ETUDE INTERNE
AUTEUR : HAMZA EL YOUMNI & AZIZ MAAOUIA
In the realm of credit risk management, financial institutions are used to assess borrowers’ repayment capacity through the construction of a credit scoring. To do so, they commonly rely on logistic regression models due to their simplicity and their clear interpretability. However, recent research has highlighted that logistic regression may be less performing than alternative machine learning algorithms.
Although these algorithms offer higher prediction performance, their outputs often lack explicit interpretability.
To address this issue, we propose in this paper to use a interpretable algorithm, the Catboost model. Nevertheless, the classic framework of backtesting used by financial institution is not fully adapted to assess this new scoring framework, particularly in terms of effectively controlling the contributions of variables. To accommodate this novel approach, we propose and compare two methodologies to adapt the standard backtesting.
