The pandemic has been for banks’ credit departments a textbook case game-changer, and we still haven’t seen all of its effects: a real accelerator in credit risk modeling for banks willing to know which company to support first in times where moratoria and public guarantees may hide the real current situation of the borrower and postpone the showdown with NPLs.
This message was at the foundation of the last Prometeia virtual session about ‘Improving PD models predictivity and reactivity with transactional information and Machine Learning’, where we presented our recent project developed with Unicredit - one of the biggest European banking institutions – to create a PD model for the small business segment. A process that leveraged more than 550 million transactions, using advanced text analytics techniques to categorize free-text transaction purposes, and machine learning to turn categorized transactions into a score for each firm.
A cutting-edge model that is proving to be very performing during the pandemic, even stand-alone, and very reactive, able to capture deterioration signals right from the start of lockdowns and restrictive measures. A strong real-time component, differentiating transactional credit scores from more traditional ones, that is allowing banks to more effectively detect companies that do need the bank’s support even without a visible credit deterioration.
The use case presentation, by Emanuele Giovannini (Unicredit) and Giangiacomo Sanna (Prometeia), preceded a high-level discussion among some very qualified matter experts in the field: Chiara Capelli, head of credit risk modeling at Unicredit, Dmitri Kraynov, head of risk modeling at Sberbank, Sid Dash, research director at Chartis Research, and Marco Stella, Prometeia’s partner.
Among other issues, the reactivity of transactional and AI-based credit risk models was cited as one of the most crucial features amid the pandemic and accelerated digitalisation. The increase in predictive power, as well, is a big advantage, especially for clients usually leaving little ‘official’ real-time information about themselves (small businesses, individuals).
Not surprisingly, the use of transactional data in risk modeling has lately been strongly recommended also by the ECB, the Eurozone regulator about banking risk models.
Such an implementation implies several challenges, of course, both from an ethical to a technological standpoint: for instance, organizing such broad sets of data requires a great computational power. But probably – this has been one of the loudest messages of the panel – the main challenges are around the explainability and interpretability of models.
Beyond being regulatory and business needs, these features have proved to be a modeler’s exigence as well: especially in times of pandemic, adapting the credit rating model has been possible just thanks to its interpretability, enabling to net it from all variables that were not reliable any more (like payment of taxes in Italy, where they have been postponed by authorities, and so lost their predictive meaning).
Moreover, interpretability has been said to be crucial to change the attitude of many internal and external stakeholders of the bank, who are extremely diffident towards black boxes. This stance makes investing in explainability even more compelling.
One thing is for sure. The road traced by these complex techniques can open up new business horizons for intermediaries. Think, for example, of the extension from transactional data to PSD2 and open banking data. A wealth of information that, framed in a model that can be explained and interpreted in every single item, could mark the new frontier for advanced loan origination and monitoring processes.