order. Disable only with --no-safeguards (not recommended).
• Reversibility. Open positions exit automatically when price > SIMMER_WEATHER_EXIT_THRESHOLD (default 0.45), or via client.cancel_order() / a manual sell. ENTRY_THRESHOLD is an upper bound (buy below). If you raise it above the exit default (e.g. entry 0.50 vs exit 0.45), the skill will try to sell the same position on the next cycle — raise EXIT_THRESHOLD too, or own exits yourself.

If anything above isn't clear, stop and ask the user before passing --live.

Strategy logic

Weather market outcomes are discrete: a temperature bucket ("34-35°F") either matches the actual high on resolution day or it doesn't. The strategy works when the NOAA forecast is more accurate than what the market has priced in.

Test before going live. The $SIM venue gives you a fully virtual sandbox at real market prices — recommended before any --live run.

Risk monitor. Stop-loss and take-profit thresholds are user settings (configurable at simmer.markets/dashboard → Settings → Auto Risk Monitor), shared across all skills under that user account. Per-position overrides via client.set_monitor(market_id, side, stop_loss_pct=..., take_profit_pct=...).

External wallet users: monitors emit alerts via the briefing endpoint — your agent must be running for sells to execute. Managed wallet users: server executes directly.

When to Use This Skill

Use this skill when the user wants to:

• Trade weather markets automatically
• Set up gopfan2-style temperature trading
• Buy low on weather predictions
• Check their weather trading positions
• Configure trading thresholds or locations

What's New in v1.23.21

• Per-market position cap (SIM-5499). SIMMER_WEATHER_MAX_BUYS_PER_MARKET (default 1) checks held positions before entry so a replay backtest and live trading measure the same strategy — replay previously DCA'd into every underpriced bucket every tick (mean 34 buys/market on a full-tape run) while live is naturally throttled by balance/backoffs. Run summaries now also print a per-location skip-reason breakdown so a location entering zero markets is diagnosable from the run's own output.

What's New in v1.23.19

• Replay archive builder hardening (SIM-5440). Open-Meteo HTTP 5xx responses now retry up to 3 attempts with backoff while 4xx remains fail-closed. Long windows are fetched in 10-day chunks and merged only when utc_offset_seconds is identical across chunks. --stations us|intl|all lets US-only backtests avoid international DST guards and unnecessary requests.

What's New in v1.23.17

• Timezone-correct leads (SIM-5434). Lead N is the smallest 1–3 such that every hourly issuance for the event day precedes the tick: N = ceil((E_end_utc − tick) / 24h), where E_end_utc is the event date's 23:59:59 in the station offset. Builder records Open-Meteo utc_offset_seconds per station in _meta. Missing offset assumes UTC−12 (tz=assumed). Each date must carry hours 00:00–23:00 exactly once.

What's New in v1.23.16

• Lead-aware archive (SIM-5434). Builder fetches temperature_2m_previous_day{1,2,3} in one request per station. Each date is {high, low, leads} with top-level high/low = lead 1. Under replay, _station_forecast(..., event_date=) picks lead = (event_date − tick.date).days + 1 in 1–3; further-out events skip. Incomplete hourly arrays abort the build. Provenance appends leads=1-3 when present. Hand-built / sample files without leads keep the old shape.

What's New in v1.23.15

• Historical forecast archive builder (SIM-5434). scripts/build_replay_forecast_archive.py fills the SIM-5429 loader plane from Open-Meteo Previous Runs. Tick D sees the D-1 forecast (temperature_2m_previous_day1 hourly → daily high/low). US stations are °F; international stations are °C. The file may include _meta (source, fetched_at, lead=previous_day1); the loader ignores it. Auto-load is still only fixtures/replay_forecasts.json (uncommitted). .sample.json stays shape-only and is never loaded.

What's New in v1.23.14

• Replay forecast archive (SIM-5429). This release ships a loader, not an archive. Under SIMMER_REPLAY=1 the skill fills _REPLAY_FORECASTS from SIMMER_REPLAY_FORECASTS=/path.json, or from a user file fixtures/replay_forecasts.json when present (uncommitted). Copy your archive to fixtures/replay_forecasts.json (uncommitted); the .sample.json is shape reference only and is never loaded. Live NOAA stays dark. A set-but-missing path fails the tick. One forced provenance line (force=True, survives --quiet): path, station count, min/max date — or no archive: NOAA dark, 0 entries is FIX.
• KEEP/KILL path. A full-tape run stays FIX until a real historical-forecast file covers the window. Do not treat the sample temps as history.

