Modern Banking: The Role of AI in Fraud Detection, Credit Assessment and Portfolio Optimization
Author
Bjerring, Anders
Term
4. semester
Publication year
2025
Abstract
This thesis examines how advanced AI techniques compare to traditional statistical and machine learning methods across three core banking domains: fraud detection, credit risk assessment, and portfolio optimization. Using structured datasets that emulate real banking environments, it evaluates ensemble models, deep learning architectures, and applications of generative AI with attention to predictive performance as well as practical factors such as explainability, organizational readiness, and regulatory compliance. Empirical findings indicate that AI-based models, particularly ensembles, generally outperform logistic regression on classification tasks. In contrast, results for forecasting tasks used in portfolio construction were more mixed: deep learning sometimes reduced prediction errors but did not consistently translate into better portfolios, underscoring the challenges of financial forecasting and the constraints of market efficiency. SHAP was applied to enhance transparency, though its reliability remains a concern in high-stakes settings. Generative AI appears presently more useful for internal operations and productivity-focused functions than for predictive financial modeling. Within Western banking, where AI adoption is growing but cautious, interpretability, model risk, and evolving regulation temper deployment. The findings are not broadly generalizable; overall, realizing value from AI requires balancing accuracy with explainability, governance, and sector-specific conditions.
Specialet undersøger, hvordan avancerede AI-metoder klarer sig i forhold til traditionelle statistiske og maskinlæringsmetoder i tre centrale bankdiscipliner: svindeldetektion, kreditvurdering og porteføljeoptimering. Med brug af strukturerede datasæt, der efterligner virkelige bankmiljøer, evalueres ensemblemodeller, dybe læringsarkitekturer og anvendelser af generativ AI ud fra både forudsigelsesydelse og praktiske hensyn som forklarbarhed, organisatorisk parathed og regulatorisk overensstemmelse. Empiriske resultater indikerer, at AI-baserede modeller, især ensembler, generelt overgår logistisk regression i klassifikationsopgaver. For prognoseopgaver i porteføljekonstruktion var resultaterne mere blandede: Dybe modeller gav til tider lavere fejl, men dette omsattes ikke konsekvent til bedre porteføljer, hvilket understreger udfordringerne ved finansielle forudsigelser og markedsefficiens. SHAP blev anvendt for at øge transparensen, men pålideligheden er fortsat en udfordring i beslutninger med høj risiko. Generativ AI vurderes p.t. mere nyttig til interne arbejdsprocesser og produktivitetsopgaver end til forudsigende finansiel modellering. Inden for vestlig banksektor, hvor AI gradvist integreres, begrænses udbredelsen af bekymringer om fortolkelighed, modelrisiko og skiftende regulering. Fundene er ikke bredt generaliserbare, og specialet konkluderer, at realisering af værdi kræver en balance mellem nøjagtighed, forklarbarhed, governance og sektorspecifikke rammevilkår.
[This apstract has been generated with the help of AI directly from the project full text]