Machine learning
Interpretable flood-risk modeling
A synthetic-data proof of concept exploring predictive accuracy, robustness, and interpretable triggers for parametric flood insurance.
The question
How can a flood-risk model balance predictive performance with interpretable and robust payout-trigger rules?
My contribution
Project Manager and Risk Strategist, as documented in the project’s contributor roles.
The approach
- Benchmarked linear regression, XGBoost, and sparse optimal decision trees across 20 environmental, infrastructure, land-use, and governance features.
- Explored feature importance with SHAP and evaluated sensitivity to reporting bias and shifted feature distributions.
- Simulated tiered payout rules and compared trigger thresholds for different stakeholder priorities.
What this explores
The proof of concept makes the tradeoffs between accuracy, interpretability, data quality, and stakeholder priorities explicit. It connects model evaluation to the decisions that a predictive system might support.
Scope & limitations
This project uses synthetic data. Its feature rankings and payout rules are experimental and require validation on credible real-world data before supporting insurance or disaster-response decisions.
Team: Lingyue Hao, He Jiang, Ammy Lin, Myla Simmons, and Anvita Suresh. Completed for IDS 705: Machine Learning at Duke University.
Project summary and attribution are based on the repository documentation.