Claims Cost Forecast Lab (Insurance PLPM)

Problem An insurance client needed to project health insurance claims costs (PLPM - Per Life Per Month) for the next 12 months. Inaccurate estimates directly caused poor premium pricing, leading to major underwriting losses or uncompetitive rates.

Solution Developed a robust machine learning time series pipeline using stacked boosting approaches (XGBoost and LightGBM). Built scheduled pipelines for training and deployed an automated drift monitoring framework to track prediction error and feature changes over time.

Challenges * **Data Non-Stationarity**: Healthcare claims are highly volatile and affected by seasonal changes, inflation, and changing policy structures. * **Solution**: Engineered lag features (rolling averages, YoY differences) and isolated seasonal indicators. * **Concept Drift**: Model accuracy degrades as claimant behaviors shift. * **Solution**: Implemented an automated monitoring system that tracks data/concept drift (using PSI and KS-tests) and triggers automated retraining when thresholds are violated.

Results * Forecast accuracy achieved under **10% MAPE**. * Saved hundreds of engineering hours through self-monitoring, self-healing retraining pipelines. * Successfully integrated directly into client premium pricing dashboards.