The financial services industry has been leveraging artificial intelligence (AI) – specifically, machine learning (ML) – to support payments for over a decade. For example, early investments by PayPal in leveraging predictive ML engines to identify fraudulent behavior have allowed the company to protect its customers from having their money stolen, even with massive transaction volumes (nearly 5 billion payments in 2015 for 188 million customers from more than 200 countries).
Despite these early and significant investments, the use of AI in the payments industry is still early in its evolution, but it promises to significantly change how customers and transactions are monitored and analyzed, even as payments processes move closer and closer to real time. The best is definitely yet to come.
What Is Machine Learning?
Before we explore how this technology can be deployed in the payments industry, it is worth a quick review of what ML actually is. ML and the broader AI are based on the concept of a predictive mathematical model and provide systems with the ability to automatically learn and improve from experience without being explicitly programmed.
Although the use of statistical modeling for prediction isn’t new, the approaches taken to train an ML algorithm can consume many more dimensions of input data and work on larger transactional data sets while being significantly better at predicting outcomes. ML software also adapts and improves over time based on changes in the input conditions and results sets.
Machine learning is a mature technology, and although innovations are emerging rapidly from universities and venture-backed startups, the basic research and corresponding mathematics have been around for many years.
ML can be approached in many different ways, ranging from simple, focused regression and tree-based models aimed at specific use cases, to neural networks and deep learning approaches that can be used to work on extremely complex problems.
At its core, machine learning involves establishing a data set that consists of known input conditions and known results and then “training” the model to predict the output to a high degree of accuracy. These techniques allow image search engines to differentiate between a picture of a domestic cat and a tiger, or between a tree and a stalk of broccoli. They also allow Amazon’s Alexa or Apple’s Siri to understand a question posed in a conversation, and an autonomous car to realize that the vehicle in front it is a school bus, is likely to stop often, and can’t be passed while its red lights are flashing.
ML’s Impact Is Accelerating
As noted earlier, ML is a fairly mature technology, and although innovations are emerging rapidly from universities and venture-backed startups, the basic research and corresponding mathematics have been around for many years. Yet the application of ML to financial services generally, and more specifically to payments, has accelerated in the last two to three years, due in part to three supporting trends:
- Cheaper compute power and storage - The price of compute power and storage has continued to drop year over year, and the hardware that is required to train and deploy a complex ML model has fallen to levels that nearly any corporation, big or small, can afford. Additionally, the widespread repurposing of graphical processing units (GPUs), the specialized chips that power the graphics in high-end gaming computers, to support ML applications has allowed companies to harness what is really a massively parallel supercomputer to drive their ML projects without breaking the budget.
- Open source and lower barriers to entry - Over the last five years, there has been a significant influx of open source code into the ML universe. Combined with standardization, this has significantly reduced the barriers to entry for companies to experiment on becoming more proficient at ML techniques. Some of the notable open source frameworks include Google’s TensorFlow, Amazon’s MXNet, and Facebook’s Torch. Furthermore, the online education market for machine learning is growing at a rapid pace, driven by demand and increasing awareness. Key vendors such as edX, Udacity, NobleProg, and Ivy Professional School have grown quickly during the last few years.
- Accessibility via the cloud - The availability of massive cloud computing capability (e.g., Amazon AWS, Microsoft Azure, Google, etc.) – including specialized ML services and GPU grids on demand and at modest cost, and supported by terabyte- and petabyte-size data infrastructures – has allowed complex ML models to be trained on massive data sets without the need for correspondingly massive capital investments.
These trends continue to accelerate, allowing companies to deploy increasing levels of compute power even more easily and cheaply. This acceleration will ensure that accessibility of machine learning by business and technology teams will also get more straightforward. It doesn’t hurt that billions of dollars of venture capital are being used to further develop AI, generating innovation and improving the proprietary and open source tool sets available to companies.
Fraud Is Still a Key Payments Use Case
Not surprisingly, transaction-fraud detection and avoidance continues to be an area where machine learning is being deployed within the payments industry. What’s changing is that many more companies beyond early adopters such as PayPal are starting to leverage the technology, and the data sets being utilized are growing beyond those that traditionally were used for fraud detection.
Historically, fraud-detection algorithms consumed a fairly small set of information on the parties involved in a payment: the amounts, and the source and destination of the funds. This lack of sophistication stemmed from the absence of more extensive data sets and a lack of compute power to do anything complex within legacy fraud engines because of the massive transaction quantities that need to be processed very quickly.
It doesn’t hurt that billions of dollars of venture capital are being deployed into AI, generating innovation and improving the proprietary and open source toolsets available to companies.