What's New in v1.23.13

• Replay forecast archive (SIM-5429) first cut. Loader + sample. Round 2 (1.23.14) stopped auto-loading the invented sample.

What's New in v1.23.12

• Live price path unchanged. _market_yes_price is external_price_yes or 0.5 when not in replay. Replay-only fallthrough to yes_price / current_probability.
• Gate script does not print KEEP. Unit tests are a path check. KEEP needs a full-tape simmer backtest ... --q temperature with evals > 0, entries > 0, and an honest forecast.

What's New in v1.23.11

• Backtest-as-gate (SIM-5428). Replay discovery already used q=temperature. The entry path now works under the same harness: frozen tick (SIMMER_REPLAY_NOW) for horizon/date parse, replay yes_price (not a silent 0.50), city-station fallback when the tape omits resolution_criteria (Dallas still excluded), no live NOAA/Open-Meteo (look-ahead), and preflight skipped so WALLET_UNVERIFIED cannot block SimState fills. Import is skipped under replay so it does not burn the eval budget.
• Keep / kill / fix for the real-capital path is below. 90-day Simmer P&L is still the lock. Dogfood MIN_HOURS=12 is a separate canary — this release does not change it.
• Agent glue: python scripts/run_backtest_gate.py runs the pinned replay tests and prints the verdict table.

What's New in v1.23.10

• Discovery horizon tracks SIMMER_WEATHER_MIN_HOURS_TO_RESOLVE. The existing tags=weather fetch and location keyword import were newest-first / same-day heavy. They now also query +1/+2 calendar days so a morning heartbeat with MIN_HOURS=24 still sees markets that can clear the floor. Horizon is max(MIN_HOURS, 48h) — raising the hours knob widens discovery; there is no separate horizon env. Do not lower the hours floor to "fix" morning no-fills.

What's New in v1.23.9

• Replay-compatible discovery. Under simmer backtest (SIMMER_REPLAY=1) the skill lists markets with q=temperature instead of tags=weather&status=active. Replay rejects those filters (422). Live still uses the tag.
• Fail-closed on listing failure. A failed market fetch now exits non-zero (MarketFetchError) so bundle.clean is not green on a 0-eval tick. An empty listing after a successful fetch is still a clean no-trade — that is an empty tape, not a fetch failure.
• Tape follow-up. Default HF volume slices often have 0 weather markets (vol floor ~$201k in the #368 dogfood). --min-volume / a weather-capable tape query is server-side work; this skill does not invent weather on a high-volume slice.

What's New in v1.23.7

• Min entry price. SIMMER_WEATHER_MIN_ENTRY_PRICE (default 0 = off) rejects lottery-ticket mids below the floor. SIMMER_WEATHER_ENTRY_THRESHOLD remains an upper bound only (buy when price is below it).
• Hours-to-resolve is now an env knob. SIMMER_WEATHER_MIN_HOURS_TO_RESOLVE overrides the previous hardcoded 2h time-decay safeguard. Same check, same check_context_safeguards path — raise it (e.g. 24) to skip resolve-day entries. Entry-only: exits keep the original 2h floor, so a raised value never blocks a sell. Discovery looks ahead max(this, 48h) so morning runs still see tomorrow's markets.

What's New in v1.23.3

• Event grouping now keys on event_ref (the canonical parent-event id, present on every market) instead of the legacy event_id, which SDK-imported markets historically lacked. Fixes temperature buckets silently dropping out of their event group (missing buckets when event_id came back null).

What's New in v1.21.0

• Per-market resolution source. Each market is now routed to the specific weather station Polymarket actually reads (parsed from the market's resolution_criteria field). Previously the skill used a hardcoded city → station map, which silently traded against the wrong forecast in a few cases (notably Dallas, where Polymarket resolves on Love Field / KDAL but the skill assumed DFW / KDFW). Markets that name a station the skill doesn't know are now skipped with a log line — better to skip than to trade a stale oracle. Robust to Polymarket swapping airports.
• Expanded NOAA station coverage. KLGA, KJFK, KEWR, KNYC, KORD, KMDW, KSEA, KATL, KDAL, KDFW, KMIA, KBOS, KDCA, KIAD, KPHX, KLAS, KSFO, KLAX, KDEN, KMSP, KPHL.
• Expanded international coverage. Adds Madrid, Milan, Amsterdam, Taipei to Open-Meteo routing (alongside existing Tel Aviv, Munich, London, Tokyo, Seoul, Ankara, Lucknow, Wellington).
• Requires the new ?include=resolution_criteria flag on /api/sdk/markets (live on Simmer backend 2026-05-03).

