Explainable AI in Fintech and Insurtech

The growing application of black-box Artificial Intelligence algorithms in many real-world application is raising the importance of understanding how models make their decision. The research field that aims to look into the inner workings of the black-box and to make predictions more interpretable is referred to as eXplainable Artificial Intelligence (XAI). Over the recent years, the research domain of XAI has seen important contributions and continuous developments, achieving great results with theoretically sound applied methodologies. These achievements enable both industry and regulators to improve on existing models and their supervision; this is done in term of explainability, which is the main purpose of these models, but it also brings new possibilities, namely the employment of eXplainable AI models and their outputs as an intermediate step to new applications, greatly expanding their usefulness beyond explainability of model decisions. This thesis is composed of six chapters: an introduction and a conclusion plus four self contained sections reporting the corresponding papers. Chapter 1 proposes the use of Shapley values in similarity networks and clustering models in order to bring out new pieces of information, useful for classification and analysis of the customer base, in an insurtech setting. In chapter 2 a comparison between SHAP and LIME, two of the most important XAI models, evaluating their parameters attribution methodologies and the information they are capable of include thereof, in italian Small and Medium Enterprises’ Probability of Default (PD) estimation, with balance sheet data as inputs. Chapter 3 introduces the use of Shapley values in feature selection techniques, with the analysis of wrapper and embedded feature selection algorithms and their ability to select relevant features with both raw data and their Shapley values, again in the setting of SME PD estimation. In chapter 4, a new methodology of model selection based on Lorenz Zoonoid is introduced, highlighting similarities with the game-theoretical concept of Shapley values and their variability decomposition attribution to independent variables as well as some advantages in terms of model comparability and standardization. These properties are explored through both a simulated example and the application to a real world dataset, provided by EU-certified rating agency Modefinance.