Workers’ Compensation Underwriting ML Quote App
May
2025
Ongoing
Problem / Purpose
The existing quote model used for workers’ compensation underwriting was ineffective at predicting risk—especially for new clients or clients with volatile claim histories—resulting in limited utility for pricing strategy.
Solution
Redesigned the model from scratch. Created a new dataset combining historical workers’ comp premiums, wages, class codes, and claims data. Developed claim development factors from 10+ years of historical captive data and applied NCCI’s split point methodology to generate a target using primary and ratable excess losses. Built and refined a frequency-severity LightGBM model in DataRobot. Integrated the model into a custom app for the Pricing Team to quote new clients based on estimated annual claims.
Key Achievements / Impact
Delivered a highly improved model that accurately predicts claims for both established and new businesses. Enabled the Pricing Team to stratify risk and assign more accurate underwriting rates. Model performs well for expected claims under $100K and clearly flags high-risk clients over that threshold. App still in active use.
Key Technologies / Tools Used
DataRobot, SQL, Excel, DBT, Insurance Underwriting, Workers' Compensation, Workers' Compensation Underwriting, Feature Engineering, Data Modeling, Snowflake, ML Models, ML Models (Frequency-Severity)
Role
Data Scientist
ProService Hawaii