Page 1 – Credit Risk Modelling: Concepts & Targets

1.1 Overview of Credit Risk

Credit risk represents the potential for financial loss arising from a borrower’s failure to meet contractual obligations. Supervised machine learning provides a flexible framework for estimating creditworthiness across retail, SME, and corporate portfolios.

1.2 Core Risk Metrics

• Probability of Default (PD) – likelihood of default within a time horizon.
• Loss Given Default (LGD) – proportion of exposure lost if default occurs.
• Exposure at Default (EAD) – amount outstanding at the time of default.
• Expected Loss (EL) – EL = PD × LGD × EAD

1.3 Common Targets for ML Models

Supervised learning can be used to estimate:

• Binary default indicator (default vs non-default).
• Continuous LGD (bounded regression problem).
• Cure prediction (recoveries or restructuring outcomes).
• Fraud / non-fraud classification.
• Churn / retention in wealth and banking products.

1.4 Economic Data Structure

Credit modelling typically uses panel, behavioural, and static customer features:

• Demographics (age, region, customer type)
• Financial metrics (income, liabilities, credit utilisation)
• Repayment history (delinquencies, arrears, roll rates)
• Behavioural signals (spending patterns, payroll deposits)
• Product characteristics (loan type, maturity, limits)

1.5 Industry Practice

Banking models must meet regulatory expectations under:

• Basel II/III – internal ratings-based approaches.
• IFRS 9 – expected credit loss (ECL) estimation.
• Consumer lending regulations – fairness & transparency.

Interpretability and stability are crucial, influencing model choice.

Page 2 – Model Families & Training Pipelines

2.1 Interpretable Baseline Models

Financial institutions frequently begin with interpretable linear models:

• Logistic regression – default classification baseline.
• Linear regression – LGD and EAD estimation.
• Scorecards – monotonic transformations of features.

2.2 Tree-Based & Ensemble Models

More complex relationships can be captured by:

• Decision Trees
• Random Forests
• Gradient Boosted Trees (XGBoost, LightGBM, CatBoost)

Benefits:

• Excellent for tabular financial datasets.
• Handle nonlinear interactions.
• Robust to missing data and outliers.

2.3 Neural Networks in Credit & Fraud

Neural networks are used selectively where complexity is justified:

• Multilayer perceptrons for behavioural scoring.
• Temporal networks (LSTMs, TCNs) for credit histories.
• Deep anomaly detection architectures (Autoencoders).

2.4 Handling Imbalanced Data

Defaults and fraud cases are rare. Techniques include:

• SMOTE oversampling
• Class weighting in the loss function
• Anomaly detection frameworks instead of classification

2.5 Feature Pipelines

A typical financial ML pipeline includes:

1. Extract features from data sources.
2. Perform encoding, scaling, outlier checks.
3. Split by time: train / validation / test.
4. Train model and tune hyperparameters.
5. Perform backtesting for time-based validity.
6. Generate explainability reports.

Page 3 – Fraud Detection & Anomaly Classification

3.1 Nature of Fraud Problems

Fraud detection differs from standard classification:

• Fraud patterns evolve rapidly.
• Fraud datasets are highly imbalanced.
• Fraudsters actively adapt to detection systems.

3.2 Supervised Fraud Detection

Supervised approaches use historical labelled data:

• Gradient boosting methods (LightGBM, XGBoost).
• Neural networks for behavioural modelling.
• Cost-sensitive classifiers for rare events.

3.3 Unsupervised & Semi-Supervised Techniques

When labels are missing or incomplete:

• Autoencoders – reconstruction error signals anomalies.
• Isolation Forest – isolates rare points more easily.
• One-class SVM – learns normal behaviour boundary.

3.4 Real-Time Fraud Decision Engines

Modern fraud systems require:

• Low-latency decision latencies (<50ms).
• Streaming feature pipelines.
• Behavioural profiles updated in near real time.

3.5 Fraud Explainability

Because false positives create customer friction, models must provide:

• Reason codes for decisions.
• Feature attributions.
• Confidence thresholds for escalation.

Page 4 – Evaluation, Stability & Regulatory Metrics

4.1 Metrics Beyond Accuracy

Financial models require specialised evaluation metrics:

• Precision: Of predicted positives, how many are correct?
• Recall (Sensitivity): Of actual positives, how many were detected?
• ROC–AUC: Ranking ability across thresholds.
• PR–AUC: Critical in imbalanced data settings.

4.2 Calibration

Credit risk models must output well-calibrated probabilities. Techniques include:

• Platt scaling
• Isotonic regression
• Temperature scaling

4.3 Stability Over Time

Financial data changes due to:

• Economic cycles
• Product changes
• Regulatory interventions

Monitoring metrics include:

• Population stability index (PSI)
• Characteristic stability index (CSI)
• Drift scores in continuous monitoring

4.4 Fairness Metrics

In regulated applications:

• Equal opportunity
• Demographic parity
• Adverse impact ratio
• Group-level calibration

4.5 Stress and Scenario Testing

Credit and fraud models must be tested under:

• Macro shocks (e.g., rates +300bps)
• Market crashes
• Credit downturn scenarios

Stress testing reveals model fragility and supports regulatory filings.

Page 5 – Deployment, Monitoring & Governance in Finance

5.1 Deployment Architectures

Typical financial ML deployment pipelines include:

• Batch scoring for risk and reporting cycles.
• Online scoring for real-time credit or fraud decisions.
• API-based scoring engines with audit logs.

5.2 Monitoring & Alerts

Critical monitoring dimensions:

• Feature distribution drift
• Prediction drift
• Model latency
• Error spikes and alert thresholds

5.3 Regulatory Documentation

Financial institutions require documentation such as:

• Model development documents (MDD)
• Model validation reports
• Ongoing monitoring reports
• Fairness and compliance attestations

5.4 Human Oversight

AI systems in finance must include human decision points:

• Overrides for borderline decisions
• Manual review for flagged transactions
• Risk committee escalation procedures

5.5 Summary

Module 2 introduced supervised learning as the backbone of modern financial classification tasks. It covered:

• Credit risk target variables and modelling frameworks.
• Fraud detection methods and anomaly detection architectures.
• Model families and training pipelines.
• Evaluation metrics, stability, fairness, and stress testing.
• Deployment considerations and governance requirements.

Next, Module 3 will examine time-series forecasting, algorithmic trading, volatility modelling, and signal engineering.