Weather research / United States
Weather models, measured against reality.
Editable.ai develops regional forecasts and severe-weather hindcasts, measures them against observed conditions, and turns weather records into case-specific evidence packets.
Current work
Three programs. One standard of evidence.
Forecast skill, severe-weather risk, and case-specific evidence are different problems. Each is evaluated against the records and metrics appropriate to the task.
Regional weather forecasting
A 2024 benchmark across temperature, dewpoint, wind, and precipitation at +6 to +24 hour lead times.
Severe-weather hindcasts
Peril-specific models for hail, thunderstorm wind, and tornado probability, evaluated against observed spatial patterns.
Traceable weather reports
Case-specific radar, report, and station records organized into a traceable timeline, score, coverage review, and evidence packet.
Forecast benchmark
Six variables. Four lead times. One matched test set.
All curves use the same 702 available initializations from 2024. Lower RMSE is better.
Figure 01 / Forecast error
Click to enlargeHow we work
Start with the test.
A convincing map is not enough. The evaluation protocol is part of the model.
-
01
Fix the comparison
Match initialization times, lead times, spatial support, and observations before ranking systems.
-
02
Use task-specific metrics
RMSE works for continuous fields; rare tornado events require probability and ranking diagnostics.
-
03
Publish the limits
Resolution changes, reporting bias, calibration gaps, and non-operational use are stated beside the result.
Contact