4 machine learning methods that fight fraud
As the world moves toward a digital payment landscape, fraudsters have evolved their methods from stealing cash and checks to exploiting electronic payments, ACH and wire transfers.
E-commerce sales in the U.S. have increased to $3.5 trillion in 2019 from $1.3 trillion in 2014, and are expected to reach $6.5 trillion by 2023, according to a recent Aite Group report, “The Bull’s-Eye on Financial Transaction: Keeping the Fraudsters in the Crosshairs.”
With this rise in adoption and usage of digital payments, fraud has also spiked. Fraudulent activities in which a card is not present have risen in the U.S. to $5.5 billion in 2019 from $3.2 billion in 2015, and are expected to reach $5.9 billion this year, the report stated. Financial institutions, fintechs and merchants are deploying AI in the form of machine learning to fight rising transactional fraud, according to the report. This is how the top four methods identified in the report work:
- Machine Learning
Algorithms based on data sets produce a fraud probability score that can be used to automatically detect fraudulent transactions. These algorithms are used by data scientists to identify the characteristics of fraudulent transactions and detect in a data set. “The larger the data set, the more accurately the model will perform,” the report stated.
Unsupervised machine learning, on the other hand, does not require an existing data set or a data scientist, and instead depends upon the model to look for anomalies and learn on the fly how to determine fraudulent activity.
- Rules Engine
A rules engine determines whether a transaction is fraudulent. If the criteria set by merchants or FIs for a transaction are not met, that transaction can be denied.
Rules are static or “velocity-based.” Under velocity-based rules, multiple transactions in rapid sequence may be denied. When machine learning is integrated into a rule it “tends to deliver the highest level of fraud prevention with the least number of false declines.” This is because machine learning can pick up individual buying behavior, the report stated.
- Alert Management and Alert Engine
Alert managements are deployed for transactions with risk that is elevated but not high enough to be immediately declined. Transactions that appear to be fraudulent are then confirmed by the customer. Alert engines communicate with the consumer to verify suspicious transactions via text, email and phone.
- Orchestration Hub
The orchestration hub method for evaluating data combines and detects multiple data sets, such as device fingerprinting, email address profiling, behavioral and physical biometrics in order to detect fraudulent transactions in real time, according to Aite.
Bank Innovation Build, which takes place Sept. 9-10 as a virtual experience, is a must-attend industry event for professionals overseeing financial technologies, product experiences and services. Register here.
