Welcome to the Pondera FraudCast, a weekly blog where we post information on fraud trends, lessons learned from client engagements, and observations from our investigators in the field. We hope you’ll check back often to stay current with our efforts to combat fraud, waste, and abuse in large government programs.
Almost everyone is familiar with antivirus software. Not everyone is familiar with how it works though. Even fewer have examined how we can apply the way antivirus software works to combat fraud. I believe that there are important lessons here which can improve our approach to fraud detection and prevention.
At a high level, antivirus software performs two important functions prior to opening a file on your computer: 1) It compares the file to known viruses and other forms of malware, and 2) It checks the file for suspicious code which may indicate a new, previously unknown virus.
The first function depends on a network of users willing to share known viruses and a system that is able to collect the virus data, design a fix, and disseminate the fix to other users prior to them being infected. The second function depends on heuristic programmers that can design systems to learn and even anticipate potential problems. Working together, this is one of the most effective ways to address the constantly changing nature of Internet malware.
Government fraud prevention, when done properly, works in a very similar manner. By examining known bad actors, bad transactions, and bad behaviors, systems can quickly compare ongoing program data to identify suspect transactions. Modern fraud detection systems also include predictive algorithms that can detect anomalies, trends, patterns, and clusters that may indicate fraud.
Unfortunately, many governments are unable, or unwilling, to share data. This limits the “network” effect that antivirus software uses so effectively. If more states and programs shared fraud schemes and findings, the library of known bad actors and methods could detect fraud and prevent it from moving from state to state and program to program.
The good news is a number of states are moving toward state-wide fraud prevention efforts and a number of government subsidy programs are moving toward cross-state fraud prevention efforts. I am confident that the future success of these efforts will promote additional sharing, leading to a larger network, and more efficient governments.
Big data analytics and the predictive engines they spawned give Internet companies a way to monetize the online experience. By tracing online behavior, companies can target advertising, promotions, and point of sale opportunities based upon past buying decisions. Online habits of Web users can be associated with ideologies, interests, and values. Increasingly sophisticated probability engines predict future buying decisions with enough accuracy to fuel a dramatic increase in online sales and commerce over the past decade. Analysts use temporal versions of these tools to forecast market trends and evaluate risk.
In the fraud detection market however, early attempts to detect fraudulent behaviors using these same probabilistic engines have achieved limited success.
What makes detecting fraud different than detecting interests, values, and ideologies? The simple answer: Fraud is binary in nature–either a particular sequence of behaviors is fraud or it is not. For example, if an individual provider of medical services submits claims to an insurance company for 5,000 hours of services in a week (an instance from actual data), there had better be around 100 employees licensed to provide that service. If there are only three or four employees with the required licenses, the provider has committed fraud. Probabilistic engines struggle to detect fraud because they are not capable of modeling this“all or nothing”nature of violating a law.
At Pondera, we still make use of predictive analytics. But rather than detecting absolute fraud, we use the algorithms mostly to inform our fraud scores and to detect emerging fraud methods.
Once reliable methods of detecting fraud have been developed, predictive engines can also play an important part in helping insurance companies, financial institutions, and government agencies prioritize targets of investigation. Predictive models can identify the highest value targets that will recover the most money or disrupt the largest criminal organizations.