Machine Learning for Fraud Detection in Insurance Claims using Time-Series Anomaly Detection
Rahul Nimbalkar · International Journal of Emerging Research in Engineering and Technology · 2024
Insurance fraud is a significant challenge that continues to grow before the eyes of the insurers, causing disruptions in the scale of billions of dollars a year. Rule-based traditional systems are unable to detect customized evolving and rather complex fraudulent patterns, especially with a massive dataset. This paper aims at giving attention to the machine learning (ML) approach to investigating fraud detection in insurance claims from the viewpoint of time-series anomalies. By operating on the temporal dimension of claim filing, time-series ML models build the intervening link that leads from normal claim behavior to temporary extensions in deviation, which mark the spotlight of improbability_cont_ . We have used real and synthetic datasets to compare algorithms for anomaly detection in various criteria: accuracy, precision, recall, and AUC_ROC. Results suggest that time-series-based models are uniquely suited to detecting dynamic and complex forms of fraud, especially in the form of late reporting, charge inflation, and claim bursts, whereas traditional classifiers fail. We also look at how unsupervised learning can be a serious contender in instances where the label data is scant or non-existent. With respect to the existing literature, this work proposes an end-to-end framework to integrate the ML anomaly detection into workflows for insurance fraud detection. The work ends with a discussion of interpretability issues, deployment challenges, and ethical considerations for algorithmic decision-making in financial services