Independent research, 2026
Forecasting weather prediction markets
A temperature forecasting model scored against 4,433 Kalshi markets.
In [12]: brier(contested)
Lower is better. n = 1,216 contested of 4,433 settled markets.
The model turns day-ahead temperature forecasts for 7 U.S. cities into probabilities and compares them with Kalshi market prices.
The first paper backtest returned 7.7%. The Brier score, which measures calibration, showed the market was much better, so I did not treat the return as evidence the model works.
What I found
Across 4,433 settled markets, on the 1,216 where the market was unsure, the market scores 0.2114 and my model scores 0.3047. Paper return is -3.37% on 212 picks. My rule for real money has not been met, so none has been traded.
| Settled markets scored | 4,433 |
|---|---|
| Contested markets | 1,216 |
| Brier, market | 0.2114 |
| Brier, my model | 0.3047 |
| Paper return | -3.37% on 212 picks |
Lower Brier is better. Full write-up and code are public.
Next case study: Segmenting online store customers in SQL