​CREDIT SCORING MODELS

​

Module 25: Credit Scoring Models by Product

​

• Origination Credit Scoring

  • Credit Card Score

  • Mortgage Score

  • consumption scores

  • Car Score

• Behavior Score (BS)

  • Temporal horizon

  • Dashboard data information

  • Panel data regression

  • Cox regression

  • Behavior Score with macroeconomic variables

  • transition matrices

  • Behavior Score with transition matrices

  • Transaction Score

  • Machine Learning Models

  • BEHAVIOR SCORE ON CREDIT CARDS

• Exercise 55: Behavior Score Logistic Regression in Python data 2

• Exercise 56: Behavior Score Support Vector Machines in python

• Exercise 57: Behavior Score Random Forest in python

• ​Exercise 58: Behavior Score Gradient Boosting Trees in python

• Exercise 59: Behavior Score Deep Learning LSTM in python

​

Module 26: Typology of Scoring models

​​​

• Response Score

• Income score

• Churn Score

• Origination Fraud Score

• Behavior Fraud Score

• Collection Score

• Recovery Score

• Big Data Scoring

• Exercise 60: Fraud Score with neural networks

• Exercise 61: Income Score

• Exercise 62: Collection Score

• Exercise 63: Recovery Score

• Exercise 64: Quit Score

​

CALIBRATION OF PD MODELS

 

Module 27: Calibration of the Probability of Default PD IRB

​

• PD estimation

  • econometric models

  • Machine Learning Models

  • Data requirement

  • Risk drivers and credit scoring criteria

  • Rating philosophy

  • Pool Treatment

• PD Calibration

  • Default Definition

  • Long run average for PD

  • Technical defaults and technical default filters

  • Data requirement

  • One Year Default Rate Calculation

  • Long-Term Default Rate Calculation

• PD Model Risk

  • Conservatism Margin

• PD Calibration Techniques

  • Anchor Point Estimate

  • Mapping from Score to PD

  • ​Adjustment to the PD Economic Cycle

  • Rating Philosophy

• PD Trough The Cycle (PD TTC) models

• PD Point in Time PD (PD PIT ) models

• PD Calibration of Models Using Machine and Deep Learning

• Exercise 65: PD Calibration in Machine Learning Models

​

Module 28: Machine Learning models to estimate Lifetime PD under IFRS 9

​​

• Credit scoring models to estimate Lifetime PD

• PD Lifetime in IFRS 9

• Impact of COVID-19 on models

• Climate Risk Impact

• Inflation impact

• Impact of rising prices

• Regression Models

  • Logistic regression

  • Logistic Multinomial Regression

  • Ordinal Probit Regression

• VAR and VEC models

• Machine Learning Model​

  • SVM: Kernel Function Definition

  • Neural Network: definition of hyperparameters and activation function

  • deep learning

  • LSTM

• PD Calibration of Models Using Machine and Deep Learning

• Exercise 66: PD Lifetime using logistic regression

• Exercise 67: PD Lifetime using multinomial regression in R

• Exercise 68: PD Lifetime using SVM in Python

• Exercise 69: PD Lifetime using Deep Learning in Python

• Exercise 70: PD Lifetime using Deep Learning LSTM in Python

​

VALIDATION OF PD MODELS

 

​Module 29: Validation of PD models

​

• Definition of PD Backtesting

• PD Calibration Validation​

  • normal test

  • Binomial Test

  • Traffic Light Approach

• Traffic Light Analysis and PD Dashboard

• PS Stability Test

• Forecasting PD vs. Real PD in time

• When should we recalibrate or reestimate a credit scoring model?

• Re-development

• Re-estimation

• Model Risk in PD

• Machine Learning to validate PD models

• Artificial Intelligence to recalibrate and rebuild models autonomously

• Exercise 71: Backtesting PD in Excel

• Exercise 72: Forecasting PD and actual PD in Excel

​

AUTOMATION OF CREDIT SCORING AND PD WITH AI

 

Module 30: Automation of Credit Scoring and PD Modeling

​

• What is modeling automation?

