Year: 2026
Type: dissertation
Abstract: This thesis develops data-driven methodologies to support decision making in financial services and actuarial applications. Across three studies, the thesis examines how reinforcement learning, counterfactual learning, risk-aware criteria, and alternative data can support credit allocation and insurance risk prediction. The first study focuses on credit limit adjustment decisions in a credit card portfolio. The central question is whether a customer’s credit line should be increased or maintained, a decision that affects both access to credit and expected losses. The problem is formulated as an expected profit maximization task that balances revenue and provisions, and using data from a Latin American super-app an offline reinforcement learning framework is developed. The results show that a tuned Double Q-learning agent outperforms competing strategies, including the institution’s policy. They also show that alternative data do not always improve balance prediction accuracy. The second study extends the credit limit problem by considering not only whether an increase should be granted, but also how large it should be. This setting is formulated through counterfactual learning to assess whether potential outcome estimation alone is sufficient for optimal policy selection. To account for model uncertainty, risk-aware treatment selection criteria based on Value-at-Risk and Conditional Value-at-Risk are occupied. Semisynthetic experiments show that these criteria produce lower policy error than recommendations based only on point estimates. The third study examines the value of geographic information for zone-level claim frequency prediction in Motor Third-Party Liability insurance. Using BeMTPL97, the study evaluates model performance in postcodes not observed during training and incorporates alternative geographic data into the prediction task, including environmental features from OpenStreetMap and CORINE Land Cover, as well as historical orthoimagery. The results show that these features improve predictive accuracy, especially when combined with coordinates at a 5 km scale. When structured environmental features are absent, pretrained vision-transformer embeddings also improve accuracy and stability. Overall, this thesis contributes to credit allocation by modelling credit limit decisions as both a reinforcement learning problem and a counterfactual dosage problem. It also contributes to insurance risk modelling by showing how constructed geographic information can improve claim zone-level frequency prediction within an interpretable actuarial framework.
Author: Alfonso Sánchez, Sherly Paola
Language: English
Subfield: Accounting
Domain: Social Sciences
Open Access status: green
OpenAlex ID: W7208804607