Detecting Fraud in Motor Insurance Claims Using XGBoost Algorithm with SMOTE

Abstract

Fraudulent claims in motor insurance policies continue to be a big menace to insurance companies. Fraudsters are devising new tactics of fabricating claims to make them appear valid. This makes insurance companies register huge losses in billions of money every year. The insurance policyholders bear these losses through increased premiums thus having negative social and economic ramifications. Numerous approaches have been proposed and applied in detecting and preventing fraudulent claims. The traditional approaches have become complex, time-consuming, and with low success ratio. To improve on fraud detection, the existing historical data can be used to train prediction models. To optimize the performance, this data require feature engineering to ensure only relevant features are used and handling of class imbalance. In this paper, we propose a model that is built on XGBoost algorithm. In data preparation, we propose to handle class imbalance by oversampling, using SMOTE. We aim at comparing the effect of class imbalance and oversampling on the performance of our model. The results obtained reveals that XGBoost performs well with SMOTE compared to imbalanced training dataset and also compared to other algorithms. Once the model is deployed, insurance companies will be able to detect and identify perpetrators of fraud and take necessary action. This will reduce their loss adjustment expenses and thus increase their profits.