Extreme Weather & Utility Risk: An Event Study Baseline

This repository contains an end-to-end data pipeline and analysis exploring how extreme weather events act as physical shocks to the U.S. Utility Sector, using the XLU ETF as a financial proxy for grid stress.

This project serves as a technical baseline for the GROWER VIP team, demonstrating the transition from administrative disaster data to actionable risk signals.

🛠 Project Structure

• scripts/build_weather_events.py: Python ETL script to pull and clean data from the FEMA OpenFEMA API.
• data/weather_events.csv: The processed event-level dataset (Aggregated by State & Event).
• 01_event_study_utility_risk.ipynb: Core analysis including "Market Shock" aggregation and volatility response modeling.

🔍 Key Data Engineering Decisions

During the development, I encountered several "real-world" data challenges that required specific handling to ensure the statistical integrity of the Event Study:

1. Filtering Logic (Meteorological vs. Other Risks)

Although the raw Southeast dataset (GA, FL, SC, NC) contained various disaster types, I explicitly excluded Fire events.

• Reasoning: In the Southeast, grid-impacting disasters are predominantly hydro-meteorological (Hurricanes, Tropical Storms). Fire declarations in this region often relate to smaller-scale management grants (FMAG) which lack the systemic "market-moving" impact of a major hurricane.

2. Handling "Administrative Clones"

FEMA data often records multiple entries for the same physical event in the same location. For example, in South Carolina (SC), Hurricane Ian appeared twice: an Emergency Declaration (EM) in September followed by a Major Disaster Declaration (DR) in November.

• Solution: I implemented a de-duplication layer to ensure each "onset" of a disaster is only counted once per region, preventing signal inflation.

3. The "Market Shock" Aggregation

A major hurricane (e.g., Hurricane Helene) often sweeps across multiple states, triggering separate FEMA disaster numbers for FL, GA, and NC.

• Strategy: Financial markets are forward-looking and typically "price in" the shock upon the first landfall. To avoid diluting the volatility signal, I aggregated these state-level records into a single Market Shock based on a normalized event name and date.

4. Severity Proxies & Data Limitations

• FIPS Coding Issues: Initially, I attempted to use fipsCountyCode for high-resolution mapping, but found many records contained "0" or missing values in the API response. I switched to a robust count of unique designatedArea to maintain consistency.
• Dynamic Filtering: Instead of hard-filtering small events during ETL, I kept the full dataset and applied a threshold (e.g., affected_areas_count > 5) within the Notebook to allow for sensitivity testing without losing raw information.

📈 Analysis Methodology

By constructing a Composite Risk Score (standardizing 30-day rolling volatility and maximum drawdown), the analysis reveals a visible "Risk Spike" in the XLU sector following the onset (T=0) of major weather shocks.

🚀 Future Work (Connection to GROWER)

• From Financial Proxies to Physical Truth: Replace XLU market data with real-time Outage Data (duration/frequency) to build a direct resilience model.
• Spatial Precision: Moving from state-level aggregation to high-resolution GIS analysis, addressing the fact that a high county count does not always correlate with the total impacted geographic area.
• Resource Optimization: Use the Risk Score framework to develop predictive "Readiness Triggers" for emergency response and resource pre-positioning.