Last week I took part in the Credit Scoring and Credit Control XVI conference in Edinburgh. The conference is known to be Europe’s premier convention for credit scoring and related topics and is one of the most important credit conferences in the world.
More than 450 participants from 45 countries took part this year. The conference programme included an impressive number of panels: touching topics ranging from IFRS9 to Machine Learning (ML) applied to credit risk modelling, and even new data sources such as open banking transactional data. The conference was well-attended both by academia as well as by the industry.
It was also my first time in Scotland and, obviously, my first time in Scotland during Brexit. During the Gala Dinner, Brexit was always present: dancing on the tongues of the people at the table, or when Professor Jonathan Crook (Professor of Business Economics) during his closing speech slayed the whole room with his English humor.
Another topic as frequent as Brexit was ML applied to credit risk modelling. Many panels somehow touched the topic, each offering a different perspectives. Just to mention some panel titles: “Rise of the Machines – The Future of Credit Scoring?”, “Estimation of Probability of Default with Machine Learning Techniques: A Comparative Approach” and “Applying Deep Learning to Credit Scoring: Our Findings So Far”. Interestingly, some were held by practitioners working at early adopter banks or service providers; the others instead, who were from academia, presented a whole set of complex mathematical formulae.
Although most of the delegates agreed that ML is here to stay as it provides incredible performance, participants were still very sceptical when it comes to real world implementation.
Firstly, before being able to apply ML extensively across the bank, risk officers will need to solve the Model Risk Management (MRM) issue. This issue was presented very well during the presentation of one of the keynote speakers, Dr. Diederick Potgieter (Bank of England). MRM is a fairly new field and a very hot topic nowadays given the sheer number of models that are being used by banks to run their business and take strategic decisions.
The potential for adverse consequences from a decision based on incorrect or misused model outputs is huge: Mr. Potgieter brought the well-known examples of the London Whale and LTCM to the attention of the audience. He suggested that the bank’s board needs to be engaged and should drive culture for MRM, which should include not only the models but also the remuneration, training, etc… His closing remark was that it is crucial to develop a culture of MRM, otherwise AI and ML will not find their way in the industry.
Also Alan Forrest (Clydesdale Bank) chaired a series of panels focused on Model Risk. He presented the parallelism between Model Risk and Credit Risk. He suggested to think of the model inventory as a portfolio of related models, where one model can fit or interact with another. Models can have correlations as well as be components of larger and more comprehensive model systems. Indeed, people working in MRM should learn from credit risk to quantify model risk and ask themselves the question “What if?”.
The second barrier when it comes to applying ML to the credit space is explainability, in IT jargon: “the blackbox issue”. In this regard some panels were focused specifically on this and proposed a set of methodologies to overcome the limitation. Explainability is vital to consumers, lenders, and regulators.
According to John Oxley and Eric McVittie (Experian), explainability is presenting textual or visual artefacts that provide qualitative understanding of the relationship between the instance’s components (e.g. application data, words in text, regions in an image) and the model’s prediction.
In this regard, some service providers are experimenting with existing methods, while others are developing (and patenting) their own. It is important to stress that solving “the blackbox issue” and setting up an appropriate MRM is necessary but not enough.
The third and probably trickiest barrier to kick down, as always when it comes to innovation, the culture of the people working in credit risk offices at banks and financial institutions, which is rather conservative by definition. They will try to avoid to lose control of their models. Especially during a downturn period, not knowing how a model works inside-out could be frightening indeed.
Another hot topic was open banking data, which is bringing fundamental changes on how scorecards are built. Infact, together with ML, Open Banking is one of the most important changes of the last few years and many people will be busy rebuilding their models.
Experts demonstrated that by combining or supplementing credit bureau with open banking data, significant scorecard performance improvements are achievable. This is because transactional data offers a completely new perspective and additional insights by i) providing more granular information that is not available in traditional credit bureaus like luxury items vs utility bill spending, rental payments, holiday frequency, saving usage, and gambling activities; ii) giving the possibility of cross-checking key features like “income”; iii) identifying seasonality patterns, life events, and triggers.
It’s worth mentioning though that giving the novelty of the field, much work still needs to be done in terms of normalisation and categorisation of data in order to build a meaningful intelligence layer. Moreover, lots of transactions do not appear in the current account but in the credit/debit card, which are not necessarily accessible to third parties as PSD2 isn’t pushing to expose credit card APIs. For the time-being the main use-case seems to be consumer lending, experts suggest not to push open banking credit scoring to all customers but only to a part of the population.
Finally, as to confirm the international depth of the conference, the panel “Sesame (Zhima) Score: ‘Social Credit Score’ or FICO-like Credit Score?” held by Professor Xinhai Liu (Peking University) tried to bring some clarity on the different credit scoring systems currently in use in China.
Zhima score was born out of the financial department of the Alibaba Group (now Ant Finance) in 2015 as the first public credit score in China and the use soon became widespread. Being more about paying capability than paying intention, digital platforms and digital merchants relied heavily on Zhima.
However, the wide application of Zhima score in fields like house rental, hotel booking, and recruitment raised the eyebrows of the People’s Bank of China, resulting in Ant Finance stopping the business activities regarding the application of the score in the finance field.
It’s fascinating to see how the word “credit” can have different meanings and applications across the globe. Indeed, the ongoing construction of a social credit score in China is a clear example. I think that this is a particularly relevant topic since, as John Thornhill (Innovation Editor and founder of Sifted, FT) clearly stated “in the digital realms, the West is facing a formidable rival in China, which is fast emerging as an AI superpower determined to set its own rules.”
In his FT article, Mr. Thornhill interviewed Song Bing (Director of the Berggruen Institute China Center), who states that chinese AI ethicists prioritise values that are open, inclusive and adaptive, speak to the totality of humanity and reject zero-sum competition. Collective good is just as important as individual rights.
China has adopted what Kai-Fu Lee (Author of the book AI Superpowers) calls a “techno-utilitarianism” approach, emphasising the greatest good for the greatest number rather than a moral imperative to protect individual rights. That, he suggests, is one reason why Chinese consumers are less concerned about installing facial recognition devices in supermarket trolleys to personalise shopping trips, or in the classroom to spot inattentive students. China makes a different trade-off between surveillance and convenience than the West.
Using an expression of Mr. Thornhill, should our future banking App contain a mini-Confucius or a mini-Kant?