Kreditriskmodellering under IFRS 9: En jämförande analys av prediktiva modeller för Probability of Default och deras påverkan på Expected Credit Loss

With the introduction of IFRS 9 in 2018, financial institutions became obligated to proactively forecast their Expected Credit Losses (ECL) through Probability of Default (PD), Loss Given Default (LGD), and Exposure at Default (EAD), marking a shift from the previously backward-looking approach. On behalf of zeb, this study focuses on the construction and evaluation of five different PD prediction models: Logistic Regression (LR, serving as a benchmark model), Support Vector Classification (SVC), Random Forest (RF), XGBoost, and Artifical Neural Network (ANN). In addition to evaluating model performance using PR AUC and ROC AUC metrics, the study applies a smoothing-based technique to assess how well each model's predicted PDs align with estimated true default probabilities. Lastly, a Monte Carlo simulation approach is used to assess the impact of each model on portfolio-level ECL estimates, offering further insights into their practical implications under IFRS 9. In summary, the results across all assessments indicated that no single model was universally superior within the constraints of the dataset. Nonetheless, all models outperformed the benchmark LR model, suggesting that more advanced machine learning techniques can yield improved predictive accuracy. Some models also exhibited distinct strengths in which the ANN demonstrated the best overall performance in PD classiciation, SVC produced the most well-calibrated probability estimates, and RF delivered the most accurate portfolio-level ECL estimates when compared to the simulated Realized Credit Loss.