Severe Weather Probability Forecasts
Hazard-level outlooks up to 14 days ahead, updated 4 times daily
Explore All Regions
Atlanta, GA
33.75°N, 84.39°WSpot forecast data is not yet available for this run.
Spot data will appear after the next pipeline cycle completes. Check back soon.
2024 Verification
Full-year retrospective verification of Anvilcast forecasts against SPC filtered storm reports. 366 days of 2024, 40 km verification proximity.
Skill vs Lead Time
Reliability
ROC Curves
Monthly Skill
Methodology
Forecasts are verified against SPC filtered storm reports using a 40 km proximity approach (standard SPC verification). Each forecast day corresponds to the SPC convective day (12Z to 12Z). Daily probabilities are composited from 6-hour window forecasts using Pdaily = 1 − ∏(1 − Pstep).
Metrics shown: ROC AUC (discrimination), Brier Skill Score (calibrated accuracy relative to climatology), AU-PRC (precision-recall, better for rare events), CSI (threat score at 5% threshold), and SEDI (base-rate independent skill).
Per-step verification of raw XGBoost predictions at 0.25° grid resolution. Each forecast step is verified independently against SPC storm reports within ±30 minutes of the valid time.
Skill vs Forecast Step
Reliability — By Lead Day
ROC — By Lead Day
Reliability — By Time of Day
ROC — By Time of Day
Methodology
Each 6-hour forecast step is verified independently at native 0.25° grid resolution. SPC storm reports are snapped to the nearest 0.25° grid cell and matched within ±30 minutes of the step valid time. Probabilities are the raw XGBoost output before spatial or temporal scaling.
Metrics shown: ROC AUC (discrimination), Brier Skill Score (calibrated accuracy relative to climatology), AU-PRC (precision-recall, better for rare events), CSI (threat score at 5% threshold), and SEDI (base-rate independent skill).
Full-year retrospective verification with quantile-mapped WN2 inputs (v3). Same methodology as v2 but with input feature correction applied before XGB inference.
Skill vs Lead Time
Reliability
ROC Curves
Monthly Skill
Methodology
Same as v2, with the addition of per-feature quantile mapping applied to WN2 inputs before XGB inference. This corrects systematic distribution differences between WN2 forecasts and the ERA5 reanalysis data the models were trained on.
Per-step verification with quantile-mapped WN2 inputs (v3). Same methodology as v2 spot verification but with input feature correction.
Skill vs Forecast Step
Reliability — By Lead Day
ROC — By Lead Day
Reliability — By Time of Day
ROC — By Time of Day
Methodology
Same as v2 spot verification, with quantile mapping applied to WN2 inputs before XGB inference.
About Anvilcast
What is Anvilcast?
Anvilcast shows you where severe weather is heading — up to two weeks out. Individual probability maps for tornado, hail, damaging wind, and lightning, updated four times a day, with global coverage.
Most severe weather outlooks stop at Day 2 or 3. Beyond that, you are left reading model data and guessing. Anvilcast fills the gap: the same kind of hazard-level detail you expect on Day 1, extended through Day 14. If a major severe weather setup is developing 8 days from now, you will see it here first.
Anvilcast is an independent project rooted in meteorology and data science, built as a contribution to the severe weather community.
How it works
Anvilcast takes ensemble forecasts from leading AI weather models, processes the atmospheric fields associated with severe convection — instability, shear, moisture, forcing — and runs them through machine learning models trained separately for each hazard type and intensity threshold.
The system also produces conditional intensity forecasts: given that a hazard is expected, how likely is it to be significant? This lets you distinguish between a marginal day and a potential high-end event, well before traditional outlooks are issued.
The visual design of Anvilcast follows the Storm Prediction Center’s outlook format — the gold standard for communicating severe weather risk. That design language has been developed and refined over decades of operational forecasting and research. Anvilcast uses it because it is familiar, effective, and well understood by the community it serves.
Calibration
All models were trained on over 20 years of SPC storm reports and atmospheric reanalysis data, then validated against held-out seasons spanning different climate eras. The probabilities are calibrated — when the model says 15%, events occur roughly 15% of the time across many forecasts.
In retrospective testing, the system captures the key signals of high-impact events — major tornado outbreaks, significant hail episodes, and widespread damaging wind events — at lead times well beyond what is typically available.
Get in touch
Questions, feedback, bug reports, or just want to say hello: [email protected]
Find us on X, Bluesky, Instagram, and Threads. Share your forecasts and tag @anvilcast.
Frequently asked questions
What does a 15% tornado probability mean?
It means that within the highlighted area, there is approximately a 15% chance of a tornado occurring within 40 km of any given point during that forecast day. These probabilities follow the same framework used by the Storm Prediction Center in their convective outlooks, so they can be interpreted the same way.
How far ahead can you forecast severe weather?
Anvilcast produces forecasts through Day 14. Skill is highest on Days 1-3 and decreases with lead time. Longer-range forecasts are best treated as broad guidance — useful for identifying whether a significant severe weather pattern is developing, but not precise enough for specific threat details.
How often are forecasts updated?
Four times daily, at approximately 00Z, 06Z, 12Z, and 18Z. Each update incorporates the latest ensemble model run, so the forecasts evolve as new atmospheric data becomes available.
What is the difference between categorical and individual hazard maps?
The categorical map shows the overall severe weather risk level (marginal, slight, enhanced, moderate, high) — a composite of all hazards combined. The individual hazard maps (tornado, hail, wind, lightning) show the probability of each specific threat separately. Use categorical for a quick overview and individual maps for detailed threat assessment.
Are forecasts outside the United States reliable?
The underlying models were trained primarily on U.S. severe weather reports and atmospheric data. Forecasts for other regions involve extrapolation and should be treated with more caution. The atmospheric physics driving severe convection are universal, but local climatological factors and reporting biases mean that non-U.S. forecasts are experimental.
What data sources does Anvilcast use?
Forecasts are derived from ECMWF AIFS ensemble forecasts and Google WeatherNext ensemble forecasts. All source data is publicly available from the respective providers.
Important disclaimer
Do not use these forecasts to make decisions that affect life or property. For official severe weather watches, warnings, and advisories, always refer to your local or national meteorological authority.
Anvilcast is an independent, automated, experimental system. It is not affiliated with, endorsed by, or a substitute for any national weather service, government agency, or official forecasting body — including but not limited to the Storm Prediction Center (SPC), NOAA, the Bureau of Meteorology, ECMWF, or any equivalent body in any country.
Forecasts outside the United States are experimental. The underlying models were trained primarily on U.S. severe weather reports. Applying them to other regions involves significant extrapolation and should be treated with an extra degree of caution.
These forecasts are provided as-is for general information, education, and personal interest only. No warranty of any kind is made regarding their accuracy, completeness, or fitness for any purpose. The creator of Anvilcast accepts no liability for any loss, damage, or injury arising from the use of or reliance on these forecasts. Use this site entirely at your own risk.
Data Sources
© 2024-5 Google LLC, whose machine learning models were used to create the experimental data made available under the following licence terms. This data is intended for experimental modelling only and is not intended, validated, or approved for real world use.
Contains modified Copernicus Climate Change Service information. ECMWF AIFS forecast data licensed under CC BY 4.0.