Interpretable AI in banking

Language models and their pitfalls

Artificial intelligence has long since arrived in everyday life, from credit scoring and fraud detection to chatbots in customer service. In public perception, however, the spotlight is often on ChatGPT & Co. and thus on a special AI model. The use of such language models also has many disadvantages that users should be aware of. To be more precise: GPT (Generative Pretrained Transformer) is a language model that generates texts using the transformer architecture. For each input, it estimates the probability of the next word, calculates a probability distribution over all possible tokens, selects a token from this and appends it to the sequence. This extended sequence is run through the model again to predict the next word, iteratively up to the maximum length or until an end-of-sequence token appears. Such large language models can be used in many areas of a bank where texts are analyzed or created. 

The models are trained on large text datasets. They are suitable for writing texts, for example, but reach their limits when it comes to strictly logical tasks, such as providing exact proofs or performing complex calculations. Roughly speaking, the model does not „think“ in symbols or formal rules, but works with probabilities and patterns. As a result, answers can sound plausible without necessarily being logically correct. For this reason, critical decisions in the banking sector, such as the assessment of risks and the prediction of loan defaults, should not be based on language models. 

Specialized models

This is precisely where other, preferably interpretable methods come in, such as Generalized Matrix Learning Vector Quantization (GMLVQ), a prototype-based classification method. Due to its robustness and ability to deliver interpretable results, GMLVQ has proven itself in many practical classification problems. Fields of application in banking arise wherever classification problems occur. Two publications published in 2025 [1] and [2] show a special application of GMLVQ in the evaluation of network reliability. In these publications, reliability levels of consecutive k-out-of-n systems are classified with GMLVQ. The approach is generally transferable to coherent systems and can even be adapted to estimate the union probability of finitely many events from single and pairwise intersection probabilities. This underlines the breadth of application of GMLVQ. A particularly exciting area of application for GMLVQ in the future could be in the evaluation of credit default probabilities or risk modeling. 

As a project partner, TeleskopEffekt GmbH supports the introduction of GMLVQ as an interpretable AI method, from the idea in the specific use case to data evaluation, and illustrates the advantages over alternative AI methods. For further insights and individual advice, our colleague Dr. Mandy Lange-Geisler and can be contacted at any time.

Sources

[1] K. Dohmen, M. Lange-Geisler, and T. Villmann. Network reliability analysis by means of generalized matrix learning vector quantization, 2025. research square (https://www.researchsquare.com/article/rs-7014031/v1) 

[2] K. Dohmen, M. Lange-Geisler, and T. Villmann. Learning of probability estimates for system and network reliability analysis by means of matrix learning vector quantization. In ESANN 2025, 2025. 6 pp.(https://www.esann.org/sites/default/files/proceedings/2025/ES2025-67.pdf)