Among bank transaction data, quantitative signals — balances and cash movements — are already used as features in credit-risk models, while qualitative signals — what the transaction was for, and with whom — remained unusable, defeated by the many-dialect problem of “bank language.” Applying a domain-specialized LLM to label those qualitative signals, we confirm they can be structured at roughly human-level accuracy — opening a path to bridge them into features a credit model can consume.

Predicting corporate bankruptcy from bank account balances and cash movements. A Bayesian hierarchical model captures segment-level heterogeneity by sales size and firm age, outperforming traditional logistic regression. Published in the International Journal of Financial Engineering.