Feedzai has one mission: to make banking and commerce safe. This mission drives every individual and department at Feedzai, but perhaps none more so than the Product and Research Teams. In 2019, they personified our mission by developing two groundbreaking technological advancements: omnichannel case manager and automatic model monitoring. Let’s take a look at each of these to understand what they are, why they’re critical, and how they help make banking and commerce safe.

Omnichannel Case Manager

The problem with traditional case managers for financial crime

Case managers, the applications used by fraud analysts and investigators to review alerts, are burdensome at best, and impede fraud detection and prevention, at worst. The problem lies with the siloed functionality case managers provide. Today, each customer channel — ATMs, debit cards, credit cards, online banking, mobile app banking, customer segments — has its own case manager. Multiple case managers mean analysts must manually cobble together a holistic view of a customer’s behavior and transactions to determine if those transactions are fraudulent or not.

For example, say that John Smith withdraws $500 from an ATM in London, though he usually withdraws $100 from an ATM in New York. Perhaps John is traveling, and the transaction should be approved. On the other hand, maybe there’s a fraudster at play? To gain an omniscient transactional view of John, the ATM fraud analyst would have to open up the case managers for several different channels, find John’s transactions in those case managers, and determine if there are any other suspicious activities related to John. By manually comparing John’s transactional behavior across case managers, she decides if the transactions should be approved or denied. In this case, by researching several case managers, she saw that “John” made a credit card transaction for online purchases with an IP address from Mumbai within five minutes of his ATM transaction. Now the analyst has the complete fraud story, rejects all transactions, places a hold on John’s accounts, and reaches out to the customer. Before she can celebrate her fraud-fighting victory, she has hundreds of other alerts to look through, which means thousands of case managers to review.

The reality with industry-standard case managers? It takes a fair amount of diligence and manual detective work on the part of the fraud analyst (which is ripe for human error) and far too much time to reach this conclusion. If we’re talking about retail fraud analysts, who are under incredible pressure to review an infinite amount of alerts in the shortest amount of time, this process creates undue stress while inhibiting success for both the analyst and the institution. Case manager applications, by their very nature, are the antithesis of efficiency.

Creating an omnichannel case manager

The obvious solution would be to create an omnichannel case manager (OCCM) that supports several workflows and reports on multiple use cases and channels. Of course, what’s obvious isn’t always easy.

Single-channel case managers exist because different client segments and channels have different fraud types, APIs, models, and rules, which all necessitates a specific data science environment for each channel. To further complicate the issue, institutions utilize a variety of tools, systems, and platforms across functions and departments.

Undoubtedly, this is the reason (despite marketing claims to the contrary) that a functional OCCM didn’t exist. Until now.

How to approach creating an omnichannel case manager

Feedzai’s determination to create an OCCM was born of a client’s desperate plea to help streamline their processes and improve operational efficiency, not to mention employee morale.

In accepting the challenge, the Product team took a hands-on approach. “We traveled to several clients and sat side by side with their fraud analysts. We wanted the firsthand experience of their jobs and pain points,” said Joana Goncalves, Feedzai’s Case Manager Product Manager. “It was only by working as fraud analysts day in and day within a variety of financial institutions, that we understood the needs and precise requirements for this project. Everything we learned and experienced informed the development of the OCCM.”

This real-world research provided the team with practical insights. It allowed them to understand that the analysts looked at transactions in groups and required the historical context of the transactions, along with the geographic location.

Once they completed the research phase, the Product team gathered in Feedzai’s Lisbon office to map out their approach to the OCCM project. When asked how those first few days were, Joana laughs. “Chaos and panic,” she says. “Lots of it. OCCM presented an enormous technical challenge. But we did it.”

Indeed they did. Feedzai’s new omnichannel case manager provides a single case manager for all channels and systems, including competitor platforms. Analysts have an integrated view of customers, increase their efficiency, and improve their accuracy, and because they can easily access historical data.