What's New in v1.20.1

• Safety rails section first. Bounding contract surfaced at the top — paper-default, --live requirement, configurable caps, server-side risk monitor, strategy-side safeguards, reversibility.
• Risk monitor framing genericized. Stop-loss / take-profit thresholds are described as configurable user settings rather than specific percentages. (See FAQ at docs.simmer.markets for current defaults — they're user-tunable in the dashboard.)
• Wallet setup link genericized. Points at docs.simmer.markets/wallets instead of a named cross-skill.

What's New in v1.20.0

• SDK 0.13.0 integration — uses SimmerClient.from_env() (auto-reads SIMMER_API_KEY, raises a clear RuntimeError with a dashboard pointer if unset). Requires simmer-sdk>=0.13.0.
• Slim per skill catalog reshape (Phase 3) — duplicated wallet-setup / changelog / decorative content removed; SKILL.md trimmed to focus on what's specific to this skill.
• Dead code removed — retired AUTOMATON_* env reads (the automaton runtime was retired 2026-04-20).

Setup

For wallet setup, see docs.simmer.markets/wallets.

Required environment:

• SIMMER_API_KEY — get from simmer.markets/dashboard → SDK tab
• WALLET_PRIVATE_KEY — Polymarket wallet private key (the SDK signs orders client-side)

Then pip install --upgrade simmer-sdk (>=0.13.0) and configure tunables below.

Configuration

Legacy env var aliases (still accepted for backwards compatibility): SIMMER_WEATHER_ENTRY, SIMMER_WEATHER_EXIT, SIMMER_WEATHER_MAX_POSITION, SIMMER_WEATHER_MAX_TRADES

Supported locations (city-name filter applied to market questions): NYC, Chicago, Seattle, Atlanta, Dallas, Miami, Austin, Houston, Denver, Tel Aviv, Munich, London, Tokyo, Seoul, Ankara, Lucknow, Wellington, Madrid, Milan, Amsterdam, Taipei, Beijing, Shanghai, Guangzhou, Shenzhen, Chengdu, Chongqing, Wuhan, Qingdao, Zhengzhou, Singapore, Kuala Lumpur, Manila, Busan, Toronto, Buenos Aires, Sao Paulo, Mexico City, Cape Town, Helsinki, Jeddah, Warsaw, Paris, Panama City. The actual oracle station is parsed per-market from resolution_criteria — see "Resolution-source routing" below.

Resolution-source routing

Polymarket weather markets carry a resolution_criteria field that names the exact station the market resolves on (e.g. "Chicago O'Hare Intl Airport Station" with wunderground.com/.../KORD). v1.21.0+ parses that text per-market and routes to the matching forecast station instead of a city default. If a market names a station the skill doesn't know, the event is skipped with a log line. Add new stations to STATION_ID_TO_NOAA (US) or INTERNATIONAL_STATION_COORDS (international) in weather_trader.py to extend coverage — PRs welcome.

SDK initialization

from simmer_sdk import SimmerClient

client = SimmerClient.from_env(venue="polymarket", live=True)

from_env() (added in simmer-sdk 0.13.0) reads SIMMER_API_KEY from the environment and raises RuntimeError with a dashboard pointer if unset. If OWS_WALLET is set, it auto-routes through the OpenClaw shared wallet.

