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-2026-2601
version: 1
kind: certified
reason: Certified into the record
timestamp: 2026-11-12T08:01:02.080804Z
status: published
lens: g_space
sector: climate
headline: Evaluating the Skill of Medium-Range Machine Learning–Based Convective Hazard Forecasts Derived from Convection-Allowing Ensembles
dek: Abstract We assess the skill of daily machine learning (ML)-based probabilistic convective hazard predictions generated from two experimental medium-range (i.e., extending to day 8) convection-allowing ensemble (CAE) systems to determine if such predictions are skillful over the CONUS. CAE forecasts were generated in real time for events in spring 2023 and 2024, and convective hazard predictions were produced using neural networks, trained with CAE output and observed storm reports (i.e., reports of tornadoes, h…
gain_title: In a study of real-time forecasts from spring 2023 and 2024 over the continental US, machine learning hazard forecasts built from convection-allowing ensembles outperformed both updraft-helicity forecasts and climatology through day 8.
problem_title: (none)
trace_subject: (none)
gain_reading: In a study of real-time forecasts from spring 2023 and 2024 over the continental US, machine learning hazard forecasts built from convection-allowing ensembles outperformed both updraft-helicity forecasts and climatology through day 8.
gain_evidence: The CAE ML-based hazard forecasts outperformed both the UH-based forecasts and climatology through day 8.
problem_reading: (none)
problem_evidence: (none)
quick_read: Researchers tested machine learning forecasts of tornadoes, hail and convective wind gusts across the continental US, built from two experimental convection-allowing ensemble systems that ran in real time during spring 2023 and 2024. Neural networks were trained on ensemble output and observed storm reports. The ML hazard forecasts outperformed both updraft-helicity forecasts and climatology through day 8.

Against ML forecasts built from NOAA's operational Global Ensemble Forecast System (GEFS), the convection-allowing versions were skillful for days 1 to 4. Beyond that, differences were small and not statistically significant, so the extra benefit is limited to the first few days.

All forecasts did better in spring 2024 than in spring 2023, likely because stronger large-scale forcing made storms more predictable. That stronger forcing shrank the advantage over GEFS: the benefit reached days 3 to 4 in 2024 but extended through day 5 in 2023.
limitation: 
tag: Evidence-backed gain
key_points: Forecasts covered the continental US and extended to day 8, longer than the day or two typical of the most advanced storm-scale models. | Two experimental convection-allowing ensemble systems ran in real time for spring 2023 and 2024 events. | Neural networks were trained on ensemble output and observed reports of tornadoes, hail and convective wind gusts. | The ML forecasts beat both updraft-helicity forecasts and climatology through day 8. | Against GEFS-based ML forecasts, the ensemble-based forecasts were skillful for days 1 to 4; later differences were not statistically significant. | Spring 2024 forecasts were more skillful than spring 2023, likely due to stronger large-scale forcing. | Interpretability tests found storm-scale predictors weighted more for days 1 to 2 and environmental predictors for days 3 to 5.
rundown: The work asks whether storm-scale ensemble forecasts, usually limited to a day or two ahead, can guide forecasters at longer lead times. Researchers generated real-time ensemble forecasts in spring 2023 and 2024 and used neural networks to turn them into probabilities of tornadoes, hail and convectively induced wind gusts.

Skill was measured against climatology, updraft-helicity forecasts from the same ensembles, and ML forecasts from NOAA's operational GEFS. The ensemble-based ML forecasts beat climatology and updraft helicity through day 8, and beat GEFS-based forecasts on days 1 to 4.

Year-to-year weather regimes affected the comparisons. Stronger large-scale forcing in 2024 improved all forecasts but reduced the ensemble advantage over GEFS, which extended to days 3 to 4 in 2024 and through day 5 in 2023. The authors say this matters when comparing weather prediction systems.

Interpretability experiments showed the networks relied more on storm-scale predictors for days 1 to 2 and on environmental predictors for days 3 to 5.
sources:
- peer_reviewed | Weather and Forecasting | https://doi.org/10.1175/waf-d-25-0243.1 | 2026-07-23
prev: 0000000000000000000000000000000000000000000000000000000000000000

SHA-256 hash

9084e3d3af6b496ad46880c7dd4e36b7a7347bb19c25df745950671c6e78c908

Previous version’s hash

0000000000000000000000000000000000000000000000000000000000000000