Transfer Learning for Loan Recovery Prediction under Distribution Shifts with Heterogeneous Feature Spaces
arXiv
Loan recovery rates are hard to predict when few defaults have been observed in the target portfolio and the available features differ across portfolios. We develop a tabular Transformer with a mixture-density output and test it in simulations and in transfer from corporate loans to bonds. The gains are clearest with few target observations, overlapping features and sufficiently similar recovery distributions, while pronounced shifts in the recovery distribution limit the benefit and the loan-to-bond gains are modest.
Natural Language Processing in Bank Marketing: A Systematic Review and Strategic Gap Analysis
International Journal of Bank Marketing
Research on language processing in bank marketing is scattered across disciplines. We systematically review 109 studies and use text embeddings to map themes, comparing the evidence across the customer journey and the marketing mix. Only eight studies in the review directly address language processing in bank marketing, with retention receiving more attention than acquisition, personalised engagement and the use of external text sources.
Multimodal Document Analytics for Banking Process Automation
Information Fusion
When extracting information from German company register extracts, both the words and their position and appearance on the page matter. We fine-tune LayoutXLM to classify individual tokens and test how performance changes when layout or image features are removed or when fewer labelled pages are available. In this case study, layout supplied the larger gain beyond text, image features added a smaller gain, and fine-tuning on 250 labelled pages yielded an F1 score above 0.75.
Company2Vec — German Company Embeddings Based on Corporate Websites
International Journal of Information Technology & Decision Making
Industry codes group companies into categories but give little detail on how similar individual firms are. Company2Vec uses Word2Vec and dimensionality reduction to turn German corporate website content into vectors for peer comparison and clustering. The vectors allow finer comparisons than NACE categories alone, though their accuracy in predicting NACE codes falls at finer levels of the classification.