The weather market-making strategy has been running for a while, but recently I noticed the drawdown doesn't seem right. Upon investigation, the issue lies in two assumptions of the probability model.

First: The calibration uses grid reanalysis data, but Polymarket settles using on-site measured data. There is a systematic bias between these two data sources, so the calibrated σ is inaccurate from the start.
Second: A more subtle point is that the model assumes unbiased forecasts. In reality, each city's forecast has directional bias—some cities tend to systematically be colder, others systematically hotter. The model doesn't account for this, leading it to repeatedly bet in the wrong direction.
For example, if a city's forecast is nearly 2°C lower than actual, the model believes "the temperature won't reach X" with high probability and buys a lot of NO positions. But in reality, the temperature is always higher than forecasted.
Initially, the instinct was to cut off cities with poor performance. After doing so, I found that one-third of them had been disabled. That made me realize: to survive, I need to cut only one-third to cover the risk, and the problem isn't with the cities but with the model.
I changed the calibration data source to use on-site observations from the same source as the settlement, and added bias correction to the probability calculations. Then I restored all the cities that had been cut—cities with high σ, where the model reduces signals on its own, so manual banning isn't necessary.