Introduction: From Market Risk to Credit and Counterparty Exposure

In Lesson 6, we examined quantitative risk management with a focus on market risk, Value at Risk (VaR), and macroeconomic stress testing.

However, institutional financial stability also depends heavily on credit risk and counterparty risk, which measure the possibility that a borrower, bond issuer, or financial counterparty fails to meet its contractual obligations.

Traditionally, credit assessment relied on static rule-based systems such as the FICO credit score.

Today, modern financial institutions use:

  • Machine learning
  • Alternative data sources
  • Graph analytics
  • Real-time fraud detection systems

These technologies enable more accurate credit assessments and faster fraud detection.

This lesson covers:

  • Probability of Default (PD)
  • Loss Given Default (LGD)
  • Exposure at Default (EAD)
  • The Merton Structural Credit Risk Model
  • Machine Learning for Default Prediction
  • IFRS 9 and CECL Provisioning
  • Graph Neural Networks (GNNs) for Fraud Detection

Part 1: The Core Credit Risk Triad (PD, LGD, and EAD)

Under global regulatory standards such as Basel III/IV, expected credit losses are calculated using three key components.

1. Expected Loss (EL)

The expected loss for a loan or credit portfolio is calculated as:

 
EL = PD × LGD × EAD
 

Where:

  • EL = Expected Loss
  • PD = Probability of Default
  • LGD = Loss Given Default
  • EAD = Exposure at Default

This equation forms the foundation of institutional credit risk management.

Probability of Default (PD)

Probability of Default is the likelihood that a borrower will fail to repay their debt over a specified period.

Typical prediction horizons include:

  • 12 months
  • Multiple years (for long-term lending)

PD is expressed as a probability between 0 and 1 (or as a percentage).

Example:

If PD = 0.04,

there is a 4% probability that the borrower will default during the prediction period.

Loss Given Default (LGD)

Loss Given Default measures the proportion of money the lender loses if a borrower defaults.

It accounts for:

  • Collateral recovery
  • Asset liquidation
  • Legal costs
  • Seniority of debt

Formula:

 
LGD = 1 − Recovery Rate
 

Example:

Recovery Rate = 70%

 
LGD = 1 − 0.70 = 0.30
 

The lender loses 30% of the total exposure.

Exposure at Default (EAD)

Exposure at Default represents the total amount owed when default occurs.

It includes:

  • Outstanding loan principal
  • Accrued interest
  • Unused committed credit facilities
  • Revolving credit balances

Example:

A borrower defaults while owing:

  • Principal = $90,000
  • Interest = $2,000
  • Unused credit line = $8,000
 
EAD = $100,000
 

2. Machine Learning for Default Prediction

Traditional credit scoring models use only a limited number of borrower characteristics.

Modern financial institutions employ advanced machine learning algorithms such as:

  • XGBoost
  • LightGBM
  • Deep Neural Networks

These models analyze thousands of variables simultaneously.

Examples of input features include:

  • Credit history
  • Income stability
  • Cash flow volatility
  • Digital subscription payments
  • Banking transaction history
  • Spending behaviour
  • Open Banking API data

The objective is to estimate the borrower’s Probability of Default (PD) more accurately than traditional scorecards.

Part 2: Structural Credit Risk Models (The Merton Model)

Machine learning models rely on historical data.

Structural credit risk models instead use financial theory to estimate default risk.

The most widely known structural model is the Merton Model, developed by Robert Merton.

1. The Firm as a Call Option

The Merton Model views a company’s equity as a European call option on its assets.

Definitions:

  • Vâ‚œ = Market value of the firm’s assets
  • D = Face value of debt
  • T = Debt maturity date

Two possible outcomes exist at maturity.

Case 1: Assets exceed debt

 
V_T > D
 

The company pays its debt.

Shareholders retain the remaining value.

No default occurs.

Case 2: Assets are less than debt

 
V_T ≤ D
 

The company cannot repay its obligations.

Default occurs.

Shareholders receive nothing.

Bondholders assume ownership of the firm’s remaining assets.

2. Distance to Default (DD)

The Merton Model estimates how far a company is from insolvency.

Formula:

 
DD = [ln(V₀ / D) + (r − 0.5σ²_V)T] / (σ_V√T)
 

Where:

  • Vâ‚€ = Current asset value
  • D = Debt obligation
  • r = Risk-free interest rate
  • σ_V = Asset volatility
  • T = Time until debt maturity

Interpretation:

Higher Distance to Default

→ Lower default probability

Lower Distance to Default

→ Higher default probability

Probability of Default

The Probability of Default is calculated as:

 
PD = N(−DD)
 

Where:

  • N() represents the cumulative standard normal distribution.

