Future of SupTech: AI and machine learning in regulatory reporting

In the past, using machine learning in regulatory reporting meant mathematics experts in labs using supercomputers to process data. Thanks to technological developments, today the benefits of machine learning are available to financial regulators worldwide using ordinary laptops. That means computers can process the data while regulators focus on making important supervisory decisions. 

There has been a surge in industry confidence concerning AI and its subfield machine learning. These technologies help regulators to identify trends in large data sets and make predictions based on these trends. As AI techniques have become more mainstream, their usability and accessibility have improved. 

Regulators can now have the confidence to take advantage of this technological shift by integrating machine learning into their processes. 

How can regulators use machine learning in regulatory reporting? 

After the global financial crisis in 2008, regulators introduced stringent reporting requirements financial institutions had to meet to ensure operational transparency. These requirements led to an explosion in the volume of data requested by and provided to regulators. It also means that data of higher complexity had to be delivered much more frequently. 

This creates a problem: it is likely that the amount of data collected will continue to grow, and it is not feasible for the bank to continue taking on new staff to process it. Innovative solutions can reduce this burden and allow regulators to take advantage of the insights in the data. AI and machine learning offer regulators the opportunity to harness the power contained within this data. 

The evolution of the traditional data journey 

Some trends are not easily found - even by subject matter experts - when they are completing exploratory data analysis manually. Using machine learning in regulatory reporting gives a sophisticated level of insight into the trends contained within data.  

The traditional exploratory data analysis process follows the path: 

• Review data in the form of information. 
• Query the information to extract knowledge. 
• Interpret the gained knowledge using wisdom obtained by experience to make a decision 

In the machine learning world, the data is automatically interrogated to find trends. These are provided to the human decision-maker and the significance of each trend is highlighted.  This insight can be achieved in a much shorter time frame and provide a much greater level of detail than traditional methods. It enables better decisions to be completed in a shorter timeframe. 

This paradigm allows regulators to collect a much more granular level of data in close to real-time. The regulator receives the records of each transaction completed and collates the data as they require it.  

The evolution of machine learning in regulatory reporting 

Early iterations of machine learning were difficult to implement, but with recent advances, it has become: 

• Accessible – Machine learning knowledge is easier to ascertain 
• Commoditised – Multiple machine learning models are now available off the shelf 
• Transparent – New techniques allow previously opaque decision logic to be provided 

Until recently, machine learning was confined to research labs. The knowledge needed to implement such techniques was held by a small pool of highly specialised individuals. It also required access to supercomputers with the power to validate the techniques.

With advancements in computing power machine learning techniques are being successfully deployed on desktop and laptop computers. With the added availability of high-performance computing environments in the cloud, even models that require additional computing power are accessible without the need for large infrastructure investment. 

This allows regulators to harness the power of the data they are already collecting. We already have evidence of this in action; it is only a relatively short amount of time since our regulators successfully moved to digital submissions from paper submissions. Machine learning in regulatory reporting has shown its awesome power with significant increases in model accuracy.  

Challenges of machine learning in regulatory reporting 

Until recently, the most accurate models lacked a degree of transparency. Machine learning techniques were extremely good at telling you what to do, but not as good at telling you why you should do it. This was a challenge because transparency is a cornerstone of regulation. Regulators are required to evidence any regulatory action they take, so using black-box algorithms with no details of the logic used to make their decisions is not an option. 

This has recently changed with the development of the eXplainable A (XAI) discipline which provides interpretable and explainable techniques. Interpretable AI and eXplainable AI can now present the logic used in making automated predictions. So, regulators can now take advantage of the highly accurate machine learning methods while using these new techniques to provide the transparency they require.  

Implementing machine learning in regulatory reporting 

Downstream reporting  

The first step for those who want to use machine learning techniques in regulatory reporting has been to extract historical data sets from their data warehouses. Machine to machine reporting is then used to verify previously considered trends. 

Regulators then work to identify previously unknown trends contained within this data that might assist in the future prediction of issues. They also use machine learning to forecast future performance, which enhances industry stability.  

Upstream decision making 

Building on the successes found by training machine learning models on downstream data, this knowledge can also be transferred upstream to assist in real-time regulation. Regulators can integrate machine learning models into the data collection platform to get predictions.  

The earlier in the cycle that machine to machine regulatory reporting is introduced, the more beneficial it can be to the humans involved. From a regulatory point of view, the human in the loop is vital for several reasons:

• Ensures accountability for decisions 
• Processes privileged information and decisions in a safe and secure manner 
• Ensures adequate evidence exists to deliver consequential regulatory actions 
• Builds trust in machine learning models. 

We do not envisage machine learning replacing regulators. The job of a human regulator will become streamlined by the availability of insights from machine learning models.  

That enables the regulator to be more agile. Machine learning allows humans to focus on high-risk areas, and to identify concerns much earlier than was previously possible. Machine learning in regulatory reporting flags potential issues; financial entities, users, or behaviours that are outside of the norm. That can be a red flag for issues that may require a more detailed investigation. 

Regulatory machine learning in the United Kingdom 

The Bank of England is one of the regulators leading the charge with machine learning. It is proactively investigating how machine learning can be used to assist in its supervisory process and regulatory data lifecycle.  

The bank initially published a foundational paper on Machine learning at central banks in 2017. It has built on this with several additional publications that place a greater focus on the need for explainability within the regulatory environment. These include: 

• Shapley regressions: A framework for statistical inference on machine learning models (2019) 
• Predicting bank distress in the UK with machine learning (2019) 
• Machine learning explainability in finance: an application to default risk analysis (2019) 
• Credit growth, the yield curve and financial crisis prediction: evidence from a machine learning approach (2020) 

Regulatory machine learning in the United States 

Meanwhile, in the U.S. the SEC has been ratcheting up its use of machine learning since the global financial crisis. The SEC now uses decisions by machine learning models to guide regulators to entities whose behaviour required more detailed analysis. This uses a method called 'keeping the human in the loop', which takes advantage of the power of machine learning but maintains a level of human accountability. All privileged decisions involve a human actor as opposed to autonomously by a machine. 

Future of machine learning in regulatory reporting 

The world has undergone a monumental shift in the digital age. The time is now for innovation to unleash the power of the insights contained within this data. 

Within financial supervision, the next step is to integrate this technology into existing data collection pipelines. Sharing the available insights with the regulator as early as possible in the cycle will be the key to success.