Kalshi runs a deep set of weather markets - temperature thresholds in major cities, snowfall totals, hurricane formation. They settle against NWS/NOAA data. We build bots that combine ensemble forecast inputs (ECMWF, GFS), historical analog matching, and your own conviction overlay. Technically: ingest NOAA's public forecast products + run an ensemble probability model, compare to Kalshi mid-prices, fire when our model and the market disagree by >X.
Weather is one of the few markets where there's a public, audit-able truth source.
Pull ECMWF + GFS ensemble forecasts on schedule. Each model run gives a probability distribution for the threshold.
Blend ensemble forecasts with historical analog matching (similar days in last 20 years). Output: a probability for the YES side of the Kalshi contract.
If your model says 70% and the Kalshi market says 50%, that's a 20¢ edge. Fire - sized against orderbook depth and your conviction in the model.
Pure quant signal. No 'I think it'll be cold' overlays unless you explicitly add a conviction weight.
# Weather trading config - Boston JAN 32°F example market: "HIGHKBOS-26JAN15-T32" # Will Boston high hit 32°F on Jan 15? model: inputs: - "gfs-ensemble" - "ecmwf-ensemble" - "historical-analog-10yr" blend: "weighted-mean" uncertainty_margin_pct: 5 # model error budget trigger: min_edge_cents: 5 # model vs market gap min_hours_to_resolution: 24 # avoid noisy near-term resolutions position: size_usdc_per_cent_edge: 100 # 10¢ edge = $1,000 position max_per_market_usdc: 3_000
5-day forecasts settle close to climatology. 1-day forecasts are increasingly accurate. We size positions by the forecast horizon.
On a 3¢ edge, you're losing money even if you're right. We hard-code a minimum-edge threshold that accounts for round-trip fee + slippage.
ECMWF and GFS ensemble spread is the honest signal. A 70% YES on the model with wide spread isn't a 70% YES - it's 'somewhere between 60-80'. We size by the lower bound.
If we don't have 2+ years of similar markets to backtest the model, we won't deploy it. Period.
Where the model has signal vs. noise.
ECMWF/GFS are very accurate at this horizon. Best Sharpe markets in our backtests.
Climatology-anchored. Mean-reverting. Lower-frequency but stable edges.
High volatility, episodic. Profitable when you can monitor forecast updates closely.
One city / one weather variable, fully modeled and deployed.
Portfolio of weather markets with cross-market correlation handling.
Tell us which Kalshi weather markets interest you. We'll come back with a backtest excerpt and a build estimate.