• that is automated

• Automation of machine learning processes

• Optimizers and Evaluators

• Modeling Automation Workflow Components

  • Summary

  • Indicted

  • Feature engineering

  • Model generation

  • Assessment

• Hyperparameter optimization

• Reconstruction or recalibration of credit scoring

• Credit Scoring Modeling

  • Main milestones

  • Evaluation and optimization

  • Possible Issues

• PD calibration modeling

  • Evaluation and optimization

  • backtesting

  • Discriminating Power

  • Stability Tests

• Global evaluation of modeling automation

• Implementation of modeling automation in banking

• Technological requirements

• available tools

• Benefits and possible ROI estimation

• Main Issues

• Model Risk

• Genetic algorithms

• Exercise 73: Automation of the modeling, optimization and validation of credit scoring hyperparametry

• Exercise 74: Automation of PD modeling and validation

​

Module 31: Explainable AI, Counterfactual Credit Decisions & Actionable Scoring

​

Modern credit scoring requires more than predictive accuracy. Banks need to understand why a customer receives a particular score, which variables drive the decision and what realistic changes could alter the outcome.

This module develops advanced explainability techniques specifically designed for lending decisions and moves beyond traditional feature importance towards actionable explanations and counterfactual analysis.

Topics

• Explainable AI in Credit Scoring

• Global vs Local Explainability

• SHAP Values

• SHAP Interaction Effects

• Partial Dependence

• Individual Conditional Expectation

• Counterfactual Explanations

• Actionable Recourse

• Immutable vs Mutable Variables

• Decision Path Analysis

• Explanation Stability

• Model-Agnostic Explainability

• Explainability for Neural Networks

• Explainability for Gradient Boosting

• Reason Codes

• Credit Decision Documentation

​

🧪 Exercise 75: AI Counterfactual Credit Decision Engine

Develop a credit decision engine that explains why an applicant was declined and calculates the minimum realistic changes required to become eligible.

Example:

Current Decision: Decline

The AI engine identifies:

• Debt-to-income ratio

• Utilisation

• Recent arrears

• Income stability

and generates alternative scenarios such as:

“Reducing revolving utilisation from 82% to 58% while maintaining current income would move the applicant above the approval threshold.”

The exercise combines SHAP + Counterfactual Optimisation + Actionable Recourse.

​

Module 32: Fairness, Bias Detection & AI Governance in

Credit Decisions

​

From Model Accuracy to Responsible Lending

AI credit models may unintentionally generate different outcomes across customer populations even when sensitive variables are excluded.

This module develops quantitative frameworks for identifying, measuring and mitigating bias without reducing model governance to a purely qualitative exercise.

Topics

• Algorithmic Bias

• Proxy Variables

• Fair Lending

• Group Fairness

• Individual Fairness

• Demographic Parity

• Equal Opportunity

• Equalised Odds

• Disparate Impact

• Calibration by Group

• Fairness–Accuracy Trade-off

• Bias Detection

• Reweighing

• Fairness-Constrained Optimisation

• Adversarial Debiasing

• Explainability and Fairness

• Human Oversight

• AI Governance for Credit Decisions

🧪 Exercise 76: AI Fair Credit Scoring Laboratory

Compare the approval and default performance of a model across several customer groups.

The student calculates:

Approval Rate

False Positive Rate

False Negative Rate

Default Rate

Calibration Error

and applies:

• Reweighting

• Threshold optimisation

• Fairness constraints

• Adversarial debiasing

The final task is to construct an Accuracy–Fairness Frontier rather than optimising accuracy alone.

 

Module 33: Causal Machine Learning for Credit Scoring & Lending Decisions

​

From Prediction to Cause-and-Effect

Traditional credit scoring answers:

“Who is likely to default?”

