Artificial intelligence plays various roles in real-time detection and prevention of common instances of fraud. Among the most advanced examples of machine learning (ML) for fraud detection currently in use is PayPal's system, which uses a deep neural network that H2O.ai developed. The system disallows fraudulent transactions proactively by looking for patterns of behavior that deviate from customers' past transactions. Developers train the networks to recognize new patterns as fraud artists develop new methods.

In response to large-scale financial crimes, governments require banks to rigorously monitor their systems for evidence of money laundering and financing of terrorism. In response to regulatory tightening, financial institutions have adopted analytics software to help them detect such crimes and take appropriate action. Some companies are using and developing advanced analytics software that uses machine learning to assist analysts and investigators in detecting, investigating, and stopping illegal financial operations.

For example, Fico's Anti-Financial Crime Solutions software suite uses a model based on unsupervised Bayesian inference techniques. The software aggregates customer transaction data and creates archetype labels that are based on customer behavior and change in real time, and notifies analysts about suspicious activity. Another example of analytics software using ML has been developed by VC-backed start-up Ayasdi. Their solution uses topological data analysis to group customer data graphically based on financial behavior, make predictions based on connections between data, and provide justification for flagging suspicious activity. The software platform provides detailed information for investigative teams and suggests updates to customer group profiles based on behavior. In a pilot of the software, Ayasdi reports that it helped HSBC reduce the number of fraud false positives, saving the bank money that would have been spent on further investigation.

Other companies purportedly applying machine learning techniques in available financial-crime analytics software include Intel, Nice, and QuantaVerse. Oracle indicates it is delaying market entry until it can use ML to provide explainable results and to automate regulatory filings.

Implications

Although machine learning and deep learning could potentially improve analytics software, whether the degree of improvement will be meaningful enough to justify switching away from existing software and practices is an open question. Compared with ML solutions, companies may find that simpler analytics systems are easier to implement and allow easier verification of regulatory compliance. Alternatively, if analytics systems can apply deep learning techniques transparently, they may be better suited to addressing evolving financial crime that continues to be more complex and geographically widespread. Ultimately advanced analytics systems will likely be measured on their ability to correctly determine cases of fraud, present step-by-step reasoning, and reduce costs of regulatory compliance.

Impacts/Disruptions

Analytics software that uses deep learning techniques may be able to perform new, advanced operations on large sets of data from various sources and in the future automate many tasks previously completed by humans. Advanced analytics systems may be able to autonomously analyze data, learn from investigations, and ultimately embed entire processes in systems. In the future, systems may not only reduce the cost of compliance, but may also entirely replace human labor. It may eventually be possible for these advanced systems to be less transparent if they are ultimately more effective than traditional analytics software.