A Thing
The Green SheetGreen Sheet

The Green Sheet Online Edition

September 13, 2021 • Issue 21:09:01

AI's evolution from predictive to prescriptive

By Dale S. Laszig

Twenty years ago, artificial intelligence (AI) was considered a breakthrough technology; today, AI is an integral part of connected commerce. Next-generation AI models are more autonomous than early models and frequently make decisions independently of humans. Their ability to parse anomalous behavior from massive data sets makes them invaluable to numerous endeavors from fraud prevention to financial services.

The Green Sheet interviewed technology leaders who leverage advanced AI technologies to protect the commerce value chain and improve the customer experience. In this article, they detail AI's remarkable evolution and offer insights into the increasingly sophisticated interplay between predictive AI machines and their human creators.

Innovation, opportunity

Sam Bobley, chief executive officer at Ocrolus, a fintech automation platform, works with numerous clients employing data scientists and has seen AI drive innovation and opportunity. "The reasons make sense: you don't build a differentiated lending business through brute force; there is no intellectual property associated with simply hiring more loan officers and making more calls," he said. "That effort is table stakes. What is significantly more valuable in the mid and long term is building proprietary risk models using loan performance data."

Andy Renshaw, vice president, strategy and solutions at Feedzai, a cloud-based financial risk management platform, observed that AI adoption is rising. "Interest in predictive AI has been increasing with proven success stories from early adopters, increased availability of data and low cost of cloud computing," he said. "The adoption of predictive analytics was accelerated by the COVID-19 pandemic as organizations, including financial institutions, had to rethink business processes and digital transformation objectives."

Felix Berkhahn, chief data scientist at AML software provider HAWK:AI, attributes growing AI adoption to consumer acceptance, which has changed in recent years, particularly in the financial field. "We see new customers who are more willing to use AI, but we also see it on the regulatory side," he said. "Here in Berlin, several big financial institutions are working together to build institutional monitoring solutions. You can really see, from all angles, we have changed."

Scalable foundations

Like a weather report or sales forecast, predictive analytics is more intuition than exact science; it's a way to anticipate outcomes based on current and historical data, which machines can do in seconds, researchers noted. On the Opportunities and Risks of Foundation Models, a study published Aug. 18, 2021, by the Center for Research on Foundation Models, drew sharp distinctions between traditional AI and today's advanced AI capabilities.

"Most AI systems today are powered by machine learning, where predictive models are trained on historical data and used to make future predictions," CRFM researchers wrote. "The rise of machine learning within AI started in the 1990s, representing a marked shift from the way AI systems were built previously: rather than specifying how to solve a task, a learning algorithm would induce it based on data—i.e., the how emerges from the dynamics of learning."

AI foundation models use neural networks and self-supervised learning to assess massive swaths of data, CRFM researchers noted while also cautioning against misuse. "Foundation models are scientifically interesting due to their impressive performance and capabilities, but what makes them critical to study is the fact that they are quickly being integrated into real-world deployments of AI systems with far-reaching consequences on people," they wrote.

Defining characteristics

David Snitkof, vice president of analytics at Ocrolus, mentioned the biggest change he has seen in predictive analytics is the sheer mass of diverse and disparate data sets, including data from unstructured documents used in credit risk modeling. "[O]ur world is awash in data, but much of this data is unstructured, noisy or in hard-to-reach places," Snitkof said. "With the use of AI to extract valuable signals from documents as well as digital data feeds, it is possible to leverage a greater variety of data and use it to offer more personalized financial products."

Nicole Newlin, vice president of solutions at Ocrolus, agreed, stating solution providers need to focus on their inputs. "The human plays a big part in providing the data a machine learns as it becomes self-sufficient," she said. "[Solution providers] need to remove bias or discrimination and manage incomplete data to build the best possible model."

Snitkof further noted that legacy AI solutions typically focus on credit bureau data, which may not present a borrower's full financial picture. He advised lenders to constantly evolve data collection, heightened probability and feedback loop processes.

Dynamic analytics

"The best lenders collect a lot of data, both in terms of quantity as well as variety," Snitkof said about data collection. "They evaluate signals from many different sources and explore how they correlate with outcomes, individually as well as in concert." Regarding heightened probability, he added, "Better than trying to predict the future (you know you'll be wrong, just not by how much) is to understand the distribution of potential outcomes and simulate the likelihood, impact and subsequent actions for potential future scenarios."

Explaining feedback loop development, Snitkof said, "[T]he best organizations build a learning machine, investing in data quality, thoughtfully designed data architecture, and the ability to construct a feedback loop that can increase the accuracy of analytics over time."

Berkhahn stated that next-gen AI is generally more contextual than prototypical AI solutions, which helps service providers authenticate legitimate customers and thwart criminal activity. "We want to know the true owner of a company, what type of company and the risk threshold of the company's transaction monitoring," Berkhahn said about contextual analysis. "That contextual information helps identify and block criminal activity."

Renshaw cited automation and machine learning as additional components of next-gen AI. Regarding automated processes, he said, "Predictive analytics solutions should also look for efficiencies beyond model performance. Automating various stages of the data science life cycle, from data migration to model creation to continuous model governance, will reduce operational overheads and ensure that the solution continues to stay robust."

