Weather prediction markets on Kalshi tie their resolution to NOAA data. A market like "Will NYC's high temp be above 80°F on Wednesday?" resolves based on NOAA's Automated Surface Observing System (ASOS) reading at Central Park. Because NOAA also publishes its own probabilistic forecasts for the same location and day, there's a natural model-vs-market comparison available.
What NOAA publishes
The US National Weather Service runs two main model families:
- GFS (Global Forecast System): NOAA's own global model, updated every 6 hours, forecasts out to 16 days
- NAM (North American Mesoscale): higher-resolution regional model for North America, updated every 6 hours
Both models are ensemble , they run multiple times with slightly perturbed starting conditions, and the spread of outcomes gives you a probability distribution. Instead of "tomorrow's high is 82°F," it's "80% chance of 80-84°F, 15% chance of 75-79°F, 5% chance of 85°F+."
How markets respond to forecasts
The typical pattern:
- Market opens a week out, price reflects historical climatology (for NYC in April: maybe 40% chance of 80°F+)
- As the event approaches, each new NOAA run updates the ensemble probability
- Trader sentiment catches up to the new forecast; price moves accordingly
- Morning-of, the market converges toward actual current conditions
The interesting gap is between steps 2 and 3 , sometimes the NOAA models say 80% but the market is still at 55% because traders are slow to update. That's the window where "mispricings" exist.
How we compute edges
Our internal system runs the NOAA ensemble every few hours for every active Kalshi weather ticker and stores the resulting probability. When we render this site, we compare:
- Market implied probability (from Kalshi's current bid-ask midpoint)
- NOAA model probability (from our latest ensemble run)
The difference is the "edge." A market priced at 40% YES with NOAA saying 85% YES has a +45 percentage-point edge.
Why edges aren't guaranteed profits
- Models are imperfect: NOAA ensemble spreads are narrow close to the event and wider far out. If the market is 5 days out and the model is 50% +/- 30%, the market is pricing in scenarios the model may not see.
- Liquidity cost: to actually trade the edge, you have to cross the bid-ask spread. A "30% edge" can shrink to a 15% edge after paying the spread and fees.
- Late-arriving information: markets sometimes move just before our model refresh. The market's "slow update" can be resolved by the time you see the edge.
What we publish
- Biggest current NOAA-vs-market edges , sorted by absolute edge size
- Per-market NOAA panel on every weather market page, showing current model probability, market price, and edge
- Daily digest emailing the top edges to subscribers
Caveats
This is research data, not trading advice. Weather models can be wrong. Markets can be wrong. The biggest edges tend to be in thinner markets where individual traders haven't moved the price toward consensus yet , but those same markets have wider spreads and less reliable pricing. Do your own homework.