Causal Machine Learning addresses a different question:

“How will a lending decision change customer behaviour?”

This distinction is particularly relevant for limits, pricing, restructuring, collections and credit interventions.

Topics

• Prediction vs Causality

• Potential Outcomes Framework

• Treatment Effects

• Confounding

• Propensity Scores

• Matching

• Double Machine Learning

• Causal Forests

• Meta-Learners

• S-Learner

• T-Learner

• X-Learner

• Heterogeneous Treatment Effects

• Uplift Modelling

• Policy Learning

• Causal Feature Selection

• Lending Intervention Analysis

🧪 Exercise 77: Causal AI Credit Limit Optimisation

Estimate which customers would benefit from a credit-limit increase and which customers would become riskier because of it.

The model estimates:

Individual Treatment Effect

Δ Default Probability | Credit Limit Increase

The participant compares:

• Conventional propensity model

• Causal Forest

• Double Machine Learning

• Uplift model

The objective is no longer simply to predict PD, but to determine who should receive the intervention.

 

Module 34: Graph Neural Networks for Relational Credit Risk

Creditworthiness Beyond the Individual Borrower

​

Many borrowers are economically connected through ownership, transactions, supply chains, directors, households or corporate groups.

Traditional scoring treats observations largely independently. Graph Machine Learning allows the model to incorporate these relationships directly.

Topics

• Graph Theory for Credit Risk

• Nodes and Edges

• Borrower Networks

• Corporate Ownership Networks

• Transaction Networks

• Supply-Chain Links

• Household Relationships

• Network Centrality

• Community Detection

• Graph Embeddings

• Graph Convolutional Networks

• GraphSAGE

• Graph Attention Networks

• Link Prediction

• Node Classification

• Graph Explainability

• Contagion Effects

• Relational Credit Scoring

 

🧪 Exercise 78: Graph Neural Network Credit Scoring

Construct a network:

Borrowers ↔ Companies ↔ Directors ↔ Transactions ↔ Suppliers

A Graph Neural Network produces a borrower risk score using both:

Individual Characteristics

and

Network Relationships

The participant compares:

XGBoost Credit Score

versus

Graph Neural Network Score

and identifies cases where relational information reveals risk invisible to a conventional scorecard.

 

Module 35: Multimodal AI & Generative AI for Credit Underwriting Combining Structured Data, Documents and Text

​

Credit decisions increasingly use information that cannot be represented in a traditional tabular scorecard.

This module integrates structured borrower data with financial statements, application documents, management information and textual information through modern AI architectures.

Topics

• Multimodal Credit Scoring

• Structured vs Unstructured Data

• Document Intelligence

• Financial Statement Extraction

• Embeddings

• Transformer Models

• Large Language Models

• Retrieval-Augmented Generation

• Text Classification

• Semantic Features

• Alternative Data

• Document Consistency Checks

• Information Extraction

• Entity Recognition

• Multimodal Fusion

• Structured + Text Models

• Hallucination Controls

• Human-in-the-Loop Underwriting

🧪 Exercise 79: AI Credit Underwriting Agent

Build an AI underwriting workflow that receives:

• Application Data​

• Financial Statements​

• Borrower Documentation​

• Credit Bureau Information

The system extracts and validates information, creates structured risk features and sends them to a credit scoring model.

An AI Agent then produces:

• Risk drivers

• Missing information

• Inconsistencies

• Credit score

• Decision rationale

• Analyst review points

The LLM does not replace the quantitative model; it acts as an orchestration and document-intelligence layer.

​

Module 36: Federated Learning & Privacy-Preserving Credit Scoring Training AI Without Centralising Customer Data

​

Banks frequently face a conflict between the value of larger datasets and restrictions on moving sensitive customer information.

Federated Learning provides a framework where models can learn from distributed datasets without transferring the underlying raw data to a central repository.

