Case Study

Hazelcast Powers Real-Time Fraud Detection

Competent fraud detection systems can help organizations gain a clearer view of entities, relationships, and hidden patterns as they deal with financial crimes including payment card fraud, anti-money laundering, and anti-terrorist financing. Failure to invest in fraud detection systems can be disastrous for a company’s bottom line; recently a major European bank was fined more than £500m by British and American authorities for anti-money laundering failings.

Consumer financial transactions are also a target for fraudsters who are using increasingly sophisticated systems of their own, all at a time when today’s detection systems are facing increasing pressure. In the UK alone, payment card spending exceeds to £180 billion generated from a record 3.9 billion transactions. Total financial fraud losses across payment cards, remote banking and checks have been increasing consistently, enabled by scams and online attacks.

These rates of change mean that today’s fraud detection systems must be flexible to change, not only in their core logic but also in system topology. They must report anomalies in near real-time and be highly available 24/7. The Hazelcast Platform is ideally suited to delivering such systems.

Hazelcast can store account details, recent transaction history and scoring all in one place for powerful, comprehensive checks within the request. Short SLAs at the millisecond level means that these checks must be done very fast, and Hazelcast gives customers performance in a very cost-effective manner.

Customer Success Story

A top ten U.S. bank was hitting transaction rate limits when trying to apply fraud detection algorithms against customer data sitting on its old relational database platform. This technically-imposed restriction was causing them to break fraud detection SLAs, and ultimately became a blocker on new business. Hazelcast is being used to store 2TB of customer data, which will eventually grow to 5TB. Because this data is now stored in memory, the bank can easily process 5,000 transactions per second with expectations for this to grow to 10,000 transactions. To further enhance operational availability, they are also using WAN replication to synchronize multiple Hazelcast clusters to ensure the continuation of service in the event of a regional data center outage.

The performance advantages they gained from the Hazelcast Platform also reduced fraud loss by enabling more accurate fraud detection scoring. Hazelcast gave them the performance headroom to run multiple fraud detection algorithms in parallel in the transaction processing pipeline. Each algorithm scored the transaction based on a different combination of factors, with different weightings. The extra levels of fraud detection let them calculate a composite score that was more accurate than any single algorithm. This reduced financial loss from fraud and also helped to reduce “false positives” (i.e., legitimate transactions that are mistakenly flagged as fraudulent) which helped to retain transaction fees on valid purchase attempts.

Why Most Datastores Fall ShortWhy Hazelcast Is Ideal

Most data stores are complex and costly to scale according to demand. Many data stores cannot be installed to scale without operator intervention. This is an essential feature for systems that wish to operate cost-effectively, including in on-demand cloud environments.

Scaling a solution built on Hazelcast is as easy as starting an extra process. Hazelcast takes care of the rest, distributing data and tasks fairly, while the system is still running.

Client-side reactive messaging is not an available feature in most data stores, certainly not as a vendor-supported, first-class feature.

Hazelcast provides reliable, highly available messaging across not just data events but also system events. The events can be received by clients and members of the Hazelcast cluster.

For real-time analytics on large volumes of data, NoSQL and RDBMS solutions suffer from latency issues because data must first be written to the database before the additional business-critical processing can be performed.

As an ultra-fast, optimized platform that combines real-time stream processing and in-memory data access speeds, Hazelcast can ingest, transform, and score data in microseconds.

Most NoSQL architectures have single processes for writes and other processes for backups. Worse still, during times of failure for these write processes, the system will block further updates until a new write process is elected. This means blocking writes on the entire data set. Introducing sharding to solve these issues results in highly complex topologies.

Every process in a Hazelcast cluster is responsible for handling a portion of the main data and backups, there is no single master in a Hazelcast cluster, meaning that processes can be dropped without blocking writes to entire data sets.

Datastores do not allow sophisticated code to execute in situ with stored data.

Hazelcast can embed fraud detection code written in Java in its nodes where the customer account data is located, so that a large number of rules can be run within the SLA, all without needing to shift data over the network.