Next-generation artificial intelligence tools are ramping up the defence against fraudsters and allowing banks to increase efficiency, reduce headcount in compliance and provide a better customer experience. Joy Macknight reports.
While artificial intelligence (AI) has gained a lot of attention this year – boosted by Elon Musk and Mark Zuckerberg’s very public spat – the financial services industry has long used the technology in fraud detection.
However, the AI used in the past was simple, rule based and fairly rigid, a way of automating experts’ decision-making processes with an outcome in the form of a score – the higher the score, the greater the likelihood of fraud. Although the algorithms could be tweaked, they were incapable of learning or expanding their expertise.
In addition, the true potential of AI was restricted by available computing power, which forced companies to “trim” algorithms, limiting how far they delved into the data in order to be operationally executable, according to Johan Gerber, executive vice-president, security and decision products, at MasterCard.
These constraints resulted in high levels of false positive alerts, with endless investigation hours wasted on authentic transactions.
Learning curves
Today, advanced AI techniques, such as machine learning and deep learning, are possible because of faster, cheaper and stronger computing power, as well as more sophisticated algorithms and the vast amounts of available data in an increasingly digital world.
Many banks are now replacing their old AI fraud systems and developing expertise in machine learning, a form of AI that enables computers to learn without being explicitly programmed, and deep learning, which uses multiple levels of neural networks to see, learn and react to complex situations.
Lloyds Banking Group, for example, is using machine learning for online fraud prevention. Gill Wylie, chief operating officer, group transformation, at the UK bank, says: “We use models that can detect when the person logged in to our online banking is not the customer, but rather a fraudster, or even a ‘bot’. This helps us stop the fraudsters in their tracks.”
Business trends
While specific technological advances have made next-generation AI possible, several business drivers are also positioning it as a critical solution for banks in the fight against fraud.
Nadeem Gulzar, head of advanced analytics and architecture at Danske Bank Group, believes that the shift to digital banking has spurred a need for better AI solutions. “With the digitisation of banking products has come a surge in fraud, whether in e-payments or online applications. Obviously, as attacks on our digital channels increase, we need an improved detection rate,” he says. “In addition, fraudsters are becoming more tech savvy, so we need to use machine learning, and even deep learning, to stay one step ahead of them.”
TJ Horan, vice-president of product management at FICO, an analytics software company, points to the growth in new and emerging payment types, such as instant payments, as another driver for AI adoption. With real-time payments comes the need for real-time analysis, but in a situation where a bank does not have the historical data needed to train its models and analytics. FICO has developed “self-calibrating” fraud analytics technology, which continuously adjusts in real time based on the transaction stream, according to Mr Horan.
Online surge
But it is also the sheer volume and surge in online transactions that is putting banks under pressure to replace the old rule-based systems. Global non-cash transaction volumes grew 11.2% during 2014-15 to reach 433.1 billion, the highest growth of the past decade, according to the World Payments Report 2017.
“Using AI significantly reduces the time it takes to review a transaction, including all the factors and relevant data associated with that transaction. Our defences must both harness the vast data available to capture, as well as adapt to changing patterns and new instruments very rapidly,” says Mr Horan.
He continues: “Banks want to allow consumers to have the convenience that they demand, yet also have the required protection from a regulatory perspective. The speed at which everything is changing – and by that I mean consumer behaviour, new products and services, as well as the speed of global commerce and raw processing power – is pushing AI to the fore.”
MasterCard’s Mr Gerber also cites changing customer behaviour as a business driver for AI deployment. “We are moving into an era of frictionless, or intuitive, payments,” he says, using the example of Uber, where the payment occurs automatically when a customer steps out of the vehicle.
“In the past, the focus was on keeping fraud down and limiting losses, which had the potential of introducing friction into the payment process that could cause customers to have an unhappy experience. But today, the focus is on providing the best consumer experience with the least amount of fraud – the consumer experience is paramount,” adds Mr Gerber. “Therefore, our algorithms need to be incredibly accurate to ensure we only introduce friction when it is absolutely necessary.”
In November 2016, MasterCard launched Decision Intelligence, a fraud detection service that uses machine learning to move from predictive analytics to what Mr Gerber terms “prescriptive analytics”. “We aren’t just looking to see if this transaction is fraudulent, but looking at the behaviour of the account holder,” he explains.
“When looking at our model performance results, we see a 40% increase in the accuracy of detecting fraud with Decision Intelligence, compared with the old algorithms, and a 50% reduction in false positives,” adds Mr Gerber.
Reducing false positives
As evidenced by the Decision Intelligence metrics, employing new AI techniques can lead to an impressive improvement in detection rates and a reduction in false positives, which in turn limits financial losses and reputational damage.
The main objective of Danske Bank’s fraud project was to reduce false positives. “We can now take all the information, from transactions to customer behavioural patterns, and feed it into the advanced analytics engine, and that targeted insight is then made available to our investigation officers,” says Mr Gulzar.
To reduce the potential for false positives in its transaction monitoring process, Singapore’s OCBC Bank turned to the fintech community for advanced AI solutions to complement its existing systems. In July, it announced partnerships with BlackSwan Technologies and Silent Eight, which were part of the bank’s second accelerator programme run by its fintech unit, The Open Vault at OCBC. The former uses AI to detect red flags and analyse suspicious transactions, while the latter helps transaction-monitoring analysts put together a suspicious individual’s 'dossier' within 60 seconds.
