Proprietary Credit Risk Probability Map - System Overview
1. Introduction
This document provides an overview of the Evolved Credit Risk Analysis System, a proof-of-concept designed to offer a holistic and dynamic view of credit risk. It integrates quantitative models, qualitative data, knowledge graph context, and scenario analysis to provide a comprehensive assessment framework.
The system operates on a synthetic, file-based dataset for demonstration purposes. For full technical details, setup instructions, and in-depth explanations, please refer to the Full System README.
2. System Capabilities
- Expanded Ontology & Rich Data Model: Defines core financial concepts using Pydantic models for structured data representation.
- Integrated Knowledge Graph: Captures interconnections (subsidiaries, suppliers, guarantors) and provides contextual insights like network centrality.
- Advanced Risk Models: Utilizes a Random Forest Classifier for Probability of Default (PD) and a Gradient Boosting Regressor for Loss Given Default (LGD), both incorporating enhanced feature engineering.
- SHAP Explainability for PD Model: Offers insights into the drivers of PD model predictions.
- Dynamic Pricing Model: Calculates loan interest rates based on PD, LGD, customer segments, and knowledge graph context.
- Comprehensive Risk Mapping Service: Aggregates portfolio-wide risk, including qualitative scores and KG metrics, with summaries by sector and country.
- MLOps Framework (PoC): Features automated model registration, loading of "production" models, prediction logging, data drift detection (numerical & categorical), and simulated performance monitoring.
- Sophisticated Scenario Simulation: Allows for feature-level shocks to raw data, with PD/LGD re-calculated by models to assess stress impact realistically.
3. Walkthrough / How to Use
To get started with the system:
- Ensure Python 3.9+ is installed and set up a virtual environment.
- Install dependencies:
pip install -r requirements.txt - Train the PD and LGD models by running their respective scripts (e.g.,
python -m src.risk_models.pd_model). This also registers them. - Generate system outputs like the orchestration manifest:
python scripts/generate_outputs.py. - Explore the Comprehensive Analysis Notebook for an interactive demonstration of the system's capabilities.
For detailed setup and execution instructions, please see the Setup and Running the System section in the full README.
4. "Probability Map" - Notebook Snapshot/Summary
The Comprehensive Analysis Notebook is the primary interface for visualizing and interacting with the "Probability Map". It synthesizes data from all components to provide risk insights. Below are conceptual snapshots:
Illustrative Portfolio Risk/Return Landscape
This conceptual plot positions different asset classes (corporate loans, synthetic equities, synthetic commodities) in a risk-return space. Bubble size can represent exposure or market capitalization.
Example Peer Comparison
The notebook allows for deep dives into specific entities, including comparison against synthetic peers on key metrics.
Key Portfolio Data Snippet
The portfolio overview generated by the RiskMapService provides a rich table of data. Here's a small sample of what the structure looks like (actual data will vary based on the synthetic dataset):
| loan_id | company_id | company_name | pd_estimate | lgd_estimate | expected_loss_usd | management_quality_score | kg_degree_centrality |
|---|---|---|---|---|---|---|---|
| LOAN7001 | COMP001 | Innovatech Solutions | 0.0452 | 0.3875 | 87567.50 | 8 | 0.052 |
| LOAN7003 | COMP002 | GreenBuild Corp | 0.0911 | 0.5210 | 47463.10 | 7 | 0.030 |
| LOAN7004 | COMP003 | HealthFirst Pharma | 0.1530 | 0.6050 | 92565.00 | 9 | 0.040 |
Deep Dive Snapshot: Example Company Analysis
For a selected company (e.g., Innovatech Solutions - COMP001), a deep dive in the notebook might summarize key risk factors (actual values from a run):
- Quantitative Scores: PD: 0.0452, LGD: 0.3875, Base EL: $87,567.50
- Financial Health (Latest FS): D/E: 0.67, Current Ratio: 2.20. Overall: Good
- PD Model Drivers (SHAP Top 3): debt_to_equity_ratio, company_age_at_origination, net_profit_margin.
- Knowledge Graph Context: Centrality: 0.052, Suppliers: 2, Customers: 1, Subsidiaries: 0. Implications: Moderate network activity, some supplier concentration.
- Qualitative Factors: Management Quality Score: 8.
- Stress Test Impact (Targeted Shock for COMP001): Stressed EL: $131,351.25 (Change: 50.00 %). Resilience: Moderate.
This multi-faceted view helps in forming a comprehensive risk opinion. For the full interactive analysis, please refer to the notebook.
5. Other Generated Outputs
The system can generate structured metadata and data files:
- Orchestration Manifest (JSON): Describes data sources, active models, and key services.
- Sample Risk Profiles (JSON-LD): Example company risk profiles in a linked data format.
- Consolidated Data Export (Excel) - Key data tables in Excel format.
- Excel Data Exporter Notebook - Jupyter Notebook to generate the Excel export.
- Excel Data Summary (Markdown) - A markdown guide to the contents of the consolidated Excel export.