Business Problem
Two kinds of uncertainty
Weather forecasting is rooted in using data to reduce the uncertainty of our climate — radar, satellites and predictive models let us quantify and visualize storms as they happen. But radar can't measure what an event will do to people, property and communities. That's a second, harder form of uncertainty: impact.
Economic damage from severe weather is costly and difficult to anticipate. Our team wanted to know whether the characteristics NOAA already records about a storm — before its full impact is known — carry enough signal to predict that damage on their own.
Research question: Can we predict whether a weather event will result in economic damage, based on recorded event characteristics in the NOAA Storm Events Database (2020–2025)?
Frequency isn't the same as risk
Share of events with recorded economic damage, by event type (2020–2025)What this shows: Hail is one of the most common events in the dataset, but rarely linked to recorded damage. Strong Wind and Tornado are far less frequent, yet show damage much more often — event type carries real signal.
Where the Data Pointed
Occurrence and cost are different problems
Before modeling anything, we compared how often each storm type occurs against how often it's actually tied to recorded economic damage. The patterns weren't the same — which is exactly why this was worth modeling in the first place.
The same split shows up in total dollars: Hurricane events are far less common than something like Thunderstorm Wind, but account for the largest total losses in the dataset by a wide margin ($50.6B vs. $1.7B).
Hypothesis
What we expected going in
Event type is the strongest predictor of recorded economic damage, with region, storm duration and seasonal timing contributing secondary signal. We expected storm characteristics available near an event's onset to beat a no-skill baseline by a meaningful margin — supporting their use for forward-looking risk assessment by emergency managers, insurers and local governments.
Methods
From raw records to a modeling table
Evaluation: Average Precision (PR-AUC) alongside ROC-AUC, since the positive class — events with recorded damage — is only about 23% of the data.
Train/test split: grouped by storm episode (EPISODE_ID), not by individual row. Events from the same storm system share nearly identical characteristics — region, season and duration all move together — so a row-level split would let the model see near-duplicate records across training and testing, inflating how good it looks. Grouping by episode keeps every event from one storm system on the same side of the split.
Both models beat the baseline
Average precision, higher is better. Source: project write-up, Appendix A.Random Forest also led on ROC-AUC (0.880 vs. 0.871) — but the two models are close enough that most of the predictive pattern is already visible in the simpler, more interpretable Logistic Regression.
What actually predicts damage
Random Forest permutation importance (decrease in validation PR-AUC when shuffled)Event type is roughly 2.4× more important than region. That's predictive weight, not proof of cause — the model finds association in historical records, not a physical mechanism.
Recommendations & Ethical Considerations
Who this affects — and where it could go wrong
Stakeholder impact
Emergency managers, local governments and insurers could use these patterns to prepare and allocate resources. The risk: over-relying on results for storm types or regions that are rare or less consistently recorded.
Data assumptions & bias
The model assumes NOAA's historical records are accurate and complete. Reporting practices vary by storm type and have changed over time, and the model assumes future events follow past patterns — a limitation we state plainly wherever these results are used.
Conclusion
Back to the real question
Our interest was never just whether a storm happens — it was what it costs once it does. Heat and Excessive Heat are among the most frequent events in the dataset, yet almost none carry a recorded economic outcome, while far rarer events like Strong Wind and Tornado are disproportionately tied to damage. As the mix of severe weather shifts, planning that leans on frequency alone risks preparing for the wrong thing.
Next steps: layering in FEMA National Risk Index exposure and vulnerability data, so predictions reflect who and what is actually in a storm's path, not just the storm's own characteristics.
Presentation
Milestone 3 walkthrough
The recorded team presentation covers this same project end to end — business problem, methods, results and conclusion — in about 12 minutes.
Code & Resources