Credit Research Centre

Author: Andrija Djurovic

Testing the Concentration in the Rating Grades

This set of slides outlines the basic metrics for measuring Concentration among PD risk grades, and the European regulatory requirements for reporting them. Andrija’s presentation gives helpful commentary to calculate these metrics. This should be read and explored by all students of quantitative risk management and by model managers, as an example of best practice and regulatory requirement for Credit Risk IRB.

IMF’s Partial Portfolio

This folder contains Andrija’s presentation summary of the IMF’s Partial Portfolio Approach to Concentration Risk, with illustrating data and code in R and Python. The IMF’s hybrid model of portfolio credit risk has a structure that maintains good performance even in the presence of high name-exposure concentrations, and therefore is an important model option and a mitigation of Concentration Risk. This should be read by model developers and model risk managers as an example where model structure has mitigated a potential model error, so helping manage the specification risk of the model.

Author: Denis Burakov

Scorecard Boosting Notebooks

This repository features a comprehensive collection of Jupyter Notebooks and Python scripts focused on scorecard boosting. It includes all relevant materials and developments by the original author, Paul Edwards, providing a detailed guide to the topic.

Scorecard Boosting Package (xbooster)

This repository offers an open-source Python implementation of scorecard boosting tailored for credit risk modeling. The package facilitates the creation of interpretable scorecards using XGBoost logistic regression and includes tools designed for risk management explainability.

Collection of Various Notebooks on Credit Scoring with Machine Learning

This collection encompasses a wide range of topics within credit scoring, such as class-imbalance learning, machine learning fairness, and experiments with XGBoost using custom loss functions, alongside Naive Bayes scorecards.

Documentation of Boosted Scoring Package

This document outlines the scorecard boosting package, detailing the key steps involved in constructing a scorecard from gradient boosted decision trees (GBDT). It also covers practical use cases, including handling categorical data and deploying a boosted scorecard with SQL.

Author: Denis Burakov

Collection of Various Implementation of Algorithms

This compilation includes various implementations of machine learning modelling techniques and optimization algorithms. It covers binary logistic regression (with methods such as stochastic gradient descent, Fisher scoring, weighted, and focal loss), multi-class classification using Softmax, and neural networks (including perceptrons).

Author: Joseph Breeden

Specification Errors in Stress Testing Models

This paper highlights how stress testing models for retail lending often suffer from inherent specification errors when incorporating lifecycle, credit quality, and macroeconomic effects. Drawing parallels to Age-Period-Cohort (APC) models, the authors show that these interdependent factors create ambiguity in trend attribution, which limits model reliability, especially for out-of-sample stress testing.

Incorporating lifecycle and environment in loan-level forecasts and stress tests

This paper introduces a loan-level Age-Vintage-Time (AVT) modeling approach to forecast credit risk by separating lifecycle effects, macroeconomic environment, and loan vintage quality. It extends Age-Period-Cohort (APC) models with retrending techniques and incorporates them into a Generalized Linear Model (GLM) framework. The work demonstrates, using a US auto loan dataset, that this method not only improves out-of-sample and out-of-time prediction accuracy but also provides stable and robust loan-level forecasts and stress tests, aligning with regulatory requirements like IFRS 9 and CECL.

Forecasting Behavior with Age-Period-Cohort Models: How APC Predicted the US Mortgage Crisis, but Also Does So Much More

This paper introduces Age-Period-Cohort (APC) models and demonstrates their versatility beyond traditional use in epidemiology and demography. It shows how APC models not only explained the U.S. mortgage crisis by decomposing loan performance into lifecycle, vintage, and environmental effects, but also apply to diverse areas such as wine price forecasting, user behavior, and tree-ring analysis for climate change