Improving the quality of analyses, detecting and preventing errors automatically, and enhancing the operational quality of processes: these are some of the challenges that many banks are facing – and machine learning (ML) offers promising solutions here. When banks take advantage of the opportunities offered by ML, we as supervisors can benefit from this, too. This is because efficient banks are stable in the long run. But how can banks seize the opportunities offered by ML while avoiding losing control of the risks? And how can banking supervisors minimise these risks without hampering innovation?

Characteristics instead of definitions

The key issue is to identify ML in the first place and to highlight the differences between ML and the models and processes that have been under our supervision for decades. These differences show us the areas where our supervisory practice might need to be adjusted, and we do not necessarily need a generally applicable definition of ML for this purpose. Rather, our supervisory practice, inspection techniques and inspection intensity will be guided by whether a specific method displays characteristics of ML, and if so, which ones. In addition, we will take into consideration the extent to which these characteristics shape the method in question. This approach will help us to identify and appropriately deal with ML innovations and the attendant risks – while avoiding a one-size-fits-all approach for new applications at banks.

For example, a scenario in which banks combine a large amount of new data and reach lending decisions using neural networks will involve many ML characteristics, resulting in greater risks compared to other existing models, especially in terms of explainability. However, if only secondary sources of data, such as newspaper reports, are obtained via ML and are used in the area of loan granting or early risk detection – in addition to traditional metrics – the model as a whole will display fewer ML characteristics. Correspondingly, less stringent supervisory requirements would then apply.

ML methods (do not) need to be explainable

People like to refer to ML as a black box that you cannot see inside. We often fail to recognise immediately which information these methods use as a basis for decision-making. It is important, however, to consider each case individually. The right balance is often found in a “grey box”, somewhere between the need for explainability, on the one hand, and model freedom, on the other. The key point is to understand and scrutinise the method used – not to ensure that each individual step of the process is approved by a human being.

The characteristics of ML serve as a compass for banks and supervisors. And this is something we are taking a close look at – while paying equal attention to risks and opportunities.