Insurance Fraud Detection Models: A comparative study

Published by Yves Yameni on

This post is an overview of Artificial Intelligence models used in insurance fraud detection.

       Insurance fraud refers to an illegal act or procedure on the part of either the policyholder (buyer) or the insurer (seller) of an insurance contract.

– Insurance frauds from the insurer include but not limited failing to submit premiums, selling policies from non-existent companies, churning policies to create more commissions.

– Policyholder frauds on the other hand consist of falsified medical history, exaggerated claims, post-dated policies, viatical fraud, faked death or kidnapping, and murder, faked accident or car thievery among other.

-Governments often fight frauds committed by insurers, whereas insurers must often defend themselves from policyholders’ frauds.

  1. Some key data on insurance fraud

In Europe: In 2019 report of Insurance Europe , one could read the following sentence: ’Fraud is a significant problem all over Europe. The level varies between countries due to various factors including, but not limited to, the size of the market; the type of products available; the degree of investment in counter-fraud systems and checks; the legislative and regulatory framework; and prevailing cultural attitudes.

According to Insurance Europe’s members, insurance fraud in the EU stood at approximately e13 billion in 2017.

In America (USA): The insurance industry is made of more than 7,000 companies collecting over $1 trillion in premiums per year. The size of the industry is too large, this significantly contributes to insurance fraud by providing more opportunities and bigger incentives for committing illegal activities.

 Based on FBI , the total cost of insurance fraud (non-health insurance) is estimated to be more than $ 40 billion per year . That means Insurance Fraud costs the average U.S. family between $ 400 and $ 700 per year in the form of increased premiums.

2. Machine Learning Models

Fighting insurance frauds has long be done through investigations by experts. It is a tedious and time-consuming task to investigate fraud manually. When the number of insurance claims increases, more human resources is required for investigation. Automated tools are in this case very useful in the sense that even if they might not be able to detect frauds with a 100% accuracy, at least they can safely guide investigators in their job.

       Many vendors offer fraud detection solution, either as a standalone solution or a part of their software packages. Few of them are listed as Intelligent Insurance Fraud solution ; SAS® Detection and Investigation ; aws fraud detector

            Above mentioned solutions are all built upon Machine Learning algorithm, and their confidence level depends on their performance. Accuracy, recall, precision, false negative, false positive, ROC-curve are metrics commonly used to evaluate performance of a Machine Learning model.

            Below is an architecture of a fully automated Machine Learning solution for fraud detection.

Figure 1. Automated ML solution for fraud detection.

The section highlighted in red is where machine learning model is trained and evaluated, that section is our focus in this post.

Figure 2. training section

For training purpose, we used a data set of 33 features and 11070 observations. Feature engineering leaded to deletion of few variables that have very small or no impact on the model. Still in pre-processing, One Hot Encoding was used to transform categorical features; Synthetic Minority Over-sampling TEchnique (SMOTE) was used to handle imbalance in data. The data set was split into two parts 80% and 10% representing training and test sets respectively.

       Here are the machine learning models used in the study followed by their performance evaluation using ROC-curve.

  • Extreme Gradient Boosting (Xgboost) Model
  • Generalized Linear Model (GLMnet), Logistic
  • Multi-Layer Perceptron (MLP)
  • Na¨ıve Bayes
  • Support vectors Machine (SVM)
  • Tuning Xgboost model with GridSearch, XgGSCV
  • Hybrid/ensemble models
  • Stacking Classifier
  • Voting Classifier

Details about each model is not given in this study, but can be found in dedicated references (books, articles, software’s websites and documentations). In our case, scikit-learn python library was used.

Figure 3. ROC-curves

Analysis of the left-hand side ROC-curve shows that Xgboost is the best. The right-hand side ROC-curve also shows that Xgboost performance is better than those of voting and stacking. However, staking performs very closely as Xgboost.

            One can conclude based on the outcome of this study that Xgboost model is recommended for building automated fraud detection tools.