ONLINE LOAN FRAUD DETECTION
A provider of online microloans was troubled by a high ratio of frauds, driven mainly by chronic fraudsters.
Online loan applications are inherently subject to much higher risk than traditional branch walk-ins. This is largely because conventional strict screening process is not available when applying online.
We developed a machine learning algorithm to calculate likelihood of each application to be a fraud based on the similarity of applicant's device fingerprint to those of historical fraudsters. While the core of the solution is a self-learning classification model assigning fraud probability, we also provided a proprieatary parsers of raw data and other smart features for building the algorithm. Furhter, the solution was optimized for speed to be able to score the applications in near-real time. To ensure the maximum value addued of the model, the threshold for the fraud probability was automatically optimized given the cost of missed fraud and rejected non-fraud.
The investment into developing the model was paid off within almost just one month. In addition, our solution suggested a desirable change in the decision making process - highly probable frauds and high size loans now undergo furhter screening rather than being automatically denied.