Weather Market Trading

Using publicly available forecast data to evaluate weather prediction markets

Weather markets on Kalshi

Kalshi lists daily weather markets for major US cities, primarily covering:

• Daily high temperature — Will the high in [city] be above X degrees?
• Daily low temperature — Will the low in [city] be below X degrees?
• Precipitation — Will it rain/snow more than X inches in [city]?

These markets typically settle at the end of the relevant day, based on official weather station observations. They are structured as bracket markets — a single event (e.g., "NYC high temperature on April 5") has multiple brackets (above 55, above 60, above 65, etc.), each priced independently.

The informational landscape

Weather markets are distinctive because high-quality forecast data is freely and publicly available. Unlike political or financial markets, where information is diffuse and subjective, weather forecasting has well-calibrated probabilistic models:

National Weather Service (NWS)

The NWS provides free point forecasts, hourly forecasts, and probabilistic forecasts for every location in the United States. Their gridded forecast data includes temperature, precipitation probability, and confidence intervals. This data is funded by US taxpayers and available through the NWS API at no cost.

Open-Meteo

Open-Meteo is a free, open-source weather API that aggregates data from national weather services and global forecast models. It provides hourly and daily forecasts, ensemble model outputs (which quantify forecast uncertainty), and historical verification data.

European Centre for Medium-Range Weather Forecasts (ECMWF)

Widely considered the world's leading weather forecast model, ECMWF runs ensemble forecasts that produce probability distributions for temperature and precipitation. Some of this data is available publicly; more detailed ensemble data requires paid access.

Forecast accuracy by time horizon

The accuracy of weather forecasts degrades with time, which directly affects the potential edge in weather markets:

• 1-day ahead — High temperature forecasts have a mean absolute error (MAE) of roughly 2-3°F in most US cities. This is the sweet spot for weather trading.
• 3-day ahead — MAE increases to roughly 3-5°F. Still useful, but with wider uncertainty bands.
• 7-day ahead — MAE of 5-7°F. Forecasts are directionally useful but lack the precision needed for bracket-level edge.
• 10+ days ahead — Forecast skill drops toward climatological baselines. Individual day accuracy is limited.

Comparing forecasts to market prices

The core idea behind weather market trading is straightforward: if a forecast model assigns a 95% probability to "high above 65°F" and the market prices it at 88 cents (implying 88%), there is a potential discrepancy.

However, converting a temperature point forecast into a bracket probability requires understanding the forecast's error distribution. A forecast of "high of 72°F" does not mean the temperature will be exactly 72°F — it means the expected high is 72°F, with some probability distribution around that value.

The margin concept

A useful heuristic is the forecast margin — the distance between the forecasted value and the bracket threshold. If the forecast high is 72°F and the bracket threshold is 65°F, the margin is 7°F. Given typical 1-day forecast errors of 2-3°F, a 7°F margin provides substantial confidence.

Conversely, if the forecast high is 66°F and the bracket threshold is 65°F, the margin is only 1°F — well within the forecast error range. This trade has little informational edge regardless of the market price.

Weather-specific risks

• Forecast busts — Even the best models occasionally produce large errors, especially during regime transitions (cold fronts, tropical systems). A "forecast bust" where the actual temperature deviates 10°F+ from the forecast can wipe out many successful small trades.
• Station location — Kalshi weather markets settle based on specific weather station observations. The official airport weather station may read differently than the forecast for the broader metro area, especially for precipitation.
• Correlation across cities — Weather systems are regional. Buying NO contracts on "high above 75°F" for five nearby cities on the same day is not five independent bets — a single forecast bust affects all of them.
• Microclimate effects — Urban heat islands, elevation differences, and coastal effects can cause local observations to diverge from regional forecasts.

Ensemble models and probability

Modern weather forecasting increasingly relies on ensemble methods — running the same forecast model multiple times with slightly different initial conditions. The spread of the ensemble provides a natural probability distribution.

Gneiting and Raftery (2005) describe how ensemble forecasts can be calibrated into reliable probability distributions. If 45 of 50 ensemble members predict the high above 65°F, that suggests a roughly 90% probability — though ensemble calibration is not always perfect and varies by location and season.