Unsupervised Learning Algorithms Help Identify Policies Worth $9MM in Premium with Likely Sales Frauds

Case Study

Unsupervised Learning Algorithms Help Identify Policies Worth $9MM in Premium with Likely Sales Frauds

Business Objective

Our client is a Fortune 500 company, providing life and disability insurance to more than 100 million customers across 50 countries.

The client wanted to improve sales practice fraud detection using a predictive analytics framework. The current process is limited to the compliance team reviewing the policies and agents based on official complaints. This reactionary approach resulted in higher losses and increased investigative labor to address the occurrence. Our client wanted to proactively identify fraud, thus enabling the compliance team to focus on identified outlier cases.

– Develop a predictive framework to flag possible sales fraud practices and detect agent and policy-level anomalies

– Build an interactive dashboard to speed up identification, evaluation, and interpretation of the drivers of anomalies for the Special Investigative Unit (SIU)

Challenges

  • Lack of labeled fraud data – No recorded historical frauds known to take the conventional route of building a predictive model using supervised learning methods
  • Inconsistent data – In-silo data sources with limited information, no data dictionaries, and minimal business context

Solution Methodology 

  • Considered potential fraud scenarios such as commission abuse / fraudulent business, misappropriation of premium, employee theft, etc., after discussing with SIU and business teams.
  • Built hypotheses that needed to be tested working closely with the sales, investigation, and local & global compliance teams
  • Identified key data sources such as customer data, agent data, policy details, commission information, premium history, sales transactions, etc., to help validate the identified hypotheses.
  • Developed unsupervised anomaly detection models (by exploring various techniques including One Class SVM, Local Outlier Factor, IForest, K Nearest Neighbor) coupled with a rule-based statistical model to identify Low, Medium and High-Risk policies and agents
  • Built an interactive Tableau dashboard for investigators to evaluate high-risk agents / policies

Business Impact

  • Trends related to sales, amount of policies sold, early lapse behavior, and increased presence of new customers helped investigators flag potential fraud cases worth $9MM premium
  • The new process eliminated the need for manual scrutiny of all policies and agents, limiting the investigation to <5% of the total population
Download Case Study

©2019 Tiger Analytics. All rights reserved.

Log in with your credentials

Forgot your details?