Machine Learning in Compliance

How have the challenges faced by bank compliance departments changed in recent years?

Thomas Ohlemacher: Banks and insurance companies have spent many years tackling a barrage of new regulations. They have been under mounting pressure to meet these new regulatory requirements. However, over recent years their focus has shifted away from dealing with new regulations towards greater standardization and efficiency. Banks need technologies such as machine learning to help them achieve these objectives.

Which machine learning use cases offer real benefits for compliance management?

Thomas Ohlemacher: Let me describe two specific use cases:

Payment monitoring

In payment monitoring, machine learning is able to learn from existing payment histories. It recognizes patterns of similar, unusual payments. Machine learning is able to generate very powerful models that often make it possible to filter payments more accurately and trigger fewer clarifications.

Read white paper

Name screening

My second example is name screening, comparing names against sanction and PEP lists. Machine learning software uses the clarification history to learn how compliance officers have clarified such cases in the past, then it uses this as a model for evaluating future cases. This makes it possible to assess at a very early stage whether a hit is likely to be relevant, then prioritize accordingly.

Read white paper 

Are people worried that machines could replace humans?

Thomas Ohlemacher: Of course there are concerns in the industry and, strictly speaking, they are not unfounded. Machine learning will inevitably replace humans in compliance departments, but this is a phenomenon that began when the first assembly line was automated.

However, in risk management and compliance there is a strong desire to make clarifications less time-consuming and achieve a more accurate hit rate, as this can create a significant competitive advantage.

Our experience with machine learning shows that automation is only successful when machine learning is combined with input from compliance specialists. Only they can assess the data in accordance with their particular objectives, so compliance experts are needed to manage the process and increase efficiency.

To sum up, it’s true that the steps in the process are carried out by machines, but they are always based on human decisions. So the need for experts is growing rapidly, and the enhancements merely cushion this huge demand.

How does the German Federal Financial Supervisory Authority (BaFin) approach the issue of automated decision-making?

Thomas Ohlemacher: In its report Big Data meets Artificial Intelligence, BaFin states that models cannot be categorized purely as black boxes. Accordingly, decisions based on machine learning have to be explainable and comprehensible by third party experts. BaFin also sees transparency as an opportunity to improve the analysis process.

So firms that are subject to BaFin supervision have to ensure that their compliance generally provides traceability, particularly for machine learning models.

Otherwise, why is traceability important?

Thomas Ohlemacher: Traceability helps compliance officers to recognize whether the machine learning model needs to be adapted. This is particularly the case if they feel the results are no longer plausible. For example, this might happen when operating conditions change but the machine learning models have not been retrained. Traceability can also help compliance cases to be analyzed at an early stage, even before they are on the firm’s radar. The information it provides can make it possible to detect patterns that weren’t obvious before.