Applying Automated Feature Engineering and Auto Modeling to Risk Control Scenarios

1. Background and Problem

Model development in risk control traditionally follows a multi‑step pipeline (business analysis, data preparation, feature engineering, model building, evaluation, monitoring). Feature engineering and model construction consume the majority of time—about 60% and 30% respectively—making rapid model delivery difficult.

Rong360 introduced an automated feature‑engineering and auto‑modeling solution that abstracts the most time‑consuming steps into a unified tool, improving efficiency, standardization, and model quality while shortening the end‑to‑end cycle to roughly five days.

2. Automated Feature Engineering

Manual feature engineering relies on domain knowledge and is labor‑intensive. Automated methods leverage the RFM (Recency, Frequency, Monetary) model to generate statistical and trend features from transaction‑level data, and can also construct network‑based features (e.g., number of first‑degree contacts, their borrowing behavior).

The automated pipeline first aggregates basic statistics per variable, then derives ratio and trend features in a second layer, producing a rich feature set with minimal manual effort.

3. Automated Modeling

Popular algorithms such as XGBoost, LightGBM, and Logistic Regression (LR) are integrated into the platform. The tool performs automatic EDA‑based feature filtering (high missing rate, low variance, instability), followed by IV screening, tree‑model importance, and collinearity checks, reducing thousands of features to a few hundred high‑value ones.

The LR section explains odds, probability, and the scoring card formula, emphasizing the critical role of WOE binning, which is also automated (equal‑frequency binning with monotonicity checks) while allowing manual adjustments.

For XGBoost, continuous variables are binned to reduce over‑fitting, and categorical variables are encoded (label, one‑hot, etc.). The platform also provides automated hyper‑parameter tuning via GridSearch and RandomSearch.

4. Automated Model Evaluation and Monitoring

Beyond traditional metrics (AUC, KS), the system evaluates models across dimensions (feature, model), samples (train, test, OOT, early‑performance), and versions (current vs. new). It monitors feature drift, PSI, ranking stability, and triggers alerts when deviations exceed thresholds.

5. Model Deployment and Online Monitoring

Instead of hand‑written Python scripts, models are packaged as configuration files that the Rong360 deployment platform consumes, enabling rapid rollout and automatic scoring. Post‑deployment, the platform continuously compares live metrics (KS, AUC, PSI) against baseline and checks feature binning consistency to detect data drift.

Author Introduction

Jiang Hong, Head of Risk‑Control Business Modeling at Rong360, holds a degree from Shanghai Jiao‑Tong University and has extensive experience in credit modeling, fraud detection, and data mining.

Job Opportunities

Rong360 is hiring senior data algorithm engineers (machine‑learning), senior risk‑control algorithm engineers, and senior data analysts in Beijing. Contact: [email protected].

End of Article