A single case manager for fraud and anti-money laundering

Of course, once a product goes live with a client, there are surprises. Will the product work as expected? What are the bugs that need fixing? In the case of Feedzai’s OCCM, there was a genuine surprise. Their clients’ Solutions Team used OCCM for anti-money laundering (AML) and transaction fraud (TF) in the same data science environment. Using one data science environment equips the compliance and TF teams with an exhaustive view of the customer, albeit with different permissions.

Next steps for omnichannel case manager

OCCM will launch with a new client and use case — merchant monitoring — in the next few months. Of course, as we’ve learned, OCCM accommodates multiple use cases, which means merchant monitoring and TF can exist in the same OCCM.

As OCCM goes live with more clients, Feedzai’s Case Manager Product Team will continue to develop the product for greater functionality, with an eye on eventually replacing the classic configuration case manager with the omnichannel case manager for all clients.

Automatic Machine Learning Model Monitoring

Machine learning models drive innovation across industries. From predicting illnesses and recommending treatments to detecting and preventing financial crime in financial services, machine learning disrupts the status quo and propels organizations into the future. But there’s one thing about machine learning many people don’t realize: it can be difficult to know if machine learning models produce accurate results. This is especially true in the financial crime scenario, where criminal intervention can take months to be reported.

Catch My Drift: Machine Learning Models

Machine learning (ML) models, like the ones Feedzai builds, process millions of transactions a day. We train the models on the most current fraud data so that analysts can make accurate decisions. But fraudsters also “train” based on what has or hasn’t worked for them and adjust their schemes accordingly – something known as “concept drift.”

Concept drift is when an attack results in the machine learning model adjusting the risk score’s distribution because it is factoring in the data from the attack. This happens because the attack enacts a large number of similar transactions.

Data scientists and engineers are well aware of the problem. Yet, there hasn’t been a standardized approach to address it. Concept drift remains an open issue in the financial technology industry.

Inspired by this challenge, the machine learning Research Team at Feedzai set out to develop a standardized solution.

How Feedzai created an automatic machine learning, model monitoring feature

To know when a new attack targets our model, we first have to understand what normal behavior looks like. We gained that insight by building a reference model from a score distribution based on two weeks of data. The model looks at the current distribution of risk scores and compares it with the reference model. When the model detects deviations from the pattern, it automatically triggers an alarm. This alarm includes a report with the most relevant transactions that contributed to the drift in the scores.

The ML Auto Model Monitoring (AMM) system detects changes in behavior caused by a bot attack or a glitch in the system (such as clients leaving a particular report field blank) that would otherwise go unnoticed for days.

AMM was no easy feat. As Luis Cardoso, AMM Technology Lead, explained, “The Model Monitoring algorithm developed by Research was very difficult to implement because of all the subtleties and details of the algorithm. It required close collaboration between Product and Research, but in the end, we have a solution that no one else in the industry has.”

Indeed, AMM is a significant technological achievement for Feedzai, but it’s not the completed vision. Client deployment is the next step for AMM. “We’re excited to partner with a client and go live with AMM,” Cardoso said. “This is just the beginning of our story.”

Want to understand the technical aspect of how Feedzai developed ML automatic model monitoring? Read ML-Powered Automatic Monitoring on Feedzai’s Tech Blog.

2020 vision: a peek at what’s next

In 2019, Feedzai created two groundbreaking technologies that live and breathe our mission to make banking and commerce safe. That’s an impressive year by anyone’s standards. How can 2020 compare? “We’re not just thinking of a new year, we’re thinking of a new decade,” said Saurabh Bajaj, Feedzai’s Chief Product Officer. “We know payments will continue to evolve, and fraud is no exception. As Feedzai enters this new decade, we’re not only focused on enabling FIs to improve detection; we’re empowering the strategic transformation of both established and emerging businesses. By driving decisions based on data and advanced machine learning, we’re building technology that isn’t reactive to change, but prepared for it.”