AI-Supported Anomaly Detection in Insurance

This technical report investigates the application of AI-supported anomaly detection methods in the insurance sector, focusing on ensemble-based unsupervised learning models, particularly autoencoder and variational autoencoder ensembles. The report outlines the significance of accurate data in insurance operations, emphasizing the challenges posed by anomalies due to human error, system integration issues, and fraud. Traditional methods of anomaly detection, which often rely on manual reviews and rule-based systems, are discussed, highlighting their limitations in scalability and adaptability. The study is structured into two main parts: a foundational analysis benchmarking various classical and deep learning algorithms on proxy datasets, and a pilot project applying these methods to real insurance data. The findings suggest that ensemble-based deep learning methods can enhance fraud detection processes and improve decision-making in insurance operations. Future directions include exploring additional application areas and integrating advanced techniques for more complex anomaly detection.