We introduce AI Harmonics ( AIH ), a novel metric designed to quantify the concentration of harms across stakeholder groups affected by AI systems. Unlike traditional approaches that rely on arbitrary numerical assignments to ordinal severity levels, AIH provides a principled framework grounded in inequality theory, extending concepts from the Gini index to purely ordinal data. The metric evaluates how harm is distributed among stakeholders, capturing whether severe impacts are concentrated within specific groups or more evenly spread. Experiments on annotated incident data show that the proposed metric exhibits a strong monotonic relationship with the Criticality Index (CI), preserving harm category rankings while capturing additional variation in concentration patterns. The method demonstrates high robustness, with Spearman rank correlations above 0.97 under severity perturbations and stable prioritization even under up to 80% random data removal. Political and physical harms consistently exhibit the highest concentration, indicating the need for urgent mitigation. Political harms erode public trust, while physical harms pose serious, even life-threatening risks, underscoring the real-world relevance of our approach. The AIH metric is particularly well-suited for policy-making and risk management, where only ordinal assessments are available, and it enables more informed prioritization of mitigation strategies.

Measuring distances in multidimensional settings poses a significant challenge encountered across various scientific and engineering disciplines. In this paper, we introduce a novel measure of divergence to quantify the discrepancy between two multidimensional distributions—one predicted by a machine learning model and the other expected. Our approach builds upon the class of Energy Distances and incorporates a whitening pre-processing step, resulting in a divergence that is strictly connected to the new multivariate Gini index. To validate the proposed divergence, we demonstrate its effectiveness as a loss function for training a neural network designed to predict the financial performance of small and medium enterprises.

The growth of Artificial Intelligence applications based on Natural Language Processing requires to develop risk management models that can balance opportunities with risks, especially in high-stakes scenarios. In this paper, we contribute to the development of risk models presenting two integrated statistical metrics that can measure the "Accuracy" and the "Sustainability" of Artificial Intelligence models based on text processing, in line with the requests of international recommendations and regulations, such as the European Artificial Intelligence Act. The framework is validated through experiments on three distinct binary classification tasks on widely differing data sources. By highlighting vulnerabilities and strengths across diverse NLP pipelines, the proposed metrics provide a practical tool for assessing the reliability of AI applications. These contributions aim to foster safer deployment of AI technologies in finance, by mitigating risks of potential harms to financial stability.

The aim of this paper is to propose a structural model that can explain the reasoning underlying automated decisions and, in particular, their unfairness. Moving beyond black-box approaches, our model provides transparency and interpretability, enabling a deep understanding of decision-making processes. Specifically, we build a diffusion process to explain the inequality and unfairness in credit lending. We then compare the Gini index before and after the application of the model. A substantial reduction in the Gini index indicates that the diffusion process can explain unfairness.

This study aims to evaluate the impact of different binary classification methods for credit ratings on the overall performance of the model and to identify the optimal classification threshold. Four advanced machine learning models are employed: Random Forest (RF), Gradient Boosting Tree (GBT), Stacked Ensemble Model (SEM), and Voting Ensemble Model (VEM). To assess the performance of these models, SAFE metrics, based on the Rank Graduation Box (RGB) approach, are introduced to comprehensively measure model performance. Credit ratings are categorized into five binary classification schemes, wherein specific ratings from D to BBB are designated as high-risk (value = 1), while the remainder are classified as low-risk (value = 0). A systematic comparison of the models’ performance under different classification schemes is conducted using RGA, RGR, RGE, and RGF metrics, alongside traditional measures such as AUC, feature importance, and SHAP values. This research seeks to identify the classification method that most effectively explains the credit risk and its corresponding optimal threshold. The experimental results not only reveal the effectiveness of different classification methods, but also provide a theoretical basis for selecting the best threshold, thereby offering a more reliable and interpretable framework for credit risk assessment.

To date, several explainable AI methods are available. The variability of the resulting explanations can be high, especially when many input features are considered. This lack of robustness may limit their usability. In this paper we try to fill this gap, by contributing a methodology that: i) is able to measure the robustness of a given set of explanations; ii) suggests how to improve robustness, by tuning the model parameters. Without loss of generality, we exemplify our proposal for ensemble tree models, which typically reach a high predictive performance in classification problems. We consider a toy case study with artificially generated data as well as two real case studies whose application domain is cybersecurity and more precisely the models used for detecting phishing websites.

Financial institutions are increasingly leveraging on advanced technologies, facilitated by the availability of Machine Learning methods that are being integrated into several applications, such as credit scoring, anomaly detection, internal controls and regulatory compliance. Despite their high predictive accuracy, Machine Learning models may not provide sufficient explainability, robustness and/or fairness; therefore, they may not be trustworthy for the involved stakeholders, such as business users, auditors, regulators and end-customers. To measure the trustworthiness of AI applications, we propose the first Key AI Risk Indicators (KAIRI) framework for AI systems, considering financial services as a reference industry. To this aim, we map the recently proposed regulatory requirements proposed for Artificial Intelligence Act into a set of four measurable principles (Sustainability, Accuracy, Fairness, Explainability) and, for each of them, we propose a set of interrelated statistical metrics that can be employed to measure, manage and mitigate the risks that arise from artificial intelligence. We apply the proposed framework to a collection of case studies, that have been indicated as highly relevant by the European financial institutions we interviewed during our research activities. The results from data analysis indicate that the proposed framework can be employed to effectively measure AI risks, thereby promoting a safe and trustworthy AI in finance.

Technological advancements have prompted the emergence of peer-to-peer credit services which improve user experience and offer significant reductions in costs. These advantages may be offset by a higher credit risk, due to disintermediation and information asymmetries. We postulate that network-based information can be employed as a tool for reducing risks through an improved credit scoring model that increases the accuracy of default predictions. Our research assumption is proven by means of empirical analysis that shows how including network parameters in classical scoring algorithms, such as logistic regression and CART, does indeed improve predictive accuracy.