Harnessing the future of machine learning

Machine learning (ML) techniques are creating waves within the financial services sector. The banking industry, which relies heavily on data, is increasingly adopting these techniques and has started to leverage their powerful capabilities. From chatbots to fraud detection, the banking sector is using ML not only to automate processes and streamline operations for both the front and back offices, but also to enhance the overall customer experience. ML tools, with their advanced prediction techniques and capabilities to utilize large volumes of data, are increasingly being used in risk management. They may inform quicker and more efficient credit, investment and business-related decision making. Another important area where ML is gaining significance, albeit at a slower pace, is regulatory stress testing. Traditional statistical methods of stress testing have been critiqued by investors and regulatory agencies as ‘not severe enough’, with numerous banks failing during crises. In the following pages we discuss the application of ML in risk management, as well the benefits and challenges of adoption.

Credit risk

• Enabling of a more robust assessment of customer credit history
• Providing insight that reveals additional operational, financial, or economic vulnerabilities that traditional models could not provide
• Enhanced accuracy of internal credit scoring models

Fraud risk

• ML can quickly spot patterns from several different channels in tandem and send alerts about potentially fraudulent activity for a virtually limitless number of banking clients at once
• Successfully differentiate between legitimate and fraudulent transaction within milliseconds

Market risk

• Reduction of risk in market trading
• Analyzing vast volumes of data points in seconds
• Providing traders with optimal price points
• Identification of trading patterns across multiple markets
• Enhanced accuracy of forecasting models

Enterprise risk

• Enhanced decision-making through greater predictive insight and visibility of risks for senior management
• Enhanced risk monitoring mechanisms whereby ML can be used to predict the likelihood of default
• Identification of breaches in security and/or other business controls

1. Credit risk modelling Banks frequently use traditional credit risk models to predict categorical, continuous, or binary outcome variables (default/non default). ML models are often difficult to interpret and may not be easy to verify for regulatory purposes. Nevertheless, they can still be used to optimize parameters and improve the variable selection process in existing regulatory models. AI-based decision tree techniques can result in easily traceable and logical decision rules, despite having nonlinear characters. Unsupervised learning techniques can be used to explore the data for traditional credit risk modeling. Classification methods, such as support vector machines, can predict key credit risk characteristics, including probability of default (PD) or loss given default (LGD) for loans. Financial services firms are increasingly hiring external consultants who use deep learning methods to develop their revenue forecasting models under stress scenarios
2. Fraud detection Banks have used machine learning methodologies for credit card portfolios for years. Credit card transactions present banks with a rich source of data which may be used to process and train unsupervised learning algorithms. Historically, these algorithms have been highly accurate in predicting credit card fraud due to their ability to develop, train and validate huge volumes of data. Credit card payment systems are embedded with workflow engines that monitor card transactions to assess the likelihood of fraud. The rich transaction history available for credit card portfolios presents banks with the ability to distinguish between specific features present in fraudulent and non-fraudulent transactions.
3. Trader behavior Technologies such as natural language processing and text mining are increasingly used to monitor activity and identify rogue trading, insider trading and market manipulation. By analyzing email traffic and calendar-related data, check in/check out times, and call times combined with trading portfolio data, systems are able to predict the probability of trader misconduct, potentially saving millions in reputational and market risk for financial institutions.

Artificial intelligence potentially offers faster decision making, enhanced accuracy, predictive power and more robust stress testing methods by automating human operations, leveraging the power of big data and reducing costs. Embracing a rapidly advancing new technology that disrupts business as usual is not easy. KPMG can help your organization seize the potential of artificial intelligence in the context of risk management. KPMG professionals are dedicated to working with you to create relevant, scalable solutions that drive value for your organization. We understand the issues and challenges involved in the development of a robust model risk management framework and KPMG professionals can leverage their experience to deliver tailored services. Our suite of services provides full support at every stage of development - from proof of concept to designing relevant use cases, integrating systems and operations to ongoing management support. We can support your organization in leveraging the benefits of +ML technologies and incorporate these in risk management models. From scenario design, macro-economic modeling, and stress testing to IFRS 9 PD, LGD and exposure at default (EAD) models, incorporating both Pillar 1 and Pillar 2 risks, our teams are here to assist.