Credit Scoring Cards vs Machine Learning in Financial Risk Control: Comparative Analysis and Practical Applications

In the domestic and international financial risk‑control field there are two main schools: the traditional statisticians who favor credit‑scoring‑card models, and the newer internet‑oriented explorers who apply machine learning and deep learning techniques. The speaker, from Rong360, discusses the advantages and disadvantages of both approaches and presents practical ML applications in risk control.

1. Credit Scoring Card Model

The scoring‑card model is a simple linear weighted sum regression that has been used for over a century. Feature selection prefers high coverage (>70%), strong linear correlation with delinquency, stable distribution over time, and high interpretability, typically using 8‑12 variables such as blacklist status, debt, and assets. Parameter estimation can be done via expert weights when data are scarce, or using KS/IV values when a few hundred samples are available.

2. Handling Non‑Linear and Interaction Features

Linear models cannot capture non‑linear or interaction effects. Non‑linear features are treated by WOE transformation or binning, turning non‑linear relationships into linear ones. Interaction features can be modeled through customer segmentation, similar to decision‑tree leaf nodes.

3. Emergence of Machine Learning

Machine learning addresses the weak feature correlation and low‑cost data requirements of internet finance. It offers stronger fitting ability for non‑linear and interaction features, and can combine base learners (RF, GBDT, XGBoost, LightGBM) for more robust models.

4. Financial Risk‑Control Data

Risk‑control data are divided into qualification, credit, consumption, and behavior data. Internet finance often lacks qualification and credit data, relying heavily on behavior data, which is cheap and widely available but has weaker correlation and rapid drift.

5. Advantages and Issues of Machine Learning

ML can extract value from sparse, noisy data, but differs from advertising/recommendation in sample size, prediction horizon, model update frequency, and generalization requirements. Successful ML adoption in credit risk requires careful migration and adaptation.

6. Three Directions of ML in Credit Risk

1) Use ML/AI to generate new features while keeping the scoring‑card as the final model. 2) Replace the scoring‑card with complex models (e.g., XGBoost) while retaining traditional feature selection rules. 3) Deploy large‑scale ML models with millions of samples and tens of thousands of features.

7. Practical ML Feature Engineering

Examples include social‑graph features, bipartite‑graph random walks, word2vec on user click streams, and order‑sequence modeling with LSTM. These techniques enrich feature sets and improve KS/IV metrics.

8. Model Monitoring and Compression

To ensure stability, monitor data and feature distributions, automate model retraining, and apply feature compression (high IV, low PSI selection, LDA topic modeling) to reduce model complexity.

9. Large‑Scale Machine Learning

Experiments with XGB‑LR, DeepFM, and Deep&Wide on millions of samples and tens of thousands of features show better performance and stability compared to traditional ML‑as‑a‑tool approaches, though over‑fitting remains a concern.

10. Conclusions

Machine learning in credit risk is still less mature than in advertising or NLP, with a development horizon of 5‑6 years. Successful adoption requires aligning new techniques with the specific constraints of risk control, and there remains ample space for further research and innovation.