Last week, our CTO Burkhard Heppe went back to his alma mater at the University of Cambridge to deliver a guest lecture on the use of AI in credit sales and management accompanied by our data scientist Theodore Gavriilidis. Master’s students of the Economics department followed the lecture as part of the course on AI in Finance. Burkhard presented a series of case studies on how NPL Markets uses machine learning and generative AI tools to facilitate credit sales by identifying suitable investors for non-performing loans, support the migration of complex data sets from seller to buyer or by classifying documents in the virtual data room during due diligence and extracting key values. Theodore highlighted the importance of pipeline implementation and deployment and explained the potential approaches to fine tuning large language models (LLM) to credit trading specific use cases.
Another LLM use case in our loan reporting activities is consistency check between different reports. For example, for securitisation transactions, regulatory disclosures using the ESMA templates may differ from the investor reports published as Excel or pdf files. NPL Markets has nearly EUR 100 billion equivalent of loan portfolios under reporting and valuation assignments and AI tools help with report automation and validation. The extraction and monitoring of key performance indicators, credit covenants, replenishment, and performance measures with the help of locally deployed AI tools for maximum safeguards of client data has benefitted greatly from recent technological progress.
While generative AI and LLM have a growing number of applications in finance and credit markets, the analysis of structured data sets using machine learning (ML) tools has not diminished in importance. ML tools help with predicting the timing and amount of risk- adjusted loan cash flows as part of a loan valuation exercise. However, the accuracy of the predictions will be less sensitive to the choice of model and more sensitive to the available calibration and test data. Especially for the prediction of recovery cashflows on non-performing loans training data might be scarce and subject to selection bias highlighting the need for larger reference data sets and data hubs.
