Dall'Accuratezza alla Redditività: Valutare l'Impatto Economico dei Modelli di Rating del Credito

This study focuses on the economic value of various predictive accuracy metrics in credit rating models. The general logic of the banks is that they rely on regression-based approaches, while more recently, it’s common to see the use of machine learning techniques to assess borrower risk. However, the question of whether investing in higher-performing models or not generates countable financial benefits remains underexplored, and most importantly, underanalysed. This paper addresses this gap by examining how more enhanced discriminatory power in models affects not only profitability, but also the lending quality and regulatory capital requirements of the banks. The analysis identifies three primary transmission channels, that are explained thoroughly in the main text, through which the models’ accuracy can influence different economic outcomes given that: (1) improved loan origination reduces defaults and enhances margins by better identifying low-risk borrowers; (2) stronger models mitigate adverse selection, a very vast sector in banking, helping retain creditworthy clients who might be lost to competitors otherwise; and (3) more accurate models, so more higher value metrics and, by extension, risk assessments can reduce Risk-Weighted Assets (RWA), freeing regulatory capital. In order to address this request in a more direct way, we are using simulation-based methods, generating synthetic loan portfolios (50,000 loans at 3% default and 10,000 prospects at 10% default) and evaluating models across AUROC bands from 65% gradually increasing to 90%. In order to do that, based on references in the bibliography, we are confident to use different logistic distributions that were applied 1 to mimic predictive scores, and they were calibrated to ensure consistent risk levels. In the end, the results show that defaults among top-approved loans decline sharply with better accuracy models - from nearly 6% at AUROC with 65% accuracy to less than 1% at AUROC with 90%. Proceeding to the adverse selection analysis, we can confirm that stronger models attract and retain more profitable clients. The capital impact is smaller but meaningful, with average RWA reductions of around 8% between lower-and higher-accuracy scenarios. Lastly, the profitability that was measured from the previous analysis gives further value to the model improvements. By applying a realistic arithmetical example, on a €3.5 billion retail portfolio, each 5-point AUROC percentage increase can generate approximately €0.8-1 million in addition to the annual profit, with relative gains of 5-12% depending on competitive dynamics. These effects can compound over time as new loans are added annually, while the findings show us that even with incremental improvements in model discrimination can yield and generate significant economic returns, reinforcing the strategic importance of continuous model enhancement. Banks, regulators, and model developers at the same time can use these insights to firstly justify investments, then set performance benchmarks, and also better understand the link between model validation metrics and real-world financial outcomes.