THE METHODOLOGY
Most weather traders look at a single forecast and guess. We do not guess. We run every trade decision through 62 independent weather model simulations and only act when the math is overwhelming.
Forecasts 1 – 31
GFS Physical Ensemble
The same model powering every weather app on your phone — except NOAA runs 31 slightly different versions to capture uncertainty. We pull all 31 members through the Open-Meteo API and compute probability distributions for temperature highs and lows in every Kalshi weather market city.
Forecasts 32 – 62
AIGEFS — NOAA Project EAGLE
NOAA's brand-new AI weather model, launched December 2025. Built on Google DeepMind's GraphCast architecture, retrained on NOAA's own data. We pull all 31 AI ensemble members directly from NOAA's AWS S3 bucket, parse the GRIB2 files, and extract 2-meter temperatures at each city coordinate.
Combined: HGEFS
62-Member Grand Ensemble
When both ensembles agree on an outcome, we trade with high confidence. When they disagree, the probability naturally moves toward 50/50 and the bot sits out. No forced trades. No guessing. NOAA's own research shows HGEFS consistently outperforms both the physics-only and AI-only systems individually.
[ INFO ] Fetching GFS ensemble members 1–31 via Open-Meteo API...
[ INFO ] Fetching AIGEFS ensemble members 1–31 from NOAA S3 (GRIB2)...
[ INFO ] Parsing 2m temperature grids for 14 Kalshi market cities...
[ CALC ] Building probability distributions — 62 members...
[ COMP ] ENSEMBLE_COMPARE ticker=KXHIGHTDC gefs_prob=0.82 aigefs_prob=0.79 combined=0.81 agreement=True
[ GATE ] EDGE_SCORE ticker=KXHIGHTDC spread=0.91 volume=0.88 imbalance=0.76 mispricing=0.85 composite=0.85
[ EXEC ] ELIGIBLE ticker=KXHIGHTDC side=YES price=0.72 edge_score=0.85 → PLACING ORDER
[ INFO ] Scanning remaining 47 tickers... 47 rejected (threshold not met).
TRADE GATING
Every potential trade is scored on four dimensions before a single dollar is risked. Only trades that pass all four filters with a composite score above threshold get executed. In practice, the bot rejects 95%+ of opportunities. Patience is the edge.
How tight is the bid-ask spread? Tight spreads mean real liquidity, not phantom orders sitting in an empty book.
Is there actual trading activity, or is this a dead market? The bot requires minimum volume thresholds before entering.
Which side has more demand? If we are buying YES and the order book is stacked with YES bids, the market agrees with us.
How far is the market price from our 62-member ensemble probability? This is the core edge: the gap between what the market thinks and what the weather models say.
Example composite score calculation
spread=0.91 × 0.30 = 0.273
volume=0.88 × 0.20 = 0.176
imbal =0.76 × 0.25 = 0.190
misprice=0.85 × 0.25 = 0.213
─────────
composite = 0.852 ✓ ABOVE THRESHOLD → EXECUTEROADMAP
The 62-member ensemble is live and being battle-tested now. After sufficient track record data is collected, the v2.0 release will ship the following. Current buyers get the update free.
HGEFS 62-member grand ensemble
Currently in live battle-testing.
FRED API integration
Real-time energy price signals as leading indicators for weather-driven demand.
Multi-platform execution
Expanded market coverage beyond the initial city set.
Dynamic confidence scaling
Enhanced probability model with adaptive thresholds based on market conditions.