Alex Ng, head of the OCBC Bank’s group transaction surveillance, expects a productivity uplift of two to three times with AI technology. “There are many manual processes involved in the investigation process, from researching online to querying the business,” he says. “The AI solutions will automate many of these processes and make them much quicker, so that more time can be spent analysing real cases.”
“The quality improvement benefits that AI can produce are quite clear,” says Cem Dimegani, founder of appliedAI.com, a platform that connects enterprises with AI solution providers. “AI can also improve processing speed and banks can therefore approve transactions faster, with less false positives, which results in better customer experience.”
It also allows banks to cut costs by reducing the headcount in compliance, which Mr Dimegani believes is a major driver for AI adoption. But this was not Danske Bank’s aim, according to Mr Gulzar. “The primary factor for us was to be much more targeted in our investigations. And looking at our different use cases for machine learning, it is much more about how to make the bank more relevant to our customers,” he explains.
Instead of reducing staff levels, six months ago Danske Bank began retraining its investigation officers. “Whereas a rules engine was simple to understand, it might be more difficult to understand exactly why this transaction was flagged by an AI algorithm. Even though we have employed the Locally Interpretable Model Explanation architecture, which is part of machine learning and deep learning to explain why exactly was this triggered, there is still some training required,” he says.
Overcoming challenges
AI is only possible with the right staff, according to Mr Gulzar. “We need people that understand the more advanced methods but can also apply the technology in an enterprise world such as banking,” he says. “And, to be honest, it hasn’t been easy to attract data scientists for they are in high demand.”
The technology developments underlying recent AI achievements gained commercial acceptance at a time when banks were cutting costs, according to Mr Dimegani, who says: “Technology giants and private capital invested significantly, making it harder for banks to attract necessary talent to build their own solutions.”
To address this shortage, Danske Bank has developed close relationships with local universities in Denmark and Lithuania, and has been able to help curate new university courses for data scientists or engineers.
Gathering vast amounts of data into one place is another challenge banks face. Over the past few years, many have been building data lakes, according to Imam Hoque, chief operating officer and global head of product at Quantexa, a big data and analytics company backed by HSBC Ventures and Albion Ventures. The start-up helps organisations pull in third-party data, produce a single view and network the data before applying AI, a technique that many banks are now using in their data lakes, he reports.
Danske Bank is using Hadoop, an open-source software framework, to train its models, as well as execute deep learning models. “That is much closer to cutting-edge technology than we have been used to,” says Mr Gulzar. “As a bank, we had to re-educate ourselves in sense that we aren’t historically first movers, but in this space if we want to do something about fraudulent cases then we do need to be a step ahead.”
Lloyds Banking Group has built a big data platform to analyse the examples of fraud that affect its customers, for example using algorithms to detect the likelihood that a customer will fall victim to fraud. Ms Wylie says that in the future the bank will be using a cloud-based machine-learning risk engine that will be retrained frequently to combat fraudsters’ latest tricks.
A digital lead
Interestingly, this is an area where new digital banks have an edge over incumbents, according to Ezequiel Szafir, CEO of Santander’s newly launched Openbank. “By default, we are constructed from a technology point of view,” he explains. “The data architecture we use, such as data lakes and public cloud, allows us to perform real-time analysis and fraud prevention without disturbing the customer. This technology set-up also allows us to do it at a decent cost, compared with mainframe-based traditional set-ups.”
Openbank uses unsupervised machine-learning algorithms for both fraud and anti-money laundering (AML). “Compared with traditional rules-based AML/fraud mechanisms, machine learning can detect fraudulent activity or money-laundering seven to 10 months earlier, which is very impressive,” says Mr Szafir.
Fraugster, a German-Israeli payment security company that uses AI for fraud prevention, is very familiar with the importance of data. Its major key performance indicator in the first year of operation was to get “as much data as possible, no matter the cost”, according to CEO Max Laemmle. “Everyone thinks that you just turn on an AI engine and it works, but that’s not what happens. In the end it’s a data game, and the player with the best data will win, in the sense of having the best results,” he says.
The company counts large payment companies Wirecard and Ingenico among its clients. “Banks don’t understand that if they want superior technology, then they may have to sometimes share data,” says Mr Laemmle. “The payment industry is more open in that it agrees that it is better to join forces to fight against criminals.”
The future is AI
In terms of new developments, Mr Horan says that "explainable AI" is beginning to attract interest. “Being able to truly understand the reasons behind various recommendations is a research area FICO is investing in,” he says. “The difficulty is that often the AI and machine learning algorithms build predictions based on relationships within the data that aren’t immediately observable. These are sometimes hard to explain to others, except highly trained mathematicians, so this will be a continuing growth area.”
Tristan Blampied, head of product at Pelican, a payments and compliance solutions company, believes that the next big thing in AI is "cognitive automation", which can be applied across compliance, fraud and money-laundering detection. For example, banks could incorporate vision and natural language processing techniques with machine learning on free format text to automate know your customer onboarding. “KYC by definition is heavy on manual document checking, something that a human cognitive ability is needed to do, but banks could make huge efficiency gains if this process could be automated,” he explains.
While a large part of Lloyds Banking Group’s AI focus is on fraud today, Ms Wylie believes that in the future the power of machine learning and AI will allow the bank to improve every area of its relationship with customers and the security of the bank. “From the early-warning detection of attacks on our network, through the identification our customers’ needs so that we can help them with their finances, to enabling new ways for the customer to authenticate themselves and interact with us – AI will be everywhere,” she says.
