Automated Feature Engineering and Modeling for Credit Risk: A DataFun Case Study

Model development typically involves business analysis, data preparation, feature engineering, model building, evaluation and monitoring, with feature engineering and model building consuming about 60% and 30% of the total time respectively.

Traditional credit‑risk modeling can take around 20 days, mainly because manual feature extraction is complex and time‑consuming; automating this step can accelerate the process.

Rong360’s automated feature engineering and modeling solution abstracts the most time‑intensive parts into a tool that integrates automatic feature generation, variable selection, model tuning, deployment and monitoring.

Key advantages include: ① significantly improving modeling efficiency and standardization; ② reducing labor and time costs; ③ quickly meeting business requirements; ④ enabling rapid experimentation with more model variants.

With automation, the end‑to‑end cycle from development to production can be shortened to roughly five days while maintaining high accuracy.

Automated Feature Engineering – Manual feature engineering relies heavily on domain knowledge and is error‑prone; automation standardizes the process and handles new data sources more efficiently.

Effective automation requires data with structural similarity; the classic RFM (Recency, Frequency, Monetary) model exemplifies three key customer metrics used in many CRM analyses.

In Rong360’s transaction‑level data, RFM‑derived features include statistical variables (e.g., total call duration) and trend variables (e.g., change in call count over recent months).

RFM can also be applied to relational networks to generate features such as the number of first‑degree contacts, their loan histories, and average borrowing amounts.

The automated feature pipeline first processes different data types, aggregates basic statistical features for categorical variables, then derives ratio and trend features in a second layer.

Automation frees model developers to focus on data understanding rather than low‑level feature coding.

Automated Modeling – The platform bundles widely used algorithms (XGBoost, LightGBM, Logistic Regression) and modularizes common modeling steps to boost efficiency.

Automatic feature selection uses EDA to drop high‑missing‑rate or low‑variance features, then applies IV filtering, tree‑based importance, and collinearity checks, reducing thousands of dimensions to a few hundred while preserving diversity.

Logistic Regression (LR) is often used for credit scoring; the article presents the odds‑to‑probability conversion, scoring card formula, and the role of WOE binning, illustrated with the following equations:

WOE binning, traditionally done manually, is now automated with equal‑frequency binning and monotonicity checks, while still allowing manual adjustments.

XGBoost, a leading algorithm, benefits from preprocessing such as binning continuous variables and encoding categorical ones (label encoding, one‑hot, etc.), and the platform supports extensive hyper‑parameter search (grid, random).

Automated Model Evaluation and Monitoring – Beyond AUC and KS, the system monitors cross‑dimensional (variables, model scores), cross‑sample (train, test, OOT), and cross‑model (current vs. new versions) performance, aggregating weighted scores into a comprehensive report.

It tracks metric drift (AUC, PSI, ranking), variable distribution shifts, and provides alerts when deviations occur.

Model Deployment and Online Monitoring – Instead of hand‑written Python scripts, models are exported as files with accompanying configuration files, enabling rapid, script‑free deployment.

Online monitoring compares daily metrics against baseline, detects feature drift, and triggers alerts for issues such as PSI spikes or performance degradation.

Monitoring outputs include weekly reports, real‑time alerts, and visual dashboards.