Business Context

Brillio is helping an insurance client improve claim cost prediction in its analytics platform. The current linear model underestimates high-cost claims because raw operational fields are noisy, sparse, and weakly encoded.

Dataset

You are given a historical claims dataset used to predict total claim payout amount for open claims 14 days after first notice of loss. The goal is to design a feature engineering approach that improves regression performance while remaining explainable enough for claims operations.

• Size: 240K claims over 4 years, 41 raw columns
• Target: Continuous — total payout amount in USD
• Target distribution: Heavily right-skewed with a long tail; ~6% of claims exceed $25K
• Missing data: 18% missing in repair and adjuster fields, 7% missing in customer demographics, some missingness is informative

Success Criteria

A good solution should reduce error versus a regularized linear baseline and clearly justify which engineered features matter most. Aim for MAE < $1,850 and RMSE < $4,200 on a held-out time-based test set.

Constraints

• Predictions run in a daily batch inside a Brillio analytics workflow; inference should stay under 10 minutes for 50K claims
• Claims operations needs interpretable drivers, not a black-box-only solution
• Feature generation must avoid leakage from post-outcome information

Deliverables

1. Propose a feature engineering plan for numerical, categorical, temporal, and sparse text-derived inputs
2. Build a regression pipeline with preprocessing and at least one engineered-feature baseline
3. Use a time-based validation strategy and explain leakage risks
4. Evaluate against baseline metrics and identify the most useful engineered features
5. Describe how you would productionize feature generation and monitor drift