CMStatistics 2022: Start Registration
View Submission - CMStatistics
Title: Amount-dependent fraud detection considering aggregated losses Authors:  Jorge C-Rella - Abanca servicios financieros (Spain) [presenting]
Ricardo Cao - Universidade da Coruna (Spain)
Juan Vilar Fernandez - Universidade da Coruna (Spain)
Abstract: Fraud detection is a significantly difficult problem due to an intrinsic extreme unbalance and class overlap. In classical approaches, the likelihood of belonging to the positive class is estimated and a threshold is selected in order to classify an observation. As this approach does not consider the amount, it produces suboptimal decision rules in cost-sensitive classification problems. A new approach is proposed constructing a two-dimensional decision space considering the two variables on which losses depend, namely the estimated fraud probability and the loan amount. This expansion allows more freedom to the decision region, from which a new proposed algorithm takes advantage in order to obtain the optimal decision rule in terms of aggregated losses on the sample. Classical classification rules are contained in the algorithm search, so an improvement is consistently expected, which is shown with a practical application and a series of simulations.