At a Glance
A national European energy provider
Using data to identify possible fraud without devoting large resources to manual review.
Combining dynamic time warping and anomaly detection algorithms in order to generate a probability of fraud for each customer.
Expertise & Technology
Many companies have to deal with the issue of customer fraud and lost money due to false claims and representations. Many of these companies however have large amounts of data on their customers due to the nature of their business. Using this data effectively can allow companies to quickly identify possible cases of fraud without devoting large resources to manually reviewing cases. In one case we worked on, a national European energy provider was having issues with fraudulent customers. These customers were submitting false meter data and thus underpaying for their energy usage.
SFL works to provide clients with not just a list of potential fraud cases but a ranking of all customers in order of their likelihood of fraud. This allows our clients to use their fraud detection resources in the most efficient way possible by looking at only the most likely cases. In the case of the European energy provider, SFL combined dynamic time warping and anomaly detection algorithms in order to generate a probability of fraud for each customer in the data set based only on their time-stamped energy meter data. The anomaly detection algorithms work by identifying unexpected changes in the energy data and the dynamic time warping was used to identify similarities with past customers. Thus, SFL’s approach compared the customers both to their own past behavior and to other past cases in order to detect fraud.
Hundreds of suspicious meter readings were detected, and allowed the company to inspect high probability fraudulent customers. This automated analysis saved thousands of human hours and significantly reduced fraudulent underpayment.
There is potential for both fraud and incorrectly calibrated/broken meters for almost all companies. The benefits of this type of analysis is not limited to anomaly detection in just time series data, indeed we can apply very similar models to credit card spending, insurance claims, and any type of prorated billing. In each case, companies can get fraudulent activity probabilities for each customer and from here, they have a wide range of latitude in terms of how they want to use this data. They can look at only the top few most likely cases of fraud or can work their way down the list. No matter how many of their customers they choose to manually review they can always be sure they are using their resources as efficiently as possible by looking at only the most likely cases.