In this piece, Harry Powell, head of industry solutions at TigerGraph, explores how ​​financial services institutions can lean on machine learning algorithms to improve the accuracy of their fraud detection efforts and clamp down on suspicious activity.Â
In these unpredictable times, there is one pattern that many experts agree on: When the economy goes down, fraud goes up.
We saw this during the 2008 recession, when investment fraud and consumer scamsOpens a new window were common, tapping into people’s fear of uncertainty. During this time, many banks suffered losses thanks to collusive fraud rings and credit card fraud. Years later, during COVID-19, CNP (card-not-present) fraud spiked, as did phishing attacks aimed at nervous consumers. Last yearOpens a new window broke records for data compromises, breaches, and fraud across the board. According to a recent World Bank Global Economic Prospects report, the pandemic, coupled with the war in Ukraine, has resulted in a global economic slump; global growth is expected to shrink from 5.7 % in 2021 to 2.9 %by the end of 2022.
 Financial institutions are investing in fraud detection systems, yet they are seeing diminishing returns regarding the effectiveness of these systems. Is it a data problem, an algorithm problem, or both? How can financial institutions improve the accuracy of their fraud detection efforts and clamp down on suspicious activity?
The Race Against Fraud
Every $1 of fraud loss costs Opens a new window U.S. financial services firms $4.00 — and one survey Opens a new window found that 46 % of businesses worldwide have experienced fraud within the past two years, with losses totaling $42 billion. Financial institutions fight a constant battle against fraudsters. After all, fraudsters can exploit any weakness within a complex system.
Meanwhile, financial institutions continue to invest in fraud investigation teams and detection systems —all harness the latest technology, such as machine learning algorithms, to combat fraud.Â
Financial institutions are hitting a wall regarding the performance of the algorithms that drive their fraud detection systems. One global study revealed more than 95 % of bank transaction alerts are false positivesOpens a new window . This means legitimate transactions are flagged as suspicious — and accounts and transactions are locked or shut down. The result: Billions of dollars lost each year in wasted fraud investigation time, potential regulatory fines, and reputational damage. Current algorithms are reaching their natural accuracy limits. The problem isn’t with the algorithms but the data models on which they are built.
Breaking the Accuracy Barrier
Machine learning algorithms have benefited fraud detection systems and have helped flatten the fraud curve. However, there is an accuracy barrier in fraud detection machine learning. Data scientists have tweaked the algorithms as far as they will go — given their current operational paradigm. Meanwhile, fraudsters are getting more creative by the day, and fraud rates continue to climb.
Let’s drill down on the data input into these fraud detection systems. This data consists of transactional data and information about the parties involved in a transaction. The system computes a fraud score by examining the nature of the transaction and the history of the parties involved; it uses various algorithms to generate an overall fraud score based on risk factors. However, this score is limited by the scope of the data fed into the system. Relationship-related data is missing — information about transaction parties’ relationships with other entities in the financial ecosystem. These entities may be people, devices, or high-risk organizations with fraud risk scores attached to them. This additional data would help contextualize the transaction — and any fraud detection score is incomplete without this relationship data.
Hidden Patterns, Connections, and Relationships
What’s needed to close this relationship data gap? Financial institutions should run various algorithms powered by a branch of mathematics called graph theory. These algorithms detect patterns that would otherwise be hidden within massive amounts of data — and these patterns can reveal flows of money and chains of relationships that would be hard to spot by viewing a party in isolation.
These algorithms that help uncover complicated relationships and flows include:
- Closeness algorithms – This type of algorithm reveals how close an account is to others with a high risk. An example of a closeness algorithm is the “shortest path,†or the shortest distance between Entity A and Entity B. If B has a higher fraud risk, then, based on how close it is to B, A may also have a high risk of fraud.
- Centrality algorithms – This algorithm finds an account at the center of an abnormally high number of capital flows and flags it for further investigation.
- Community algorithms – This algorithm analyzes large numbers of relationships to identify communities containing suspicious accounts. The reasoning is simple: If the fraud score of the community is high, then individual accounts within the community may be fraudulent as well.
Here’s how these algorithms connect the dots between parties and transactions to expose hidden relationships:
If Person A is linked to Account B, we can drill down to see that Account B has been accessed by Device C on numerous occasions. We then see a chain of links showing that Account B has transferred funds to Account D. Looking at Account D, we find that it has a high fraud score that could reflect badly on Account B. Therefore, we can flag Account B for further investigation.
See More: New Technology Takes on Call Center Fraudsters
Adjust the Data Model, Boost Fraud Detection Rates
When you view data through a graph-shaped lens, you can more easily identify outliers, influencers, and even communities within a vast web of information. When financial institutions add graph features into existing fraud detection systems, they tap into previously disconnected data to compute more accurate fraud detection scores (read: fewer false positives and negatives).
Also, graph analytics can generate explainable models. In other words, the algorithm generates a fraud score and reveals the specific connections that contributed to that computation. This additional information gives fraud investigators actionable fraud intelligence as they dig deeper into suspicious transactions and accounts.
As fraud numbers continue to rise, financial institutions need every available tool in their arsenal to stay one step ahead of fraudsters. These organizations can supercharge their accuracy algorithms by re-examining the data paradigm that powers their fraud detection machine learning systems. This is the preemptive and proactive way to break the algorithm accuracy barrier, maximize the power of data, and take a bite out of fraud.
Which steps have you taken to acquire fraud intelligence? Share with us on FacebookOpens a new window , TwitterOpens a new window , and LinkedInOpens a new window . We’d love to know!