Quick Commands

# Check account balance and positions
python scripts/status.py

# Detailed position list
python scripts/status.py --positions

API Reference:

• Base URL: https://api.simmer.markets
• Auth: Authorization: Bearer $SIMMER_API_KEY
• Portfolio: GET /api/sdk/portfolio
• Positions: GET /api/sdk/positions

Running the Skill

# Dry run (default — shows opportunities, no trades)
python weather_trader.py

# Execute real trades
python weather_trader.py --live

# With smart position sizing (uses portfolio balance)
python weather_trader.py --live --smart-sizing

# Check positions only
python weather_trader.py --positions

# View config
python weather_trader.py --config

# Disable safeguards (not recommended)
python weather_trader.py --no-safeguards

# Disable trend detection
python weather_trader.py --no-trends

# Enable volatility targeting (dynamic sizing based on market vol)
python weather_trader.py --live --smart-sizing --vol-targeting

# Quiet mode — only output on trades/errors (ideal for high-frequency runs)
python weather_trader.py --live --quiet

Backtest gate (keep / kill / fix)

90-day real-capital P&L on Simmer is the lock. This gate is the fast filter before more real capital. Dogfood's MIN_HOURS=12 canary is a separate seat — do not change it here.

Reuse the existing replay harness. Do not invent a second one.

# Pinned gate (no tape, no network, no API key)
python skills/polymarket-weather-trader/scripts/run_backtest_gate.py

# Same tests, direct
python -m pytest skills/polymarket-weather-trader/tests/test_replay_discovery.py -q

# Full-tape KEEP/KILL — build a window archive, copy it (uncommitted), then replay.
# Auto-load is fixtures/replay_forecasts.json only. .sample.json is shape-only
# and is never loaded.
python skills/polymarket-weather-trader/scripts/build_replay_forecast_archive.py \
  --start YYYY-MM-DD --end YYYY-MM-DD --out /tmp/replay_forecasts.json
cp /tmp/replay_forecasts.json \
  skills/polymarket-weather-trader/fixtures/replay_forecasts.json
# SIMMER_WEATHER_LOCATIONS defaults to "NYC" — a tape covering other cities
# (Seoul, London, Chicago, ...) will show 0 entries for every one of them
# unless you widen this list to match the tape's city mix (SIM-5484). That
# is expected scoping, not a station/forecast bug — see the note below.
#
# A plain `export SIMMER_WEATHER_LOCATIONS=...` does NOT reach the skill:
# the replay harness builds the bundle subprocess env from a strict
# allowlist (SIM-5067) that this var is not on, so it is silently stripped
# and the run measures NYC regardless. Use --set on the entrypoint instead
# — it writes config.json next to weather_trader.py, which load_config()
# reads before env vars and which survives the per-tick bundle copy.
#
# Name the tape's actual cities, not a generic US list. E.g. for an
# April tape weighted Seoul/Hong Kong/London/Shanghai/NYC/Paris: Seoul,
# London, Paris and Shanghai are mapped (INTERNATIONAL_STATION_COORDS) and
# will enter; Hong Kong has no station mapping anywhere in the skill and
# cannot enter no matter what's in this list.
simmer backtest skills/polymarket-weather-trader \
  --entrypoint weather_trader.py --t0 YYYY-MM-DD --t1 YYYY-MM-DD \
  --cadence 12h --q temperature --min-volume 0 \
  --args "--live --quiet --set locations=NYC,Chicago,Seattle,Atlanta,Miami,Austin,Houston,Denver,Seoul,London,Paris,Shanghai"
# Then read the KEEP/KILL table below, plus the per-city entry count the
# report prints — a KEEP/PIVOT read needs coverage across the mapped
# international cities, not just NYC. Optional in-process / pytest:
# export SIMMER_REPLAY_FORECASTS=/tmp/replay_forecasts.json

A green full-tape run that places 0 trades because NOAA is correctly dark is FIX (forecast plane), not KILL. Copy a real window-covering archive to fixtures/replay_forecasts.json, then re-run. The committed .sample.json is not loaded.

How It Works

Each cycle the script:

1. Fetches active weather markets from Simmer API (newest weather page plus dated queries out to max(MIN_HOURS_TO_RESOLVE, 48h))
2. Groups markets by event (each temperature day is one event)
3. Parses event names to get location and date; keeps events inside that discovery horizon
4. Fetches NOAA forecast for that location/date (under replay: archive only — never live NOAA)
5. Finds the temperature bucket that matches the forecast
6. Safeguards: Checks context for flip-flop warnings, slippage, time decay
7. Trend Detection: Looks for recent price drops (stronger buy signal)
8. Entry: If min-entry ≤ bucket price < entry threshold and safeguards pass → BUY
9. Exit: Checks open positions, sells if price > exit threshold
10. Tagging: All trades tagged with sdk:weather for tracking

