This project analyzes global natural disaster data to predict event intensity. Originally aimed at predicting fatalities (Regression), the project evolved to classify disaster strength due to significant class imbalance in mortality data.
Key Features & Engineering
K-Means Clustering: Segmented disasters into 5 distinct geographic risk zones based on location and magnitude.
Physics-Based Ratio: Created a mag_depth_ratio to evaluate the surface impact of deep vs. shallow events.
Temporal Analysis: Extracted year and month to account for seasonal disaster patterns.
Models & Results
Regression
Focused on predicting magnitude/deaths.
Used Random Forest to handle non-linear relationships.
Classification
Task: Predicted if an event is "Strong" (above median magnitude).
Winning Model: Random Forest Classifier.
Accuracy: ~99%.
Strategy: Prioritized Recall to ensure high-intensity events are not missed (minimizing False Negatives).
Business Logic
As part of a data-driven approach to risk management, the model focuses on physical intensity as a proxy for potential infrastructure damage, providing more consistent insights than mortality rates alone.
