Explainable AI Applied to Dynamic Credit Risk Models

About PwC

The PwC network spans the globe and employs nearly 300,000 people. As part of the global PwC Network, PwC Germany is well-established in this country having 21 locations and 13.000 employees while we are constantly on the lookout for new colleagues who are enthusiastic about innovation. Everyday PwC Germany is confronted with new challenges and issues, which we always address with our values in mind: We act with integrity, we make a difference, we work together, we care for others, and we reimagine the possible day by day.

In our subcluster 'CDO' in Financial Services Technology Consulting, we advise our clients - multinational banks, insurance companies, and finance-oriented corporates - on the digital transformation of their projects. This includes the areas of Data & Analytics, Big Data, Artificial Intelligence, Blockchain Technologies, and Credit Risk.

About the project
Are you interested in predicting the Probability of Default with an AI model that can replace traditional models and meet regulatory requirements with explainable AI?

Institutions like banks and insurances spend thousands of hours building regulatory-compliant credit risk models like PD (Probability of Default), LGD (Loss Given Default), or EaD (Exposure at Default). At the end of the day, a handful of risk drivers are included in the final model. These risk drivers are typically selected with a linear or logistic regression. These regression models involve the use of an optimization algorithm to find a set of risk drivers for each input to the model that minimizes the prediction error. The model performance is evaluated using measures like Somers'D (Gini-Coefficient). Before calculating this value, data transformation steps are performed to enhance data stability and comparability. Highly correlated risk drivers should not be included in the model at the same time. Machine Learning algorithms should be expected to recognize such correlations and even find interrelationships that cannot be found with the traditional statistical approach.

To address this problem in case of a PD model, we utilize a historical dataset containing risk drivers and a 'Default Flag' column indicating entity defaults at a specific point in time. Based on this information, we develop a deep learning network to establish a model and assign the Probability of Default to each entity in the dataset.