The estimation of Loss Given Default (LGD) is a critical component in credit risk modelling, directly influencing the calculation of expected losses and capital requirements for financial institutions. Traditionally, linear regression models have been employed for LGD prediction. However, more advanced methods, such as Extreme Gradient Boosting, Neural Networks and Random Forests significantly, outperform this standard approach in terms of predictive accuracy. In particular, tree-based models emerge as reliable and robust, minimizing the error of the predictions. Despite their superior performance, their complexity often limits transparency, raising concerns in highly regulated environments, where regulators expect banks to employ transparent and auditable risk models. This study demonstrates that financial institutions do not need to compromise on prediction accuracy to meet regulatory requirements for model transparency. To address these challenges, various Explainable Artificial Intelligence (XAI) techniques are explored, including SHAP, Expected Gradients, Integrated Gradients, ALE plots, LIME and Permutation Feature Importance. These methods are assessed based on their ability to verify economic coherence, feature importance and provide local explanations for misclassified observations. Through this analysis, it is shown that it is possible to identify the most influential factors driving LGD predictions and ensure that the models align with economic theory. Moreover, the local explanations offer detailed interpretations of individual predictions, enhancing the model's trustworthiness and facilitating its practical implementation in credit risk management. By testing and comparing different techniques, the findings reveal that SHAP for the ensemble methods and Expected Gradients for the Neural Network model are robust and efficient across all areas of application. This research highlights practical applications of XAI techniques, showing how they can be effectively used to reconcile regulatory requirements with model performance in LGD modelling. Overall, the superior predictive performance of black-box models can coexist with the need for explainability through the application of XAI methods.
La stima della Loss Given Default (LGD) rappresenta un elemento fondamentale nei modelli di rischio di credito, influenzando direttamente il calcolo delle perdite attese e dei requisiti patrimoniali per le banche. Tradizionalmente, la LGD viene stimata attraverso modelli di regressione lineare. Tuttavia, metodi più avanzati, come Extreme Gradient Boosting, Neural Network e Random Forest, offrono prestazioni significativamente superiori rispetto a questo approccio standard in termini di accuratezza predittiva. In particolare, i modelli basati su alberi decisionali si distinguono per la loro affidabilità e robustezza, minimizzando l'errore delle previsioni. Nonostante le loro elevate prestazioni, la complessità di questi modelli può limitarne la trasparenza, sollevando preoccupazioni in contesti altamente regolamentati, dove sono richiesti modelli di rischio trasparenti e verificabili. Questo studio dimostra che le istituzioni finanziarie possono mantenere un'elevata accuratezza predittiva senza compromettere la trasparenza richiesta dai regolatori. Per affrontare questa sfida, vengono esaminate diverse tecniche di Intelligenza Artificiale Spiegabile (XAI), tra cui SHAP, Expected Gradients, Integrated Gradients, ALE plots, LIME e Permutation Feature Importance. Questi metodi sono valutati in base alla loro capacità di verificare la coerenza economica del modello, l'importanza delle variabili e di fornire spiegazioni locali per specifiche osservazioni. L'analisi dimostra che è possibile identificare i fattori più influenti nelle previsioni di LGD e garantire l'allineamento dei modelli con la teoria economica. Inoltre, le spiegazioni locali offrono interpretazioni dettagliate delle singole previsioni, migliorando l'affidabilità del modello e facilitando la sua applicazione pratica nella gestione del rischio di credito. Confrontando diverse tecniche, i risultati evidenziano che SHAP, per i metodi basati su alberi decisionali, e Expected Gradients, per il modello di Reti Neurali, risultano robusti ed efficienti in tutte le aree di applicazione. Questa ricerca evidenzia le applicazioni pratiche delle tecniche XAI, dimostrando come possano essere utilizzate efficacemente per conciliare i requisiti normativi con le prestazioni dei modelli nella modellizzazione della LGD. In sintesi, le elevate prestazioni predittive dei modelli più avanzati possono coesistere con l'esigenza di spiegabilità grazie all'applicazione di metodi XAI.
Balancing Transparency and Performance: Explainable AI for Loss Given Default
RASCHI, SOFIA
2023/2024
Abstract
The estimation of Loss Given Default (LGD) is a critical component in credit risk modelling, directly influencing the calculation of expected losses and capital requirements for financial institutions. Traditionally, linear regression models have been employed for LGD prediction. However, more advanced methods, such as Extreme Gradient Boosting, Neural Networks and Random Forests significantly, outperform this standard approach in terms of predictive accuracy. In particular, tree-based models emerge as reliable and robust, minimizing the error of the predictions. Despite their superior performance, their complexity often limits transparency, raising concerns in highly regulated environments, where regulators expect banks to employ transparent and auditable risk models. This study demonstrates that financial institutions do not need to compromise on prediction accuracy to meet regulatory requirements for model transparency. To address these challenges, various Explainable Artificial Intelligence (XAI) techniques are explored, including SHAP, Expected Gradients, Integrated Gradients, ALE plots, LIME and Permutation Feature Importance. These methods are assessed based on their ability to verify economic coherence, feature importance and provide local explanations for misclassified observations. Through this analysis, it is shown that it is possible to identify the most influential factors driving LGD predictions and ensure that the models align with economic theory. Moreover, the local explanations offer detailed interpretations of individual predictions, enhancing the model's trustworthiness and facilitating its practical implementation in credit risk management. By testing and comparing different techniques, the findings reveal that SHAP for the ensemble methods and Expected Gradients for the Neural Network model are robust and efficient across all areas of application. This research highlights practical applications of XAI techniques, showing how they can be effectively used to reconcile regulatory requirements with model performance in LGD modelling. Overall, the superior predictive performance of black-box models can coexist with the need for explainability through the application of XAI methods.File | Dimensione | Formato | |
---|---|---|---|
2024_10_Raschi_Tesi.pdf
non accessibile
Descrizione: Tesi
Dimensione
3.84 MB
Formato
Adobe PDF
|
3.84 MB | Adobe PDF | Visualizza/Apri |
2024_10_Raschi_ExecutiveSummary.pdf
non accessibile
Descrizione: Executive Summary
Dimensione
662.36 kB
Formato
Adobe PDF
|
662.36 kB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/227242