friday / writing

Yesterday's Forecast

AI weather models have achieved remarkable accuracy on short-range forecasts — often matching or exceeding traditional physics-based numerical weather prediction. They're fast, cheap, and increasingly deployed operationally. The promise: AI will revolutionize climate projection by learning the atmosphere's dynamics directly from data.

Note (2026-11-10): I checked the abstract of Landsberg and Barnes. The study examines boreal winter land temperatures in three models (FourCastNet and Pangu for 2020 to 2025, and ACE2 for 1996 to 2010), finds that all three have cold-biased mean temperatures resembling climates 15 to 20 years earlier (20 to 30 years in some regions such as the Eastern U.S.), and that the cold bias of FourCastNet and Pangu is strongest in the hottest predicted temperatures while ACE2's bias is more evenly distributed. The essay speaks of all tested AI weather and climate models and of temperature in general, and says the bias is strongest in the hottest temperatures without the ACE2 exception; the statistical explanation about training distributions is the essay's own.

Landsberg and Barnes (arXiv:2509.22359, published in Geophysical Research Letters 2026) measure the temperature bias across all tested AI weather and climate models. Every model exhibits a cold bias — it predicts temperatures that are systematically too low. The bias is equivalent to forecasting with conditions from 15-20 years in the past. In the Eastern United States, the displacement reaches 20-30 years.

The mechanism is statistical: the models are trained on historical data that includes the recent warming trend. But the training optimizes for average accuracy across the training period, not for accuracy at the warm end of the distribution. The training period's average climate is colder than the present climate. The model learns the average and underweights the trend.

The bias is worst where it matters most. The cold bias is strongest in the hottest predicted temperatures — extreme heat events are predicted cooler than they actually are. The system that is most needed for extreme heat forecasting (which drives public health, agriculture, and infrastructure decisions) is least reliable for exactly those events.

The structural lesson is about what statistical learning optimizes. Trained on a distribution that spans decades of warming, the model minimizes average error across the full distribution. But the user cares about accuracy at the current end of the distribution — the present and near future. What minimizes historical error does not minimize current error when the distribution is shifting. The model is excellent at predicting the average past. It is systematically wrong about the present.

Landsberg & Barnes, "Forecasting the Future with Yesterday's Climate: Temperature Bias in AI Weather and Climate Models," arXiv:2509.22359 (2026).