SynapseDDM Spots Behavior Changes
Pattern-detection systems themselves are nothing new. They have been widely deployed for years to detect fraud, insider trading and money laundering. They have been used less frequently in marketing applications such as identifying customers at risk of attrition or ripe for new purchases. They are also a component of the federal government's proposed behavior-surveillance systems.
One constraint on pattern-detection systems is that users usually must define in advance which patterns are worth detecting. This is typically a laborious manual process based on examining historical instances of, say, credit card fraud, and finding patterns that are common among fraudulent transactions but uncommon among legitimate customers.
Selecting patterns to test requires a fair amount of hypothesizing, experience and intuition. Automated methods can generate large numbers of patterns and identify the most predictive, but even this approach will capture only patterns that repeated often enough to be distinctive. This is a problem in detecting illicit behavior, where criminals will purposely vary their actions to avoid conforming to known patterns.
Marketers face a slightly easier problem, as consumers are not usually quite so devious. But marketers still must find patterns that frequently repeat.
SynapseDDM (Synapse Technology, 704/887-5600, www.synapsetechnology.com) takes a different approach to pattern recognition. Instead of trying in advance to identify specific patterns that indicate particular behaviors, SynapseDDM simply tries to identify individuals who have deviated from their own past behaviors.
The system focuses on evaluating bank deposits - the DDM stands for Daily Deposit Manager - so an example would be a customer who repeatedly receives the same size deposit every two weeks and then doesn't. In the real world, the deposit is likely a paycheck and the halt indicates a change of position or unemployment. But SynapseDDM doesn't offer this interpretation; it simply notes the deviation. It's up to the marketer to decide how to react.
This approach avoids the challenge of defining significant behavior patterns in advance, but it brings other challenges. One is simply the volume of data that must be stored. Because the system must check constantly for new or discontinued patterns, it must examine individual transactions rather than summary measures. Synapse suggests the system maintain 25 months of transactions, which might translate to 500 transactions on an average account. All of these would be re-examined nightly.
A second challenge is the number of patterns the system must check - up to 48,000 per transaction, depending on how many transactions are available. This analysis is done at the individual and household levels so patterns that span multiple accounts can be identified.
Daunting as these quantities seem, Synapse handles them easily: The system processes 3 million transactions per minute on a desktop PC. One reason is that the transaction records themselves are quite small, less than 20 bytes, since they hold only an account number, date, amount and transaction code.
The greater challenge is determining which behavior changes are worth closer examination. SynapseDDM uses advanced statistics to measure the degree to which a transaction differs from past behaviors. The precise statistical methods are a well-guarded secret, but the general approach is to look at differences from past transactions in terms of amount, timing, channel, location, account type and other attributes.
SynapseDDM has about 100 parameters that set thresholds for the degree of deviation that is considered significant. About 25 can be set through the user interface, and the balance are internal. Settings are tuned to each client's needs during system implementation. A simulation environment, using actual customer data, lets users examine the effects of changing parameter settings.
But the most unusual transactions are not necessarily the most important. So after the anomalous transactions are identified, SynapseDDM applies business rules to screen out transactions that are irrelevant from a marketing perspective and to prioritize the rest. For example, transactions involving highly profitable accounts might be given higher priority than changes in low-profit accounts.
The final step is to do something. Though SynapseDDM lets users associate different behaviors with labels such as "attrition risk" or "cross-sell opportunity," it does not explicitly predict future behavior or suggest specific offers. Rather, it distributes alerts in the expectation that the bank will contact the selected customers to see what's going on. These alerts include customer and account information, the list of recent transactions, an explanation of what makes these transactions unusual and associated information such an "attrition risk" label.
SynapseDDM supports this approach with a simple but adequate contact management capability that can maintain lists of agents such as customer relationship managers, execute distribution rules to determine which alerts are sent to which agent, limit the number of alerts based on agent capacity, present alerts to agents and capture call results.
These results are later used to refine the alert selection and ranking rules. This refinement is a manual rather than automated process, though SynapseDDM does provide reporting and analysis tools to help. It also supports a test/control methodology to measure system impact.
SynapseDDM uses Microsoft technologies including the C# programming language, .NET Web services framework, SQL Server database and Windows operating system. Agent workstations also must be Windows machines, though no software is required beyond Internet Explorer.
The system can be purchased as a monthly service or conventional licensed software. Fees depend on the size of the bank and range from $3,000 to $25,000 monthly for the service or $50,000 to $200,000 for a license. It takes Synapse about a month to install and tune the system for a new client. SynapseDDM was released in March and has one active customer.