Tracking patterns of customer behavior seems like a great way to identify marketing opportunities, but sophisticated systems to do this have not been adopted widely. Products including Verbind (now owned by SAS), Harte-Hanks Allink Agent and Elity Insight have been sold to just a few customers.
Though many more companies have bought generic pattern-detection software, this is less likely to be the heart of a behavior-based marketing program.
While marketers have been slow to apply behavior tracking, it has gained broad acceptance in operational systems to detect fraud, money laundering, insider trading and bankruptcy risk. Vendors include Actimize, Atchley Systems, Magnify and SearchSpace. The technology is gaining a still higher profile as part of surveillance systems used in national security projects.
Whatever the application, pattern-detection systems must meet three requirements. The first is to gather data. Typically these systems work with detailed transaction data, often in huge volumes and in real or near real time. A system to prevent credit card fraud, for example, might scan millions of credit card transactions an hour, taking seconds to compare each transaction with previous ones for the same and similar customers and return an authorization or rejection.
Many applications must consolidate data from multiple sources, particularly in security screening, where travel, communication, financial and other types of information must be viewed together before suspicious patterns become apparent. Money-laundering and insider-trading detection also look at multiple accounts or individuals to find possible connections.
The second requirement is to identify relevant patterns. This has two components: defining patterns to look for, and finding when those patterns occur. Pattern definition most often is manual – that is, an analyst must pore through historical data and determine which patterns are significant.
Usually the analyst has automated data mining tools to help. Some systems do offer completely automated pattern definition, typically using neural networks or rule-induction technology. Automated systems are particularly helpful at identifying new patterns as these evolve. But even automated systems need manual supervision to avoid too many false alarms.
Once a suspicious pattern is defined, the system must be able to search for it. Since many patterns involve multiple transactions over time, access to past information is required. Many systems generate and store summary measures to avoid reprocessing all past transactions each time a new search is conducted.
For applications such as flagging identity theft or credit card fraud, patterns may be defined in terms of deviation from past behavior: If someone who never leaves home suddenly makes purchases in different cities on successive days, this is suspicious. The same behavior may be normal for a frequent traveler.
The third requirement is to react when significant behavior is detected. In marketing applications, the reaction often is no more than a message, such as a product offer or sales contact. It’s reasonable to generate this automatically; after all, the cost of an error is small. But with detection applications the stakes are usually much higher: You are about to accuse someone of doing something illicit. So the response is likely to be a manual review of the situation before further action is taken.
This brings its own set of requirements for assembling case information, routing it to an appropriate analyst, ensuring the most urgent cases are handled first, escalating or transferring cases that require further review, ensuring low-priority cases are not forgotten, recording resolutions and comments and keeping an audit trail for legal and management reasons.
Mantas Behavior Detection Platform (Mantas Inc., 866/462-6827, www.mantas.com) uses technology originally developed for the intelligence community and later applied to monitor trading compliance and best execution in the securities industry. The platform includes the core set of tools to load transactions, define and identify behavior patterns and respond to alerts. The vendor also has a half-dozen prebuilt applications for specifics such as fraud or money-laundering detection.
Mantas loads data from external systems into its database, which uses any standard relational database engine. The database is typically populated in a batch process though the system can handle real-time feeds as well. Mantas relies mainly on third-party data-loading tools, though it does include its own text-mining tool for tasks such as extracting the recipient name from a money order. The system also can create calculated or summary values during the load.
Behavior patterns, which Mantas calls scenarios, are defined in advance by an analyst through a browser-based graphical user interface. Scenarios can identify specific activity sequences or find common attributes among separate entities. They also can incorporate lists of individuals or countries flagged for special treatment.
The vendor created hundreds of prebuilt scenarios, which are included with its applications. Users can define new scenarios or modify existing ones using data analysis tools included with the system. Mantas itself continually updates the scenarios associated with each application, and many customers simply use what Mantas provides.
As data are loaded, the system generates a profile of actual behavior for each entity. It can compare the profile against past behavior, behavior of similar entities or stated goals such as investment objectives. Significant behavior deviations or scenario matches can trigger alerts, which are prioritized and routed based on rules specified by the user.
Analysts can view all alerts related to the same entity, see the specific data and scenario that generated each alert, run reports that show past alerts, profiles and transactions for the entity, specify a resolution and record their comments in a case history. Managers can view the alert queue, see audit trails and control the activities permitted for each user.
The Behavior Detection Platform runs on large, multiprocessor servers. Clients are mainly large financial institutions, in some cases feeding the system more than 100 million transactions daily. The software originally was developed by SRA International, which spun off Mantas as a separate firm in May 2001. It has about 15 installations.