HOW IT WORKS
How WeatherEdge v2 Finds Edge
The complete pipeline behind every trade signal. WeatherEdge v2 automates the entire Polymarket weather trading pipeline using a 4-model ensemble consensus, nowcasting, regime classification, and passive CLOB execution.
COVERAGE
58 cities with ICAO airport coordinates across the US and key international markets. Each city is mapped to the nearest airport METAR/TAF station for real surface observations. Station confidence tiers (0–3) gate whether a market is tradeable — no trade below tier 2.
FORECAST MODELS
4 global NWP models provide the probability backbone:
• GFS — 31-member ensemble (NCEP)
• ECMWF IFS — 51-member ensemble (European Centre)
• UKMO — UK Met Office deterministic (Met Office)
• MeteoFrance — Arpege/Arome (Météo-France)
All accessed via Open-Meteo API. Consensus blend produces calibrated probability with weighted uncertainty across 113 total ensemble members.
EDGE METRIC
The primary gate metric is theoretical_full_ev: SUM[size × (p × (1/P − 1) − (1−p))] − cost_proxy. This replaces the simple edge percentage from v1. Only positions with positive theo EV after cost estimation pass the gate. Sizing uses ¼-Kelly fraction by default.
NOWCASTING
When a market is ≤ 24 hours from resolution, the nowcasting engine activates. It runs an AR(1) residual model on the latest NWS observations with 5,000 Monte Carlo samples. Observation anchoring uses a half-life of 2h near the diurnal peak and 4h otherwise, blending real surface data with model forecasts for high-confidence last-mile predictions.
DIURNAL STAGING & REGIME
Markets are staged into pre-peak, near-peak, and post-peak windows based on latitude-derived solar timing. A 4-feature regime classifier (ensemble spread, wind shift, cloud cover, precip) tags each market as frontal, marine, convective, or clear-sky — shaping the probability distribution accordingly.
EXECUTION
All trades use passive CLOB limit orders — never market orders. Time-in-book ranges 30–90 seconds with up to 3 reprice attempts. Position sizing is capped at 15–25% of visible orderbook depth. Each trade gets a frozen model snapshot via a unique trade_group_id so post-trade analysis can reconstruct the exact signal state.