Design and Machine Learning Practices for Automotive Finance Risk Control

1. Automotive Finance Risk‑Control Process Design

The risk‑control workflow focuses on five key nodes: customer acquisition, anti‑fraud, credit assessment, limit setting, and interest‑rate determination. Designing the process revolves around these points.

Two additional critical factors are data completeness and customer‑group characteristics. Complete data (bank credit data, third‑party data, etc.) enriches feature dimensions, reduces reliance on applicant‑submitted information, simplifies the workflow, and improves approval efficiency. Rich data also expands design freedom for each risk‑control node.

Customer segmentation enables differentiated risk‑control paths: high‑quality customers receive simpler processes, while lower‑quality customers undergo more granular approval and are routed through distinct channels for tailored risk assessment.

Overall Automotive Finance Risk‑Control Flow

The end‑to‑end flow covers the entire vehicle‑finance lifecycle and consists of five stages:

Admission & channel rating

Anti‑fraud

Credit assessment

In‑loan monitoring

Post‑loan collection & back‑rating

2. Pre‑Loan Process

The typical pre‑loan flow includes anti‑fraud, credit assessment, and limit pricing. In practice, additional admission criteria and customer‑group analysis are often inserted.

Anti‑Fraud Dimensions

Blacklist

Application behavior anomalies

Negative records

Real‑name inconsistencies

Consumption behavior (e.g., bank statements)

Group fraud detection via relational analysis

Customer Segmentation for Modeling

Automotive finance typically segments customers into groups such as manufacturer‑backed finance, leasing, direct rent, used‑car loans, commercial‑vehicle loans, and car‑mortgage loans. Independent and identically distributed (i.i.d.) samples are required for each segment, so separate models are built per group.

Model Evaluation Metrics

KS (Kolmogorov‑Smirnov) – discriminative power

PSI (Population Stability Index) – distribution stability

Score distribution – near‑normal, monotonic bad‑rate across score bins

Modeling Techniques

Traditional models such as logistic regression and decision trees are widely used. Logistic regression offers interpretability and can be transformed into a scoring table. Decision trees capture non‑linear patterns but risk over‑fitting; ensemble methods (bagging, boosting, stacking) mitigate this.

Boosting, especially GBDT, is the most common in automotive finance. GBDT builds trees sequentially to fit the negative gradient of the loss function.

Improvements include XGBoost (regularization, second‑order gradients, shrinkage, column sampling, gradient‑based split search), LightGBM for large datasets, and CatBoost for categorical features.

3. Automated Machine‑Learning Platform

The platform addresses four major pain points: high entry barrier, low efficiency of manual hyper‑parameter tuning, long development cycles, and the gap between modeling and production environments.

Key features include reusable data, sample, cleaning, processing, model, and tuning pipelines; data source integration; one‑click deployment; interactive graphical interfaces; and end‑to‑end toolchain integration (data analysis, visualization, modeling, deployment).

4. Summary

The presentation shares Baifeng’s design experience in automotive finance risk control, technical accumulations in model building, challenges encountered, and the company’s attempts to overcome them through a unified, automated modeling platform.