Research projects

Project CEMAPRE internal

TitleInterpretable models of loss given default
ParticipantsJoão Bastos (Principal Investigator), Sara M Matos
SummaryMachine learning based predictive techniques are increasingly being adopted in Finance, and
particularly in the area of Credit Risk. However, due to their complexity and high number of degrees
of freedom, their predictions are often difficult to explain and validate. This is referred to as
the “black box” problem. In the area of Credit Risk, it is important that the model decisions
can be inspected and interpreted, not only to guarantee that the deployed models are compliant with
the prudential requirements defined by the regulators, but also to detect any potential bias and
discrimination. In this project we address the problem of the interpretability of machine learning
models to predict the loss given default: the proportion of outstanding credit that is lost by a
lender when the debtor defaults. This is a crucial parameter to calculate the amount of regulatory
capital a financial institution is required to hold to ensure that it does not take on excess
leverage and become insolvent.