Addressing data imbalance in insurance fraud prediction using sampling techniques and robust losses

Abstract

Fraud in the auto insurance industry is a significant challenge for insurers glob- ally. Policyholders engage in fraudulent activities like falsifying documents and creating fake evidence, leading to substantial financial losses for insurance companies. Traditionally, insurers have relied on financial examinations and machine learning models to detect fraud, which require extensive data processing. This study uses deep learning models, particularly convolutional neural networks, to tackle fraud detection. Using a dataset of 1,000 car collision claims from seven US states in 2015, we aim to demonstrate the effectiveness of deep learning, even with small datasets. Despite the limitations of small datasets, our deep learning models achieved performance comparable to traditional methods through rigorous efforts. This research supports sustainable development by promoting innovation and technological advancement (SDG 9) and contributes to fraud prevention, organizational integrity, and transparency in the insurance industry (SDG 16). Future work will focus on leveraging other advanced deep-learning techniques to enhance fraud detection further