Smart Sizing

With --smart-sizing, position size is calculated as:

• 5% of available USDC balance (configurable via SIMMER_WEATHER_SIZING_PCT)
• Capped at max position setting ($2.00 default)
• Falls back to fixed size if portfolio unavailable

Volatility Targeting

With --vol-targeting, position sizes are dynamically adjusted based on realized market volatility:

position_size = base_size × clamp(target_vol / realized_vol, min_alloc, max_leverage)

• High volatility: positions scale down → less risk
• Low volatility: positions scale up → more alpha capture
• Falls back to base size if insufficient price history (< 15 data points)

Safeguards

Before trading, the skill checks:

• Flip-flop warning: Skips if you've been reversing too much
• Slippage: Skips if estimated slippage > 15% (tunable via SIMMER_WEATHER_SLIPPAGE_MAX)
• Time decay: Skips if market resolves in < SIMMER_WEATHER_MIN_HOURS_TO_RESOLVE hours (default 2). Discovery looks ahead max(this, 48h) so raising the floor does not starve morning heartbeats.
• Market status: Skips if market already resolved

SIMMER_WEATHER_MIN_ENTRY_PRICE (default 0 = off) is the other side of the entry-price check, not this list — --no-safeguards does not disable it.

Disable the flip-flop / slippage / time-decay / resolved checks with --no-safeguards (not recommended).

Source Tagging

All trades are tagged with source: "sdk:weather". This means:

• Portfolio shows breakdown by strategy
• Trades tagged sdk:weather are excluded from generic copytrade sells.
• You can track weather P&L separately

Troubleshooting

"Safeguard blocked: Severe flip-flop warning" — you've been changing direction too much on this market; wait before trading again.

"Slippage too high" — market is illiquid; reduce position size or skip.

"Resolves in Xh - too soon" — market resolving sooner than SIMMER_WEATHER_MIN_HOURS_TO_RESOLVE (default 2h). Raise the env to skip resolve-day entries. Discovery already looks ahead max(MIN_HOURS, 48h), so a morning run with MIN_HOURS=24 can still take +1/+2 day markets. Same-day buckets will always fail a 24h floor before they have 24h left.

"Price $X.XX below min entry" — mid is below SIMMER_WEATHER_MIN_ENTRY_PRICE. Lottery-ticket floor; default is off (0).

"No weather markets found" — weather markets may not be active (seasonal).

simmer backtest / "Failed to fetch markets from Simmer API" — replay does not implement tags or status (422 since sdk 0.25.3/0.25.4). v1.23.9+ uses q=temperature under replay. A fetch failure now fails the tick (failed_ticks, not a clean 0-eval). If the fetch succeeds and you still see 0 weather markets, the HF volume slice likely has none — default --min-volume / top-volume selection is a tape follow-up, not a skill bug. Replay listings omit resolution_criteria; v1.23.11+ falls back to the city station table only when criteria is missing (Dallas excluded; present-but-unreadable still skips). Live NOAA is never called under replay.

A full-tape run with 0 entries and NOAA dark is FIX until a real archive covers the tape dates. Build one with scripts/build_replay_forecast_archive.py, then copy it to fixtures/replay_forecasts.json before simmer backtest (harness strips host env; that user file is copied with the bundle). Shape: { "_meta": {…}, "KLGA": { "2026-04-30": { "high": 72, "low": 50, "leads": { "1": {…}, "2": {…}, "3": {…} } } } } — top-level high/low is lead 1; _meta is ignored. Hand-built files may omit leads. fixtures/replay_forecasts.sample.json is that shape only — invented test temps, never auto-loaded. A set-but-missing path fails the tick. Do not treat a 0-entry tape as no-edge. See Backtest gate above.

"External wallet requires a pre-signed order" — WALLET_PRIVATE_KEY is not set. Fix: export WALLET_PRIVATE_KEY=0x<your-polymarket-wallet-private-key>. The SDK signs orders automatically when this env var is present — do not attempt to sign orders manually.

"Balance shows $0 but I have funds on Polygon" — Polymarket V2 (live 2026-04-28) uses pUSD (PolyUSD, 1:1 backed by USDC.e). Migrate at simmer.markets/dashboard (~30s). Full guide: docs.simmer.markets/v2-migration.

"API key invalid" — get a new key from simmer.markets/dashboard → SDK tab.