Improving extreme-rainfall forecasts
Leads research using machine learning to diagnose and reduce displacement and intensity errors in High-Resolution Ensemble Forecast extreme-precipitation guidance.
Independent evaluation
Formal Test-Set Model Skill
These figures summarize performance on the independent 2024–2025 test set. They are separate from the running verification of forecasts issued by this real-time website. No single metric fully describes forecast quality.
Loading finalized model-skill outputs…
Day-level risk occurrence
Risk-occurrence contingency by product
Each of the 45 test days contributes one outcome per product and threshold. A hit means the forecast and Practically Perfect both contained the selected risk; a miss means only Practically Perfect did; a false alarm means only the forecast did; and a correct negative means neither did.
| Product | Hits | Misses | False alarms | Correct negatives | CSI | ETS | Forecast risk days | PP risk days |
|---|
Issued-forecast record
Running Real-Time Verification
Running verification summarizes forecasts produced by this experimental real-time pipeline after their valid periods have ended. Statistics may change substantially when only a small number of cases are available.
Loading issued-forecast verification…
| Product | Selected metric | Forecast risk days | PP risk days | Verified days | Hits | Misses | False alarms | Correct negatives | Day CSI | Day ETS |
|---|
Model transparency
Global SHAP Explainability
SHAP values estimate how strongly each predictor shifts the model prediction away from its average prediction. Positive and negative effects are relative to the fitted model and do not establish causation.
Loading finalized SHAP figures…
SHAP describes the fitted model, not physical causality.
Correlated predictors can share or redistribute apparent importance.
Dependence patterns may reflect interactions and may not represent future distribution shifts.
About XGBFFP
XGBoosted Flash Flood Predictions
XGBFFP is experimental XGBoost guidance for flash-flood potential associated with mesoscale convective systems. It combines multiple neighborhood-radius models with WPC ERO context, post-event verification, environmental diagnostics, and model-transparency products.
Forecast guidance
The map displays 40-, 60-, 75-, and 100-km neighborhood configurations, a beta/testing 60-km V2 member, an ensemble mean, and the official WPC ERO reference.
Three evaluation populations
Formal independent 2024–2025 test-set results, running verification of issued forecasts, and individual event context are intentionally kept separate.
Important limitation
Experimental machine-learning guidance. Not an official NWS forecast, watch, or warning.
Researcher · developer · forecaster
About the Creator
Meet the creator
Learn about Tyreek J. Frazier’s meteorology research, machine-learning development, operational testbed experience, publications, presentations, teaching, and leadership.
About the Creator
Tyreek J. Frazier
Meteorologist · Machine Learning Scientist · Scientific Software Developer
Tyreek develops data-driven tools for high-impact weather forecasting and decision support. His work connects meteorology, hydrology, statistics, and software engineering—from multi-year atmospheric datasets and probabilistic models to real-time forecast products built for rapid interpretation.
6M+Gridded samples processed
2WPC FFaIR experiments
4Listed research papers
5Selected presentations
01
Core work
Research and technical impact
Reproducible, large-scale pipelines
Designed Python workflows that ingest, quality-control, regrid, join, and analyze more than six million samples spanning NWP, radar precipitation, Flash Flood Guidance, GRIB2, NetCDF, HDF5, and geospatial data.
Models built for rigorous evaluation
Developed XGBoost, random-forest, and logistic-regression systems with engineered predictors, hyperparameter optimization, independent holdouts, and statistical performance comparisons.
Forecasts from artifacts to interface
Built automated real-time flash-flood prediction and next-day verification workflows that retrieve operational data and publish interactive guidance for fast decision support.
02
Experience
Science in research, operations, and education
Graduate Research Assistant · Iowa State University
Leads interdisciplinary precipitation and flash-flood ML research; evaluates reliability, Brier and Ranked Probability Skill Scores, CSI, ETS, POD, FAR, object-based errors, and statistical significance.
Flash Flood and Intense Rainfall Experiment · NOAA WPC
Selected for the 2024 in-person and 2025 virtual operational testbeds. Produced daily experimental forecasts with self-developed ML tools, gathered forecaster feedback, and presented uncertainty and model limitations in a time-sensitive environment.
Co-Mentor · Undergraduate Research
Guides a senior meteorology student investigating how near-storm environments affect ML postprocessor performance during mesoscale convective systems.
Teaching Assistant · Iowa State University
Led lectures and laboratory sessions in undergraduate synoptic and mesoscale meteorology while supporting assessment and student learning.
Summer Intern · Federal Aviation Administration
Evaluated high-altitude ASOS ceilometer capabilities and worked with FAA, NCAR, NOAA, and National Weather Service scientists on aviation-weather applications.
03
Education
Academic foundation
04
Technical range
Tools and domains
PythonBashLinuxGit/GitHubXGBoostscikit-learnTensorFlowPyTorchNumPypandasxarraySciPyMatplotlibCartopyGRIB2NetCDFHDF5GISForecast verificationEnsemble forecasting
05
Scholarship
Publications and manuscripts
Gallus, W. A., Jr., et al. “A Climatology of Errors in HREF MCS Precipitation Objects.” Water, 17(15), 2168.
Frazier, T. J., et al. “A Machine Learning Postprocessor to Mitigate Ensemble Quantitative Precipitation Forecast Displacement Errors.” Journal of Artificial Intelligence for the Earth Systems.
Frazier, T. J., et al. “Using Machine Learning to Improve Intensity Errors in Ensemble Quantitative Precipitation Forecasts.”
Frazier, T. J., et al. “Probabilistic Machine Learning Forecasts of Extreme Rainfall from Mesoscale Convective Systems.”
06
Communication
Selected presentations
Using Machine Learning to Improve Forecasts of Extreme RainfallOral · Earth, Atmosphere, and Climate Graduate Student Symposium
A Machine Learning Postprocessor to Mitigate Ensemble QPF Displacement and Intensity ErrorsOral · NOAA WPC FFaIR Experiment
Evaluation of a Machine Learning Post-Processor to Mitigate Quantitative Precipitation Forecast ErrorsOral · AMS 105th Annual Meeting
Correcting MCS QPF Errors of Intensity, Timing, and Location Using Machine LearningPoster · AMS 104th Annual Meeting
Aviation Benefits of Allowing ASOS Ceilometers to Gather Data up to 25,000 ftPoster · AMS 102nd Annual Meeting