Ken Park, Head of Fraud Analytics & Operations and Joe DeCosmo, Chief Analytics Officer at Enova prep Fraud Force Summit attendees on what they can expect to learn during their session: how they use advanced analytics and algorithms to catch more fraud.
Enova is a growing, global online financial services company that provides quality credit to consumers in the U.S. and five other countries around the world. Because we offer our credit products online with minimal interactions, mitigating fraud risk has been a key requirement for the company's success. We have found the combined usage of advanced analytics and manual operations to be the key factor in achieving our goals. At Fraud Force we will talk about how analytics and operations, separately and together, have helped us consistently prevent fraud across our business.
We began using advanced analytics and modeling to prevent fraud a number of years ago. While we realize that not every fraud problem needs analytics, we have found analytics to be helpful in reducing manual work and scaling our businesses. For example, our small business lending product has a rigorous, manual process to verify applications. In this setting, manual operations are the sole solution, and they work well. However, even in such a process, we have applied analytic services to find areas where verification can be automated or bypassed (i.e. white-listed) in order to help the business scale and provide a better customer experience.
On the other end of the spectrum, we have our online consumer short-term and installment loan businesses for which individual loan amounts are low but overall transactional volume is large. In this scenario, we use analytics to selectively choose suspicious activities for manual investigation with great accuracy to help our businesses process thousands of applications per day as efficiently as possible.
Though the motivations to use analytics differ slightly, we have found that we generally go through two steps in using analytics to prevent fraud - generating variables and building models. We always start with a generalized framework where we categorize data sources and variables by different criteria.
For example, for each piece of data, we ask questions like, does this item exist, is it valid and does it make business sense, is there reliable history, does it occur as frequently as usual, and is it associated with other entities. In this way, we have created a very thorough list of variables that are fed into a model-building algorithm and avoid generating variables that fraudsters can easily manipulate.
Once we have the list, we feed the data into a machine learning algorithm of choice. Logistic regression and decision trees have been the popular choice for their simplicity in terms of developing, interpreting and implementing.
In this year's session, we will also talk about another approach we named Almond. Most of the known algorithms, we found, miss fraud cases often because fraud occurs rarely and is such an outlier in any data set. These algorithms often produce results that just simply ignore fraud cases or work well in the training data set but perform poorly in the validation data set. Almond is based on a very simple concept, which is to find all combinations of variables with a substantial volume that are associated with bad events with a high probability. Standard brute-force methods are very computationally and time intensive, so we took a known algorithm and invented some new features to make the collaborative filtering very fast. Because Almond produces a very extensive list – many of whose items do not make sense as a result of over-fitting – the data scientists and fraud investigators go through the list together to recommend sensible solutions. Though it is not a fully automated solution, we have found this method very helpful in finding the fraud segments that might have previously been missed.
In short, our aim is to create a faster feedback loop in which known fraud incidents can be detected, identified, analyzed and then prevented in real-time. With the generalized framework to create variables and our Almond algorithm to quickly explore them, we are getting closer to that goal. Of course, we cannot understate the significant role that the fraud operations team contributes to this shortened feedback loop. Analytics works well with rich historical data but not so well without it. The investigators help find emerging fraud cases and their common patterns that our models have never encountered. They also recommend new data sources that they have found to be effective from their day-to-day investigations. By combining both human intellect and advanced analytics, we are creating a feedback loop that is faster and cleverer every day.