Machine learning has the potential to revolutionize the compliance function in financial institutions. If used appropriately, it can slash personnel costs and massively reduce risk to the point that KYC costs can be cut by up to 57%. We have gathered 10 lessons learned from machine learning projects in compliance departments at financial institutions. Here we highlight the critical factors for success and the issues that really count.
1. The first step is the hardest: ask the right questions
The main advantage of using machine learning in compliance and KYC is the fact that it facilitates the identification of particular transactions in a huge mass of data. This is vital for banks in their tough fight against money laundering, terrorist financing and fraud, along with their need to meet regulatory requirements. But at first it can be hard to assess the potential of this technology. That’s why you have to ask yourself some basic questions: What do I want to achieve? What requirements do I have to meet? What can machine learning do better? What tools can I use? Who do I need for this? Where do I start, and how long will it take?
2. A vital combination: machine learning and expert rules
Machine learning is particularly good at identifying suspicious patterns and transactions when the parameters are unclear. But it can’t solve every problem. That’s why it has to be supplemented with expert rules that set precise parameters based on regulatory requirements. The banks’ experts define their business rules, such as certain compliance criteria, and these are then automated via special software. These two approaches go hand-in-glove: machine learning when the parameters are vague, and expert rules in clear-cut situations.
3. Combine the right data with the right approach
One of the key factors for success is combining the right data with the right approach, depending on the issue being addressed. Machine learning has six main approaches to different tasks. On the one hand there are monitored approaches, such as classification, regression, predicting time sequences, and multiclass classification, and on the other there are unmonitored approaches, such as clustering and anomaly detection. Classification, a monitored learning area, is particularly important in compliance. This involves dividing customers into two categories: those that trigger an alert and need double-checking, and those that do not require a subsequent check. However, this approach has to be based on a data pool containing a representative sample of the problem cases that require detection. If the data pool does not contain adequate numbers of these risky cases, unmonitored approaches are used – particularly customer clustering and anomaly detection.
4. Machine learning decisions shouldn’t be black boxes
Traceability is a key issue when using machine learning in compliance. Machine learning models shouldn’t be black boxes with unexplainable results. It is vital to have a clear understanding of why the model has made a particular decision. Explainability is primarily a requirement of financial regulators, but it is also important for in-house transparency, for ensuring the quality of the machine learning model, and for explaining to customers why a particular decision has been made. This means that machine learning should not be used in key areas of banking compliance unless it offers full traceability.
5. Operationalization: putting machine learning into practice
Operationalization is one of the main challenges in any machine learning project or pilot. This relates to the model’s actual, operational use in the various business processes. We know from experience that this is what causes many projects to falter, rather than the model itself. This is largely because the two worlds of development and application are very different. The data scientists who specialise in developing ML models tend to work separately from IT and operations teams, whose focus is on application, and they often have different objectives. Bringing these two worlds together requires platforms and tools that encompass both the development and the implementation of ML models.
6. The quality should be continuously improved
You have reached the first major milestone when the machine learning model has been integrated into your business processes and is delivering reliable results. But that’s not an end to it. It is important to keep improving its quality because every day brings fresh data, new behaviours by stakeholders, new decisions. This requires the progressive adaptation of existing models or even setting up new models as necessary. When used in compliance, machine learning has to be retrained at sensible intervals – quarterly, half-yearly or annually – in line with changing parameters and then restored to live operation. This ensures the continuous improvement of the machine learning model.
7. Save over 50% on clarifications
Financial institutions have the potential to make enormous savings when they introduce machine learning into their compliance function. In this respect, we have examined more than ten of our customer projects and pilots. Despite the fact that each customer was working within different parameters – such as data quality, information, customer structure and sanctions lists – we were still able to identify savings of between 30 % and 50 % across all the projects. In concrete terms, this means that only about 50% of false positives had to be checked in order to continue finding 100% of true positives.
8. Intelligent reviews reduce both costs and risks
Machine learning in bank compliance is not only about cutting costs. It also has a vital role to play in mitigating risk. With the help of intelligent reviews, this can occur in the most unexpected places.
9. Detect risks that were overlooked in the past
The machine learning models are not only applicable to new and future cases but can also easily detect risks that were overlooked in the past. This means you can look back through the entire history of cases and suspicious items that were previously classified as unproblematic are automatically flagged up. So machine learning is not just about the future. It can also shed light on the past and uncover critical cases that have been overlooked.
10. Experience makes the difference
For financial institutions that are seeking to establish machine learning in their compliance departments, it is crucial to find an experienced service provider. It is essential to have a profound understanding of the possibilities, strengths and weaknesses of machine learning in order to achieve your goals more quickly and avoid setting off down the wrong path. Experience is also indispensable when it comes to choosing the right approach to the available data and meeting regulatory requirements. Working with an experienced partner also allows you to estimate expenditure more accurately and plan with greater confidence.