Leveraging Predictive Analytics for Fraud Detection
The Association of Certified Fraud Examiners estimates that the typical organization loses approximately 5% of its annual revenue each year due to fraud. When this estimate is applied to the expected Gross World Product in 2016 this amounts to a staggering 3.8 million US dollars of global revenue lost due to fraud. With the volume and velocity of transactions being carried out in a typical business continuing to increase it is not humanly possible for the approvers of these transactions to scrutinize each purchase order with the attention it ideally deserves.
John Keells SGIT has developed a solution which addresses this issue by developing a dashboard for flagging vendors based on forensic accounting techniques as a well as incorporating real time scoring for flagging fraudulent Purchase Orders. The dashboard provides metrics using techniques from forensic accounting such as Relative Size Factor, Benford’s Law and Z Score analysis to give the organization a holistic view of the vendors it transacts with for effective mitigation of risk. The dashboard enables drilling down to the Purchase Order line item level to identify anomalies flagged by the above metrics in an intuitive, easy to use manner.
The Purchase Order approval application which is a Fiori (HTML5) application, with an easy to use interface, can be accessed either on mobile or PC. This app gives the approver a fraud score at a purchase order line item based on attributes such as the typical time period a purchase order is raised, the typical materials purchased and the amount and quantities purchased from a vendor. This score is generated using machine learning in the form of Self-Organizing Maps (SOMs) which are developed using SAP HANA Predictive Analysis Library (PAL).
Machine Learning behind the Fraud Score
The Machine Learning (ML) under the hood of the fraud score generated by the app uses SOMs which are a form of unsupervised neural network. This means that it is not necessary to have a previously set of transactions which have been identified as potentially fraudulent or not. Utilizing the transaction data, the SOM builds a model of the typical transaction profile carried out with a vendor taking in to account temporal attributes such as the date in which the purchase order was raised as well qualitative and quantitative attributes such as the quantity of goods/material to be purchased, the type of goods being purchased as well the value of goods being purchased, with the option of being able to add more attributes if needed. Then when a new purchase order is raised, a fraud score is generated for each item in the purchase order along with the flag whether the fraud score exceeds the threshold score for that vendor along with the contribution of each attribute to the fraud score.
Await details of the workings of the math behind SOM in the next post!!!
Given below are some screenshots from the dashboard and Purchase Order Approval App.