AI-Powered Credit Risk Assessment Platform: From Predictive Models to Explainable Intelligence

This AI-powered credit risk assessment platform is an end-to-end solution for banking scenarios, integrating LightGBM/XGBoost predictive models, SHAP explainable AI technology, and natural language query functions. The project achieves one-click deployment through decoupled architecture design and Docker containerization, while combining asymmetric cost-benefit modeling to ensure model decisions align with banks' risk tolerance and meet compliance audit requirements.

Original Author & Source

• Author/Maintainer: deva1702
• Source Platform: GitHub
• Original Title: credit_risk_model
• Release Date: June 2, 2026

Project Overview An end-to-end AI risk assessment platform for bank credit scenarios, integrating machine learning prediction, explainable AI, and natural language interaction. It adopts a decoupled architecture (separation of machine learning, data engineering, conversational AI, and presentation layers) and supports one-click deployment via Docker containerization.

System Architecture Four-layer architecture:

1. Client Layer: Streamlit Web Interface
2. Intelligence & Agent Layer: Core components include Groq LLaMA-3.3-70B (natural language understanding), NL-to-SQL translator, SHAP interpreter, inference engine (loads LightGBM/XGBoost models)
3. Data & Persistence Layer: SQLite database, CSV data sources, pre-trained model files

Predictive Model & Feature Engineering

• Core Models: LightGBM and XGBoost
• Data Cleaning: Filter columns with missing rate >40% (except TARGET and EXT_SOURCE scores); label encoding for categorical features, with unknown labels falling back to "Missing"
• Domain Features: Designed financial health indicators such as CREDIT_INCOME_RATIO, ANNUITY_INCOME_RATIO, DEBT_SERVICE_RATIO, and CREDIT_STRESS
• Class Imbalance Handling: Adopt class weight adjustment (scale_pos_weight=5) to improve recall rate for high-risk applicants

Asymmetric Cost-Benefit Modeling

• False Negative Cost (approving defaulting applicants): 60% of loan principal (LGD)
• False Positive Cost (rejecting good borrowers): 10% of loan principal (NIM)
• Decision threshold set to 0.30 (instead of 0.5) to optimize business costs

Model Performance Evaluation Adopt 80/20 stratified training validation, KS statistic to evaluate separation:

Metric	LightGBM	XGBoost	Winner
ROC-AUC	0.7673	0.7649	LightGBM
PR-AUC	0.2608	0.2578	LightGBM
KS Statistic	0.4089	0.4016	LightGBM

Explainable AI

• Global Explanation: Beeswarm plots and average impact plots show that EXT_SOURCE_2/3 dominate global risk signals
• Individual Explanation: Dynamic factor cards (color-coded feature impacts), waterfall charts (feature adjustment process)

Natural Language Query Supports pure English dataset exploration, ensures SQL accuracy via schema injection prompts, and has a hallucination fallback mechanism

Technical Highlights

1. Domain Knowledge Integration: Integrate financial theory into professional features like CREDIT_STRESS
2. Business Goal Alignment: Asymmetric cost modeling optimizes actual business metrics
3. Explainability Priority: SHAP integration meets regulatory and user trust requirements
4. Natural Language Interaction: Reduces the analysis barrier for non-technical users
5. Engineering Mindset: Layered architecture and containerization ensure maintainability

Project Insights Provides a reference for AI applications in the financial sector, demonstrating how to transform lab prototypes into production-grade banking solutions

Deployment Steps

1. Clone the repository: git clone <repository> && cd credit_risk_model
2. Configure environment variables: Create a .env file and set GROQ_API_KEY, DATA_PATH, MODEL_PATH, DB_PATH, ACTIVE_MODEL
3. Start the container: docker-compose up
4. Access: http://localhost:9200