PHL Risk Diagnostics Dashboard

Structure-aware model risk diagnostics for credit decision systems.

This project implements the Projection Density Law (PHL) framework and provides a practical tool for analyzing decision boundary sensitivity and structural instability in neural network models.

---

Live Demo

Open the deployed dashboard:

https://phl-risk-dashboard-rngrymxad3purafnszxkax.streamlit.app/

---

Run Locally

streamlit run app.py

Then open:

http://localhost:8501

---

What This Project Does

This dashboard analyzes model behavior beyond traditional feature importance by focusing on decision structure:

• Logit scale (sigma) → output volatility
• Mid-density (mid) → decision boundary uncertainty
• Effective rank (r) → representation capacity
• Spectrum Sharpening Index (SSI) → feature concentration risk

It helps answer:

Why is a model decision unstable or sensitive near approval thresholds?

---

Use Case: Credit Risk Explanation

Designed for bank Model Risk Management (MRM) and compliance teams.

Typical scenarios:

• Borderline applications (q ~ 0.5)
• Loan rejections requiring explanation
• High-sensitivity decision regions

PHL provides:

• Structural explanation (beyond feature attribution)
• Model risk diagnostics
• Automated PDF reporting

Supports:

• Adverse Action Explanation
• EU AI Act (high-risk AI system documentation)

---

Demo Scenario (German Credit)

Includes a built-in interactive demo (no model training required):

German Credit – Rejection Analysis

• Applicant requests €5000 loan
• Model output: q = 0.48 (near decision boundary)
• Decision: Rejected

PHL reveals:

• Reduced effective rank → limited representation capacity
• Spectrum sharpening → feature concentration
• High-sensitivity zone near decision boundary

Small changes in applicant features may significantly affect outcomes.

---

PDF Report Output

Generates a professional MRM-style risk report including:

• Executive Summary (Risk Level)
• Key Metrics (sigma, mid, rank, SSI)
• Spectrum Analysis
• Structural Interpretation
• Recommended Actions
• Compliance-oriented explanation

---

PHL PDF Risk Report – Final Enhancement (MRM-ready POC)

This update upgrades the PHL PDF report to a professional, compliance-aligned format suitable for bank Model Risk Management (MRM) teams and regulatory audits.

Key Improvements

1. Executive Summary (Business + Regulatory Focus)

• Added a strong regulatory-aligned opening referencing Adverse Action Explanation and EU AI Act high-risk AI systems.
• Explicitly interprets MEDIUM RISK in business terms: potential inconsistency in credit decisions and increased risk of appeals/regulatory scrutiny.

2. Financial Language Standardization

Translated technical terms into clear, business-oriented language:

• “Reduced model representational capacity, leading to higher sensitivity at decision boundary”
• “High-sensitivity zone near the decision boundary, where small changes in applicant features may lead to inconsistent approval outcomes”
• Recommendations rewritten as actionable banking procedures.

3. Visual and Layout Improvements

• Increased spacing and margins for better readability.
• Risk Level (MEDIUM) visually emphasized using professional color palettes.
• Key Metrics table redesigned with structured borders and audit-style alignment.
• Spectrum chart enlarged, centered, and given a clear business-oriented caption explaining spectrum sharpening.

4. Regulatory Alignment Section (New)

• Added dedicated section: “Alignment with Regulatory Requirements”
• Covers SR 11-7 Model Risk Management expectations, Adverse Action Explanation requirements, and EU AI Act technical documentation standards.

5. Compliance-safe Design

• No changes to core PHL metrics or computation logic.
• Maintains the fpdf2 + kaleido pipeline for high-quality PDF generation.
• Output remains fully compatible with Streamlit download_button (returns bytes).

Result

The PHL PDF Risk Report is now positioned as:

A decision stability and structural risk explanation document for credit model governance.

Suitable for:

• Model validation and independent challenge
• Internal audit discussions
• Regulatory technical documentation support
• Loan rejection appeal explanations

This version makes the report MRM-ready — professional enough for bank risk and compliance teams while clearly highlighting the unique value of PHL’s spectrum-based structural diagnosis.

---

Core Idea (Simplified)

We study the projection:

z = w^T h

Define:

mid = P(|q - 0.5| <= eps)

Empirically:

mid ~ eps / sigma

This links decision uncertainty to logit scale and representation geometry.

---

Project Structure

phl-risk-dashboard/

• app.py — Streamlit entry point
• ui/ — Dashboard UI modules
• analyzer.py — Core PHL metrics
• data_loader_credit.py — German Credit dataset loader
• model.py — Reference MLP (optional)

---

Installation

pip install -r requirements.txt

Main dependencies:

• streamlit
• torch
• plotly
• fpdf2
• kaleido
• scikit-learn
• pandas

---

Important Notes

• This is a research prototype (POC)
• Not intended as a standalone decision system
• Should be used with human oversight and existing governance frameworks

---

Positioning

This is not a feature importance tool.

It introduces structure-aware model risk diagnostics, focusing on:

• Decision boundary behavior
• Representation collapse
• Structural instability

---

Author

Independent Researcher
Focus: Structure-aware XAI / Model Risk / AI Governance