The exact text we saved

This is the version exactly as it was saved, character for character. The SHA-256 hash below is a 64-character code worked out from that text. Run the same text through SHA-256 and you get the same code; change a single letter and the code comes out completely different. Each version also stores the previous version’s hash, which chains them together so no version can be swapped out quietly.

TRUVACE RECORD VERSION
record: TRV-2026-2543
version: 1
kind: certified
reason: Certified into the record
timestamp: 2026-11-12T07:51:06.210657Z
status: published
lens: p_space
sector: science
headline: The Impact of Target Dataset Choice on Machine Learning Prediction of Severe Hail, Wind, and Tornadoes
dek: Abstract Supervised machine learning (ML) methods implicitly assume that their target dataset is error free. For severe weather prediction, point-based observed Storm Data storm reports that have been expanded in space and time are commonly used targets. However, other target datasets could potentially better represent severe weather occurrence, especially for situations requiring more precision. This study analyzes 1-h probabilistic severe weather hazard forecasts made at lead times from 0 to 3 h from random fo…
gain_title: (none)
problem_title: In a comparison of random forest models, severe weather forecasts trained on National Weather Service warnings scored notably worse when checked against storm reports than report-trained models checked against warnings.
trace_subject: (none)
gain_reading: (none)
gain_evidence: (none)
problem_reading: In a comparison of random forest models, severe weather forecasts trained on National Weather Service warnings scored notably worse when checked against storm reports than report-trained models checked against warnings.
problem_evidence: warnings-trained RF forecasts verified on reports have notably worse objective verification than report-trained RF forecasts verified on warnings
quick_read: Researchers compared two sets of random forest models that forecast severe hail, wind and tornado hazards one hour at a time, at lead times from 0 to 3 hours. One set was trained on storm reports expanded in space and time; the other was trained on hazard-specific National Weather Service warnings.

Forecast probabilities came out drastically different depending on the target. Warnings-trained models gave higher probabilities after 0000 UTC and lower ones before it than report-trained models, and their forecasts were sharper and more spatially focused, especially for hail and wind.

Which model looked better depended on the yardstick. Models trained and verified on warnings scored better, but warnings-trained models verified on reports scored notably worse than report-trained models verified on warnings. The authors say developers and users of ML forecasts need to weigh the target dataset when building models and reading results.
limitation: Verification results depend on which dataset is used as truth, since the two target datasets have different spatial and temporal climatologies, so neither model clearly outperforms the other overall.
tag: Evidence-backed problem
key_points: All models used predictors from the Warn-on-Forecast System, ProbSevere version 3 and the tornado probability algorithm; only the training target differed. | Both target datasets were built to have the same overall climatological frequency for each hazard but different spatial and temporal patterns. | Models tended to reproduce the climatology of their target: warnings-trained models gave higher probabilities after 0000 UTC and lower ones before it. | Warnings-trained forecasts were sharper and more spatially focused, especially for severe hail and wind. | Models trained and verified on warnings scored better than report-trained models on the same verification.
rundown: The study tested a common assumption in supervised machine learning, that the target dataset is error free. For severe weather, the usual target is point-based Storm Data reports spread over surrounding space and time, which may not capture precisely where and when hazards occurred.

To test alternatives, the authors trained random forests on either those expanded reports or hazard-specific National Weather Service warnings, holding everything else constant, and produced 1-hour probabilistic forecasts at lead times of 0 to 3 hours.

The cross-verification result is the key caution: warnings-trained forecasts checked against reports did notably worse than report-trained forecasts checked against warnings. The authors conclude that target dataset selection matters during both model development and evaluation.
sources:
- peer_reviewed | Artificial Intelligence for the Earth Systems | https://doi.org/10.1175/aies-d-26-0004.1 | 2026-09-28
prev: 0000000000000000000000000000000000000000000000000000000000000000

SHA-256 hash

1c4557399f410bc564314ba9d57457ba0a42150d6a212ed95367200c7bbec580

Previous version’s hash

0000000000000000000000000000000000000000000000000000000000000000