Banking is one of the oldest industries in the world—dating all the way back to simplistic exchange systems in ancient Assyria and Babylonia—yet it has consistently been at the forefront of technological innovation. The first wave of “fintech” actually dates back to 1866, with the completion of the first transatlantic cable and the resultant globalization of financial markets.
Today, banking is leading the way in the adoption of artificial intelligence, particularly for risk management functions. The field of risk management has dramatically increased in prominence since the global financial crisis, with banks and financial regulators now acutely focused on how risk is detected, reported, and managed. According to Accelerate Your Data-Driven Transformation, a study conducted by Forrester Consulting on behalf of RapidMiner, 88% of those surveyed from the financial services industry said that reducing risk was a very important or critical outcome for artificial intelligence, machine learning, and advanced analytics initiatives.
And because AI, especially in the form of machine learning, allows computers to analyze vast volumes of data while perpetually improving their own analytics capabilities, it’s perfectly suited to support the banking industry.
Risk Management in Banking: 3 Ways AI Is Changing the Game
Let’s take a look at three ways that AI and ML can help financial institutions identify risk in an effective and timely manner, make more informed credit decisions, and improve all aspects of regulatory compliance.
1. Real-time transaction fraud detection
Fraudulent transactions represent only a tiny fraction of financial transactions, but they’re a big problem for both the banking sector and any industry reliant upon digital payment transactions. Recent research suggests that banking services and businesses involved in e-commerce, airline ticketing, and money transfers will cumulatively lose over $200 billion to online payment fraud between 2020 and 2024.
What’s more, this problem has been compounded by COVID-19, as more people are paying for goods and services digitally, which creates an ever-growing attack surface for fraudsters. The costs of attacks are also rising. According to a LexisNexis Risk Solutions study, financial service firms end up spending $3.64 for every $1 of fraud due to direct losses from fraudulent transactions, fines and legal fees, investigation and recovery expenses, and other factors.
The detection models that banks have historically used for fraud were (a) developed for physical credit card transactions and (b) structured around rule-based systems based on insights gained from past incidents.
Today, though, online purchases do not require that a credit card be physically present. And rule-based systems have not proven adaptable enough for either the fast-paced world of modern digital commerce or the sophisticated payment fraud schemes developed by criminals
This is why so many industries involved in e-commerce, especially banks, have transitioned from rule-based systems to machine learning-based models. With machine learning, AI systems are able to constantly adjust rules and even learn new ones as more and more data is processed—thus providing an immediate advantage when confronting the many risks and forms of fraud.
Two types of machine learning algorithms are employed in this task: supervised and unsupervised:
- Supervised learning uses already annotated historical data (with instances of fraud activity manually labeled) to identify fraudulent patterns.
- Unsupervised learning deals with datasets that have not been labeled and looks for relationships and variable links that might not be apparent to human investigators.
The two approaches are ultimately complementary, as supervised techniques learn from past fraudulent behaviors while unsupervised techniques allow the detection of new types of fraud. And by combining the two, banks are able to analyze transaction flows in a holistic way, recognize subtle patterns from a user’s purchasing journey, and then accurately identify and flag fraudulent activity in real-time without the need for human intervention. And if you’ve ever gotten a notification from your bank asking if you’ve just made a large purchase, you’ve helped to train these models.
This lets humans focus on more complicated fraud cases—as well as the work needed to resolve the issues that arise from fraudulent transactions—and thus produces better results for both financial institutions and their customers.
For these reasons and more, machine learning has become a crucial component in fraud detection and prevention efforts. Juniper, in fact, predicts that spending on machine learning software by banks and key digital commerce players will reach $10 billion by 2024.
2. Lending fraud detection
Lending fraud has always been a problem for financial institutions, but the Internet age has amplified its prevalence. It’s all too easy today for fraudsters to falsify applications using someone else’s IDs, mobile phone number, and even social security number. This puts lenders in a tough spot, as consumers or businesses generally need loan funds promptly, while rigorous assessments of loan applications can take a long time.
The aforementioned LexisNexis Risk Solutions study found that identity, address, phone, and email verification are all among the top-ranked online channel challenges for mortgage lenders and smaller creditors. The study also notes that sophisticated forms of identity fraud are growing among mid-and-large-sized firms and those using mobile channels.
Lending fraud is especially pronounced for lenders serving small-and-medium businesses. A different LexisNexis study found that fraud targeting SMB lenders has increased recently, with small banks and credit unions being hit with losses of 4.5 percent of their revenues.
Spotting and thwarting fraudulent applications early is clearly important. But as explained in the transaction fraud examples above, classic rule-based systems—those reviewing past user behavior and existing identity information—have difficulty flagging complex fraud schemes, which in the case of lending fraud can mean synthetic identity fraud (in which real and fake information is combined to create a new identity), loan stacking (in which an applicant applies for multiple loans within a short period of time), and account takeover (in which a fraudster obtains access to a consumer’s account).
Machine learning algorithms, however, can spot discrepancies, inconsistencies, and unusual patterns faster and more accurately as they’re not restrained to a limited number of variables. ML can also help lenders cross-reference applications and uncover additional relevant information. And the more datasets models review, the “smarter” their predictive capabilities and risk profiles become, as learnings from past applications can be applied to future ones.
3. Regulatory compliance
Compliance and risk management are closely aligned, as compliance with federal and global regulations helps protect banks from a variety of unique risks. The process of compliance is expensive, though, costing the banking industry $270 billion a year and accounting for 10 percent of operational spending, according to a recent Citigroup report. Even the average credit union spends the equivalent of one employee’s time (for every five employees) on meeting regulatory obligations.
Machine learning is vastly superior to human beings for this task because not only can ML-powered programs process massive amounts of data in rapid time while analyzing countless variables, but they can also identify important correlations that would never stand out to us.
JPMorgan Chase is a great example of this. The bank previously employed a whole team of lawyers and loan officers to spend 360,000 hours each year tackling mundane tasks, including reviewing commercial-loan agreements. But by using an ML-powered program, COIN, the bank was able to process 12,000 credit agreements in several seconds. This not only shortened the time required to review documents, it also decreased the number of loan-service mistakes. JPMorgan Chase is now planning to deploy machine learning in more complex areas, such as credit default swaps and custody agreements.
ML solutions can also help reduce—if not effectively eliminate—the number of false-positive alerts in compliance systems. Right now, each of these alarms triggers a review by a human compliance officer. But by learning from the compliance officers’ own data, machine learning programs can increase efficiency and accuracy, streamline operations, and reduce costs by only surfacing alarms when the detection system isn’t sure and human expertise is needed.
Machine learning allows AI systems to surface insights within large, complex data sets. This technology has clear applications for banking risk management, and when implemented, can lower operational and compliance costs while providing decision-makers with more accurate credit scores. AI-powered solutions can also be incorporated into various financial risk products.
If you’d like to learn more about how to get your own AI initiative started, check out our Human’s Guide to Machine Learning Projects to make sure you’re doing things right from day one.