My dissertation
Creating Value with Artificial Intelligence in Banking: Data, Processes, Models, and Strategy
Humboldt University of Berlin · 2026 · summa cum laude
Read the dissertation in the HU repositoryAI Value Chain
The AI Value Chain connects four studies by asking what each method can establish: what information can be extracted, how it enters a workflow, when models can transfer to other data and where evidence for strategic use is still sparse. The perspectives overlap. For instance, how data are represented matters both for reading documents and for transferring models across portfolios.
Data representation
Industry codes capture only broad differences in what individual companies do. I developed Company2Vec to represent German companies as vectors, using Word2Vec and dimensionality reduction on corporate website content. In the study, these vectors captured similarities at a finer level than NACE industry codes and made it possible to group comparable firms.
Company2Vec — German Company Embeddings Based on Corporate WebsitesProcess automation
In company register extracts, a piece of text can have different meanings depending on where it appears on the page. We fine-tuned LayoutXLM to classify parts of these documents and ran ablation studies to isolate the contributions of text, layout and image features. Layout substantially improved classification in this case study, and the model achieved an F1 score above 0.75 with 250 labelled pages for fine-tuning.
Multimodal Document Analytics for Banking Process AutomationModel adaptation
Defaults are rare in many loan portfolios, and the source and target portfolios can differ in their features and distributions. For the preprint, we developed a Transformer for tabular data that can use different sets of features to provide loan-level recovery estimates and portfolio-level predictive distributions. With few target observations, it outperformed comparison models in the experiments, especially when input distributions or the relationship between inputs and recoveries changed, while shifts in the distribution of recoveries remained difficult.
Transfer Learning for Loan Recovery Prediction under Distribution Shifts with Heterogeneous Feature SpacesStrategic use
Evidence on which bank marketing decisions can benefit from text analysis is scattered. We systematically reviewed 109 studies and mapped the findings to the customer journey and the marketing mix. Only eight of the studies directly addressed language processing in bank marketing, mainly customer retention, while acquisition, personalised engagement and external text sources received less attention.
Natural Language Processing in Bank Marketing: A Systematic Review and Strategic Gap AnalysisOpen research questions
- When does transfer learning help with few target observations, and which differences between source and target erase that advantage?
- When does an additional modality, such as layout or images, justify the cost of data collection, annotation and operation?
- How can we assess usefulness beyond model performance, for example by measuring manual review effort or the quality of downstream decisions?