BrazilAI can improve operational risk management in banking
Risk management is an integral part of banking. By taking financial risks, banks are able to generate the profits that are necessary to survive. Risk management aims to control this process by making potential losses more predictable. This makes the bank more robust to external fluctuations.
Whereas profits can be made by accepting certain financial risks, operational risk is intrinsically different. Operational risks only cause losses – financially in terms of bottom-line impact and non-financially in the form of for instance reputational damage. The consequences of operational risk events can have a large impact on an organisation and the financial system as a whole as experienced during the last financial crisis. It is not surprising therefore that operational risk is receiving more attention within the financial sector, with banks trying to minimise the operational risks they take, given the resources available while keeping in mind the strategic goals of the organisation.
In the past decade, there has been major progress in the development of artificial intelligence (AI). AI algorithms excel at data analysis and have evolved to the point where they surpass human performance for a wide variety of tasks. More and more businesses exploit these technological advances to optimise different kinds of processes such as marketing, sales and e-commerce, manufacturing and logistics. In today’s growing data-driven world, this trend is expected to continue on the back of widening opportunities for use cases, including in the area of operational risk management.
Challenges in operational risk management
In June 2011 the Basel committee published the Principles for the Sound Management of Operational Risk (BCBS), which provides a framework for the development of proper operational risk management. Three years later, a survey was conducted to measure to what extent banks complied with these principles. One of these principles states that banks should write a risk appetite and tolerance statement. Banks reported that this is more challenging for operational risk than for other risk categories and attributed this to the nature and pervasiveness of operational risk. The banks that did comply with this principle often reported the use of backward-looking metrics of operational risk, such as operational losses as a percentage of gross revenue.
The above example underlines the challenges which banks face in the management and measurement of operational risk. Compared to financial risk, operational risk is a more qualitative field of study. Whereas financial risk management has been the main priority of banking for a longer time, operational risk management is much younger resulting in less extensive historical data. Predictive modeling becomes more of a challenge in this situation.
On top of that, the events in operational risk are much more diverse in scope. The Basel committee defines operational risk as the “risk of loss resulting from inadequate or failed internal processes, people and systems or from external events”. Internal fraud, data leakage and reputational damage are very different problems, yet can be very closely related as well. A lot of the data in operational risk consists of textual input which contain qualitative information. The qualitative nature of operational risk is reflected in the Basel framework, which encompasses guidelines for organisational structures, culture and awareness, and qualitative reporting.
The computer as a reader
Artificial intelligence could play a valuable role in upgrading operational risk practices, with in particular machine learning – the field of self-learning computer algorithms – showing promise. Machine learning algorithms can make predictions based on data fed to the algorithm. Recently, major progress was made in the field of natural language processing (NLP). NLP focuses on using textual data for predictions. As an example, an algorithm can learn to rate hotel reviews. By processing large amounts of reviews together with their given ratings, the algorithm can learn to give a rating to a new review it has never seen before.
“There are many opportunities for operational risk management to exploit AI and other related technological advances.”
– Lars de Ruiter and Matthias Geerse, Solid Professionals
In operational risk, many textual reports are written regarding specific risks or risk appetite of the organisation as a whole. These reports are generally written by risk managers – experts in their field. Assuming that risk reports hold information which is absent in historical loss data, this information could be extracted and used for predictive purposes. So why should computers perform this task instead of risk experts? The answer lies within the physical limitations of human being – as well as the natural biases they have to cope with. While humans may struggle to remember a piece of text, a computer algorithm easily processes thousands of books and finds structure within using statistical methods with near-perfect precision.
Towards prediction of operational risk
The performance of NLP algorithms using word embeddings has increased tremendously of late. Using the embedding as a starting point, these algorithms learn the meaning of full sentences and use this information for further predictions. This is more generally called sentiment analysis. A popular application in the financial setting is predicting stock price movement from news articles, or twitter feeds. A recent paper by Denmark’s national bank predicts corporate distress of firms by analysing their annual reports. Taking textual data under scrutiny can make a contribution to quantitative predictive modelling, even in the financial sector where the use of words is very different to day-to-day language.
In summary, there are many opportunities for operational risk management to exploit AI and other related technological advances. One possibility is the classification of risk events. As an example, a computer algorithm can read risk descriptions written by risk managers and classify them according to their impact and frequency. Combining loss data with risk reports, improved prediction of risk events might lead to more accurate prediction of future losses. A different application is the measurement of more abstract concepts such as the financial health or maybe the cultural aspects of a company.
An article by Lars de Ruiter and Matthias Geerse. Both are finance & risk consultants at Solid Professionals, a consultancy from the Netherlands.
