I run a small tool that prices Kalshi's daily temperature markets ("Will the low in Atlanta tomorrow be 49° to 50°?") against a weather forecast. For each bracket it computes a forecast probability and puts it next to the market's price. Where the two disagree, that's the interesting row.
On November 11 one row disagreed about as much as a row can:
Atlanta, daily low, bracket 49° to 50° market: 90% my model: 0% observed low so far: 48.2°F
Enter fullscreen mode
Exit fullscreen mode
The logic looked airtight. The low for the day had already reached 48.2°F. A low can only go down from there, so it can't end up between 49 and 50. Zero percent.
The market was pricing 90% with real money behind it. When a model says "impossible" and the market says "very likely," the model is usually the one that's wrong. Here it was.
Where 48.2 came from
The observations come from the free National Weather Service API ( api.weather.gov/stations/{id}/observations ).…