Project

Can AI beat the KMA?

Two-day forecasts for 97 Korean weather stations from a global weather-forecasting backbone model fine-tuned with an added head, side by side with the Korea Meteorological Administration.

More accurate than the KMA in the test period

	Temp MAE	Humidity MAE	Wind MAE	Rain F1
Our AI	0.84℃	4.5%	0.58m/s	0.41
KMA	1.08℃	6.5%	0.96m/s	0.26

Total development cost: about 20,000 KRW (roughly 15 USD) · just a Colab subscription and Claude tokens

33 issuances from July to October 12, 2026 · 97 ASOS stations · 48 hours after issuance. A limited period; it does not show wins in every season or major event. KMA values are 5 km grid forecasts; even with its per-station bias removed (1.02 °C) ours stayed ahead. Counting drizzle as rain (0.1 mm, light rain), our F1 was 0.51 against the KMA’s 0.37.

Our AI vs the KMA forecast

Loading the forecast…

Last 7 days

Past issuances checked against ASOS observations. Bold marks the winner; lower MAE and higher F1 are better. Rain F1 counts 1 mm or more; light-rain F1 counts 0.1 mm or more, drizzle included.