Financial institutions monitor Distance to Default continuously using market prices and equity volatility.

Part 3: Regulatory Provisioning (IFRS 9 and CECL)

Following the 2008 Global Financial Crisis, accounting standards changed significantly.

Previously,

banks recognized losses only after borrowers defaulted.

Modern standards require banks to estimate losses before defaults occur.

This is known as the Expected Credit Loss (ECL) approach.

1. IFRS 9

IFRS 9 uses a three-stage impairment model.

Stage 1

Loans performing normally.

Provision required:

12-month Expected Credit Loss

Stage 2

Significant increase in credit risk.

Provision required:

Lifetime Expected Credit Loss

Stage 3

Credit-impaired or defaulted loans.

Provision required:

  • Lifetime Expected Credit Loss
  • Interest income calculated using the net carrying amount

CECL (Current Expected Credit Loss)

CECL is the U.S. accounting standard.

Unlike IFRS 9,

CECL requires lenders to recognize lifetime expected losses immediately when a loan is originated.

This generally results in earlier and larger loan-loss provisions.

2. Machine Learning in Loan Provisioning

Modern banks combine machine learning with macroeconomic forecasting.

Models predict variables such as:

  • GDP growth
  • Inflation
  • Unemployment
  • House prices
  • Interest rates

These forecasts feed into Expected Credit Loss models to estimate future loan losses under IFRS 9 and CECL.

Part 4: Graph Neural Networks and Fraud Analytics

Credit risk evaluates whether a borrower cannot repay.

Fraud analytics evaluates whether someone never intended to repay or is committing financial crime.

Fraud detection focuses on:

  • Identity theft
  • Synthetic identities
  • Money laundering
  • Organized fraud rings

1. Real-Time Transaction Scoring

Large financial institutions process millions of transactions every second.

Machine learning models score every transaction almost instantly.

Common behavioral indicators include:

  • Sudden changes in geographic location
  • Unknown device fingerprints
  • High transaction frequency
  • Unusual spending patterns
  • Rapid account activity

Suspicious transactions are flagged before authorization is completed.

2. Graph Neural Networks (GNNs)

Fraud is often committed by groups rather than individuals.

Fraud networks may share:

  • Bank accounts
  • Phone numbers
  • Email addresses
  • Devices
  • Merchant accounts
  • Synthetic identities

Instead of treating each customer independently,

Graph Neural Networks model relationships between entities.

Graph Representation

Fraud networks are represented as:

 
G = (V, E)
 

Where:

  • G = Graph
  • V = Vertices (customers, accounts, devices, merchants)
  • E = Edges (relationships or transactions)

Graph Convolutional Networks (GCNs)

Graph Convolutional Networks perform message passing between connected nodes.

This enables detection of:

  • Fraud rings
  • Synthetic identity networks
  • Money laundering schemes
  • Hidden criminal organizations

Traditional tabular machine learning models often fail to detect these complex relationships.

Summary

Advanced credit risk modeling combines financial theory, machine learning, and regulatory compliance to estimate credit losses and protect financial institutions.

Key concepts include:

  • Expected Loss (EL): Calculated as EL = PD × LGD × EAD, forming the foundation of modern credit risk measurement.
  • Probability of Default (PD): Estimates the likelihood that a borrower will default within a specified time horizon.
  • Loss Given Default (LGD): Measures the proportion of exposure lost after accounting for recoveries from collateral or liquidation.
  • Exposure at Default (EAD): Represents the total outstanding financial exposure when default occurs.
  • Merton Model: Uses option pricing theory to estimate corporate default risk by treating a firm’s equity as a call option on its assets.
  • Distance to Default (DD): Indicates how close a company is to insolvency; a higher DD implies a lower probability of default.
  • IFRS 9 and CECL: Require forward-looking expected credit loss provisioning, ensuring institutions recognize potential losses before defaults occur.
  • Machine Learning: Enhances credit scoring by incorporating thousands of traditional and alternative data features to improve default prediction.
  • Graph Neural Networks (GNNs): Analyze relationships between entities to uncover fraud rings, synthetic identities, and money laundering networks that conventional models may miss.
PreviousNext