If you've spent any time in Kalshi weather markets, you've heard people reference "the models." Weather models are the foundation of temperature trading. Most traders — especially those coming from sports betting, poker, or DFS — have no idea what these models are, how they differ, or which ones matter for Kalshi markets specifically.
A weather model is a massive computer simulation of the atmosphere. It takes current observations — temperature, pressure, wind, humidity from weather stations, satellites, radar, weather balloons, and aircraft — and runs them through physics equations to predict what the atmosphere will do next. Different models use different approaches: different physics equations, different resolutions, different data sources, and different update schedules. That's why they disagree. And when they disagree, that's where trading opportunities live.
NWS is the source most Kalshi traders anchor to — it's the official temperature settlement authority, so the market tends to price around it. NWS isn't actually a model. It's a human forecast. NWS meteorologists look at multiple models (GFS, ECMWF, NAM, HRRR, and others), combine them with local knowledge and experience, and issue a single number. That's both its strength and its weakness. NWS updates their daily high/low forecast a few times per day. On a morning when conditions are shifting fast, the NWS forecast from 5 AM might already be stale by 10 AM.
HRRR is NOAA's highest-resolution operational model for the continental US at 3 kilometers (~1.9 miles between grid points). It updates every hour, with radar data assimilated every 15 minutes. Forecast range: 18 hours on hourly runs; 48 hours on the 00z/06z/12z/18z extended runs.
HRRR is the single most valuable model for same-day temperature markets. At 10 AM, the HRRR run from 9 AM has already incorporated the morning's actual temperatures, cloud cover, and wind patterns. It's not predicting from scratch — it's projecting forward from what's actually happening. GFS and ECMWF, by contrast, run only four times a day and can be 3–6 hours stale by the time you're making a trading decision.
HRRR's advantage fades with time. At 0–12 hours, it's the best US model for surface temperature. At 24–36 hours (next-day markets), the gap between HRRR and other models narrows significantly.
GFS is NOAA's primary global weather model at 13 kilometers resolution, updating four times daily (00z, 06z, 12z, 18z) with a 16-day forecast range. GFS provides the big picture and is good at capturing large-scale weather patterns: cold fronts, high pressure systems, major storm tracks.
GFS also runs a 31-member ensemble (GEFS), which means it runs the same forecast 31 times with slightly different starting conditions. The spread between those 31 runs tells you how uncertain the forecast is.
ECMWF's reputation is earned for multi-day forecasts — it consistently outperforms GFS in global verification scores. Its 51-member ensemble is considered the gold standard for probabilistic forecasting. For same-day US temperature markets, though, ECMWF's advantages don't apply — it only updates twice daily, its resolution is coarser than HRRR for the US domain, and it doesn't assimilate US radar data the way HRRR does.
For same-day markets: HRRR's hourly updates are your primary source. Check what HRRR's latest run says for your city. Compare it to the NWS forecast. If HRRR diverges from NWS by 2°F or more, that's a potential signal. The market is likely anchored to NWS. If HRRR is right, the market is mispriced.
For next-day markets: look for model consensus — or the lack of it. When GFS and ECMWF disagree meaningfully on a next-day high, that spread is a signal: the atmosphere is in a hard-to-predict state. Markets priced with high confidence in that environment are likely mispriced.
Confusing the forecast with settlement. The models predict temperature. Kalshi settles on what the ASOS station records — these are not the same thing. Treating the NWS forecast as gospel — NWS is most likely already priced into the market. Ignoring model age — always check the run timestamp before acting on model data.