Available data sets are growing almost exponentially, making it possible to consume an ever-widening set of information in an attempt to understand who is generating a payment and whether they are allowed to do so. ML-based fraud detection and prediction systems are leveraging information such as the device fingerprint of the phone, tablet, or computer used to initiate the transaction, location data, third-party data on the parties involved in the payment (social media posts, for example), and information from the deep web (unindexed but visible) and dark web (hidden, not visible). Network analysis of fraud – including waves of similar transactions with different parties, cascading transactions that bounce payments from party to party, and other multistep transactions – also are being evaluated to improve fraud detection.
The only way to effectively process all this information to determine whether a transaction is fraudulent is to leverage machine learning. The challenge is even more difficult given that fraudsters keep changing techniques so the algorithms need to “learn” to continue to be effective. The emergence of true real-time payments has further decreased the amount of time a fraud detection process has to execute its analysis.
Identity Is Becoming a Harder Problem to Solve
With the recent Equifax data breach, the information that is available to support identity verification processes has been significantly reduced. Establishing identity is at the root of all fraud processes, because account takeover is a primary way that fraudsters profit from their activities. If the person originating a real-time payment isn’t actually the authorized owner of the asset being transferred, then the transaction must be blocked before it originates, because reversing a real-time transaction is often impossible.
ML can support establishing identity in two key ways. As with transaction-fraud detection, the first is to leverage increasingly larger data sets to determine whether the person attempting a transaction is actually who he or she says. These data sets fall into three key categories:
- Information on the person originating and receiving a transaction, including demographics, the device being used, etc. This information is then pattern matched using an ML algorithm to predict whether the parties involved in a transaction are likely to be valid and authorized.
- Information on the context of the transaction, including its type, size, date and time, geographic proximity to the goods or service being purchased, etc. ML can be used to predict whether a transaction fits the pattern of what similar parties are expected to do, or whether it is outside the bounds of what is expected and therefore the parties might not be who they claim to be.
- Information on the behavior of the originating party. Some of the more interesting aspects of identity verification use ML to match the current behavior of an individual to behavior he historically has displayed. For example, is this person typing with the same cadence as he normally does?
Biometric technology is one way that identity potentially could be definitively established before a transaction is initiated without introducing significant friction in the customer experience, which happens with many two-factor authentication schemes. But biometrics such as fingerprints and face scans have not yet been proven to be infallible, so ML techniques most likely will continue to play a part in establishing identity.
KYC and AML Efforts
The risk of not effectively performing know-your-customer (KYC) efforts and identifying anti-money laundering (AML) activity is well understood in the financial services sector, given the massive fines that have been assessed by regulators for process failures. Yet KYC processes are complicated, slow, and often ineffective, and AML surveillance systems still rely on fairly basic algorithms, resulting in the need for armies of investigators to evaluate alerts. ML technology may radically change how KYC processes and AML surveillance is performed, and innovative organizations already are seeing compelling results.
There are many potential applications for machine learning within the real-time payments market that likely will be identified and exploited by innovative companies.
Today, KYC processes rely on a basic match process between the identity information provided by a prospective customer and that provided by a range of third-party data sources, including credit reports, watch lists, and other collections of consumer and business information. The problem with this approach is that if retail customers aren’t in those databases, they essentially are excluded from the global payments system. For business customers, the challenge for KYC processes is that malicious actors have many techniques to obscure their real identity, putting financial services firms at risk of processing a transaction that ultimately turns out to be in support of money laundering or terrorist financing.
In the KYC process, ML engines can consume vastly greater quantities of information on a customer than is possible with today’s methods. By expanding the scope of the review, the rural farmer or furniture maker in Africa has as much opportunity to be successfully approved to transact business as the executive who works in midtown Manhattan. ML algorithms are likely to be more effective in predicting whether either the farmer or executive is actually who he says he is.
Similar challenges exist with AML surveillance, because the information necessary to evaluate whether a transaction is illegal is limited, resulting in too many alerts and a complicated and time-intensive investigation process. ML techniques also promise to extend AML surveillance from the simplistic approaches used today to sophisticated surveillance consuming a vast set of information, not only on the individual parties in the transaction but other parties they have transacted with and the types of transactions they have performed. This will not only reduce false positives (those pesky alerts investigators need to run down), but, more importantly, reduce false negatives (activity that needs to be investigated but isn’t spotted or reviewed).
There are many other potential applications for ML within the real-time payments market that likely will be identified and exploited by innovative companies. They might include near-instantaneous analysis of payment intent and its use to better understand a customer’s needs and wants. The applications might extend to systems for accounts payable and accounts receivable, automatically facilitating payments across supply chains to both protect the parties involved from fraud but also optimize cash flow and the use of lines of credit. They could also be used to let you tell your phone to order dinner from your favorite restaurant and have it delivered as you walk into the elevator at the end of the day, with seamless real-time payment for your meal that your agent facilitates. Given the power of machine learning to consume immense quantities of data and predict outcomes, the sky is the limit when it comes to potential uses.