J.G. Wentworth - ML Platform for Financial Services
Key Highlights
- •Built 3 production ML models: enrollment, contact propensity, and churn prediction
- •Designed media mix optimization managing multi-million dollar marketing budgets
- •Developed custom MOM scoring algorithm for feature selection
- •Integrated TransUnion credit data and US Census demographics into 70+ feature pipeline
- •Delivered $500K+ budget reallocation recommendations across 50+ marketing channels
Overview
Machine learning platform for J.G. Wentworth, a debt resolution company, predicting customer behavior and optimizing marketing spend across their multi-million dollar acquisition pipeline. Built as outsourced ML engineer through CML Insights, processing 30K+ monthly leads with real-time scoring.
Architecture
End-to-end ML system with integrated data sources:
- Data Integration: Azure SQL Data Warehouse, TransUnion credit API, US Census DP03, Salesforce CRM
- ML Models: Three gradient boosting classifiers for enrollment, contact, and churn prediction
- Optimization: Polynomial regression for media mix modeling across 50+ channels
- Deployment: MLflow tracking with monthly retraining cycles
Key Technical Contributions
Custom MOM Feature Selection Algorithm
Developed Measure of Match scoring from scratch, evaluating prediction power by measuring distribution overlap between positive and negative classes. For numerical features, creates histograms calculating probability densities with overlap metric Σ min(P_pos[bin_i], P_neg[bin_i]). For categoricals, uses frequency distributions with top-N grouping for high cardinality. Handles missing values explicitly since missingness itself is predictive in financial data.
Three Production ML Models
Built gradient boosting classifiers: enrollment prediction (14-day window, tertile scoring for sales prioritization), contact propensity (70+ features with top predictors: total debt 20%, marketing channel 12%, FICO 4%), and cancellation prediction (90-day churn risk with intervention triggers). Applied stratified sampling for 8% enrollment rate imbalance, class weights in training, and probability calibration for realistic score distributions.
Media Mix Optimization
Designed polynomial saturation modeling (degree-2 polynomials via scikit-learn) capturing diminishing returns in CPL and realizable effect as lead volume scales. Built constrained optimization with scipy solving maximize Σ(revenue - cost) subject to budget, saturation limits, and contractual minimums. Validated through A/B tests and holdout analysis, delivering $500K+ reallocation recommendations (±20% adjustments across channels).
Data Engineering Pipeline
Integrated TransUnion credit reports (trade lines, balances, inquiries), Census economic indicators (DP03 by ZIP), and Salesforce interactions into 70-80 feature pipeline. Handled different refresh cadences (daily leads, monthly census), schema evolution, PII anonymization, and high cardinality through target encoding and geographic hierarchies (ZIP→county→state).
Technologies
- Python: Scikit-learn (GradientBoost, RandomForest), Pandas, NumPy
- Azure: SQL Data Warehouse for data extraction
- MLflow: Experiment tracking and model registry
- SciPy: Constrained optimization for budget allocation
- Data: 2+ years history, 100K+ observations, 70+ features
Impact
25% improvement in contact-to-enrollment conversion, 15% reduction in customer acquisition cost, 10% decrease in 90-day churn. System processes 30K+ leads monthly with automated scoring and monthly model retraining maintaining accuracy as market conditions shift.