Risk Analytics Engine

An end-to-end credit risk modeling and analytics pipeline built with Python, LightGBM, and XGBoost.

This project simulates a real-world financial risk modeling workflow, including:

• feature screening
• supervised variable selection
• tree-based feature filtering
• LightGBM / XGBoost modeling
• model evaluation
• stability analysis
• score enhancement analysis
• automated model reporting

The repository is designed as a reusable modeling framework for:

• credit risk analytics
• fintech modeling
• banking data science
• machine learning model development

---

Project Structure

risk-analytics-engine/
│
├── notebooks/
│   ├── 01.无监督筛选.ipynb
│   ├── 02.有监督筛选.ipynb
│   ├── 03.单变量初筛.ipynb
│   ├── 04.决策树初筛_不含评分.ipynb
│   ├── 04.决策树初筛_含评分.ipynb
│   ├── 05.LGBM不含评分.ipynb
│   ├── 05.LGBM含评分.ipynb
│   ├── 06.LGBM模型开发报告整理.ipynb
│   ├── 07.XGB模型优化.ipynb
│   ├── 08.最终模型结论整理.ipynb
│   └── 09.正式模型开发报告.ipynb
│
├── scripts/
│   └── generate_demo_data.py
│
├── src/
│   └── 函数代码包.py
│
├── docs/
│   └── data_schema.md
│
├── requirements.txt
└── README.md

---

Workflow Overview

01. 无监督筛选 (Unsupervised Feature Screening)

Perform:

• missing rate analysis
• dominant value analysis
• PSI stability analysis

Goal: Remove unstable or low-quality variables before modeling.

---

02. 有监督筛选 (Supervised Feature Screening)

Evaluate variable predictive power using:

• KS
• AUC
• monthly KS
• overall performance

Goal: Identify variables with meaningful risk discrimination ability.

---

03. 单变量初筛 (Initial Feature Filtering)

Combine:

• PSI thresholds
• KS thresholds
• coverage rules
• business exclusion lists

Goal: Build the initial candidate variable pool.

---

04. 决策树初筛 (Tree-based Feature Filtering)

Use:

• LightGBM feature importance
• product prefix control
• variable redundancy reduction

Two versions:

• without external score variables
• with external score variables

Goal: Reduce dimensionality and retain strong variables.

---

05. LGBM Modeling

Train LightGBM models for:

• y1 target
• y2 target
• score-enhanced versions
• non-score versions

Outputs:

• KS
• AUC
• feature importance
• model stability metrics

---

06. LGBM Report Aggregation

Aggregate:

• model metrics
• feature importance
• ranking results
• business conclusions

Generate standardized model summary tables.

---

07. XGB Model Optimization

Train and optimize XGBoost models.

Includes:

• hyperparameter tuning
• stability comparison
• score enhancement analysis
• algorithm comparison with LGBM

---

08. Final Model Comparison

Compare:

• LGBM vs XGB
• y1 vs y2
• score vs non-score models

Automatically generate:

• model rankings
• recommendation conclusions
• algorithm comparison tables

---

09. Formal Modeling Report

Generate a structured modeling report including:

• project background
• feature engineering workflow
• model evaluation
• score analysis
• ranking stability
• final recommendations

Output format:

• Excel-based formal modeling report

---

Demo Data

For privacy protection, the original business data is NOT included.

Instead, the repository provides:

scripts/generate_demo_data.py

which generates simulated datasets with similar structure and workflow logic.

The demo data preserves:

• feature naming conventions
• variable engineering logic
• pipeline structure
• modeling workflow

while removing all sensitive business information.

---

Tech Stack

• Python
• Pandas
• NumPy
• LightGBM
• XGBoost
• Scikit-learn
• OpenPyXL
• Jupyter Notebook

---

Modeling Topics Covered

This project covers many practical industry modeling techniques, including:

• PSI stability analysis
• KS/AUC evaluation
• feature importance filtering
• score enhancement analysis
• model decay analysis
• OOT validation
• decision table monotonicity
• Lift analysis
• model ranking systems
• automated report generation

---

Notes

This repository is a sanitized educational version of a real-world risk modeling workflow.

All:

• customer information
• business identifiers
• sensitive variables
• production data

have been removed or anonymized.

---

Author

Althea He

Interested in:

• Risk Analytics
• Data Science
• Machine Learning Engineering
• Financial Technology
• AI Infrastructure