Topics

• Federated Learning

• Horizontal Federated Learning

• Vertical Federated Learning

• Cross-Silo Learning

• Federated Averaging

• Distributed Credit Scoring

• Secure Aggregation

• Differential Privacy

• Privacy Budgets

• Model Updates

• Data Locality

• Non-IID Data

• Federated Model Drift

• Privacy vs Accuracy

• Federated Validation

• Multi-Bank Model Development

🧪 Exercise 80: Federated Credit Scoring Across Three Banks

Simulate three financial institutions:

Bank A + Bank B + Bank C

Each institution keeps its borrowers locally.

A federated algorithm trains a common model using distributed parameter updates.

Compare:

Local Models

vs

Federated Model

vs

Centralised Benchmark

using:

• AUC

• Gini

• Brier Score

• Calibration

• Privacy loss

The exercise demonstrates how model performance can improve without pooling raw customer records.

 

Module 37: Uncertainty-Aware Credit Decisions with Conformal Prediction

​

When the Model Should Say “I Am Not Certain”

Conventional credit scores normally produce a probability or score as if every prediction had similar reliability.

This module introduces modern uncertainty quantification so that the bank can distinguish between confident predictions and uncertain decisions requiring additional information or human review.

Topics

• Predictive Uncertainty

• Aleatoric vs Epistemic Uncertainty

• Conformal Prediction

• Conformal Classification

• Prediction Sets

• Coverage Guarantees

• Calibration

• Distribution Shift

• Selective Classification

• Abstention Models

• Confidence-Based Decisioning

• Out-of-Distribution Detection

• Uncertainty Thresholds

• Human Review Zones

• AI Decision Escalation

• Risk-Based Approval Architecture

🧪 Exercise 81: AI Selective Credit Decision Engine

Instead of producing only:

Approve / Decline

the model generates:

🟢 Approve — High Confidence

🟠 Manual Review — High Uncertainty

🔴 Decline — High Confidence

Using Conformal Prediction, the participant calculates prediction sets and automatically sends uncertain applications to manual review.

The exercise measures the trade-off between:

Automation Rate

Default Risk

Prediction Coverage

Manual Review Cost

This gives the course a very modern practical concept: AI that knows when not to make the decision automatically.

 

Module 38: Quantum Feature Selection & Hybrid Quantum-Classical Credit Scoring

​

Advanced Quantum Optimisation for Credit Risk

I would deliberately make the new quantum module different from the current course. You already have Quantum K-Means, QSVM, Quantum Feed-Forward Networks, Quantum CNNs and Tensor Networks. fermacrisk

Therefore this module would focus on a different problem: optimising the credit-scoring feature set and model structure through hybrid quantum-classical optimisation.

Topics

• Hybrid Quantum-Classical Modelling

• Quantum Optimisation

• QUBO Formulation

• Feature Selection as an Optimisation Problem

• QAOA

• Quantum Annealing

• Feature Redundancy

• Predictive Power

• Sparsity Constraints

• Model Complexity

• Quantum Optimisation Objective Functions

• Hybrid Classical Optimisers

• Quantum Feature Maps

• Quantum Kernels

• Noisy Intermediate-Scale Quantum Devices

• Classical vs Quantum Benchmarking

🧪 Exercise 82: QAOA Credit Scoring Feature Selection

Start with a dataset containing, for example, 100 candidate risk variables.

The objective is to select a smaller subset that simultaneously:

Maximises

Predictive Power + Stability

Minimises

Number of Variables + Redundancy + Model Complexity

The problem is transformed into a QUBO and solved using:

• Classical optimisation

• Genetic Algorithm

• QAOA / quantum simulator

The selected variables are then used to train the same credit-scoring model.

The participant compares:

Classical Feature Selection vs Genetic Optimisation vs Quantum QAOA Feature Selection

using:

AUC | Gini | Stability | Number of Variables | Computational Cost

This gives you a quantum exercise that is genuinely different from the four quantum credit-scoring exercises already present in the course.