On the role of machine learning, Renshaw said, "Robust and effective predictive analytics solutions should be based on the ML best practices to achieve performance, speed and adaptability." He added that solutions should be tailored to customer requirements.

Early-stage growth

Reflecting on AI's evolution, Renshaw identified three stages of growth: early-stage, predictive and prescriptive. Early-stage financial fraud solutions were driven by rules that looked for certain characteristics in transactions to identify, block and report fraud, he stated. Today, a rules-based approach is still relevant but has limitations and may even be counterproductive from an operational efficiency and management perspective, he added. 

"Legacy fraud prevention solutions detected and alerted on malicious behaviors associated with device-centric data,." Renshaw said. "Initially, this approach created ‘deny lists’ and ‘allow lists’ to identify good vs. bad users and were notoriously out of date."

As early models evolved, they became capable of detecting malicious activity by identifying malware, devices, and networks associated with fraudulent behavior, Renshaw said. From there, they attempted to identify individual users by analyzing their behavior and actions, he noted, adding that these efforts were only moderately successful, he added, because they relied on connecting identities to devices.

"Fraud solutions then evolved into predictive analytics based on artificial intelligence (AI) and machine learning (ML) models which detected fraud at different stages of the payment process, often in silos," Renshaw said. "For example, know your customer versus transaction fraud versus anti-money laundering solutions provide a fraud probability or riskiness score. ML models have evolved over time with more robust algorithms and strategies replacing legacy ones."

The predictive stage

Next-gen solutions focus on protecting end-to-end customer journeys across the global payments landscape, from enrollment and onboarding to continuous transaction and account monitoring. These solutions, which have become more sophisticated over time, are embedded into an organization's fraud and risk management processes, Renshaw stated.

Over time, advanced predictive analytics became more prescriptive by providing actionable intelligence and enhancing business outcomes when they detected fraud, Renshaw noted, adding that next-gen predictive AI uses robust ML models that are measurable and explainable.

"An example of advanced predictive capability is the Feedzai Genome—a dynamic visualization engine included in the Feedzai solution to show hidden connections between transactions," Renshaw said. "This provides an intuitive way for risk teams to proactively identify emerging financial fraud patterns and not just flagged fraudulent transactions. Risk teams can use this intelligence to put preventative measures in place."

The prescriptive stage

As advanced predictive analytic solutions evolved, they developed the ability to identify users (good or bad) from individual composite characteristics based on thousands of data points, Renshaw recalled.

They could recognize users based on digital assets such as behavioral biometrics (touch/keystroke, mouse movements, mobile spatial sensors, etc.); behavioral analytics (user journey, fast travel check, date/time of connections, etc.); and device and network data, he explained. These compiled digital assets create a unique digital fingerprint (which Feedzai calls a "BionicID") for every user, whether legitimate or bad actor.

"BionicIDs can identify legitimate users to speed them through their journey or prevent bad actors from committing fraud," Renshaw said. "Prescriptive analytic solutions utilize BionicID based data to identify good vs. bad users and can rely on biometric data alone to identify the differences between the two."

Advanced AI solutions like BionicIDs are difficult to copy, and their ability to provide a unique identifier facilitates accurate user identification and protects customers from manipulation and impersonation threats, Renshaw stated.

What's next?

Berkhahn pointed out that effective AI solutions employ federated learning and graph neural networks (GNNs), methodologies that enable technologies to operate holistically as part of a collective network. He described federated learning as a way to collaboratively build models without sharing sensitive data; graph neural networks are data structures composed of nodes connected across a network. "Research in federated learning and GNN has improved in recent years, providing additional opportunities for machine learning innovation," he said.

Renshaw summarized AI best practices as "models are robust, models are observable and explainable, model governance is enabled, model management is automated to drive operational efficiency and finally, models are fair." He urged AI resellers and end-users to incorporate these best practices. "Any AI model implemented should be explainable," he added. "Blackbox models reduce both customer and auditor trust."

Renshaw also emphasized monitoring of existing models and remediating performance issues. "AI is a powerful tool and should ultimately benefit the customers by providing fair, inclusive decisions," he said.

Snitkof expects AI to continue to expand in capabilities and scope due to its proclivity for continuous process improvement. Ocrolus leverages AI with humans in the loop, he stated, describing the process as attended ML applied to unstructured content. "The model employs predictive modeling and valuation metrics with a defined feedback loop based on human supervision," he said, adding it identifies document types, data fields and actual data, continuously improving as it learns from human input.

Newlin envisions a new generation of AI solutions in an environment resembling a vast hall of mirrors, as humans supervise machines that supervise humans. "When I think about machine learning with a human supervisor, it makes me think we could be in this loop forever, because now we would need another sort of AI and machine learning solution to check the review of the human who supervised." end of article

Dale S. Laszig, senior staff writer at The Green Sheet and managing director at DSL Direct LLC, is a payments industry journalist and content strategist. She can be reached at dale@dsldirectllc.com and on Twitter at @DSLdirect.

The Green Sheet Inc. is now a proud affiliate of Bankcard Life, a premier community that provides industry-leading training and resources for payment professionals. Click here for more information.

Notice to readers: These are archived articles. Contact names or information may be out of date. We regret any inconvenience.

Prev Next
A Thing