Insurance companies depend on accurate, high-quality data to operate effectively, whether for pricing policies, processing claims, or managing risk. However, the large volume and complexity of insurance data make it susceptible to anomalies. These anomalies may stem from human error, system integration issues, unusual customer behavior, or deliberate fraud. Detecting such irregularities is crucial for safeguarding operational integrity, ensuring compliance, and maintaining customer trust.

Traditional anomaly detection in insurance has often relied on manual review or rule-based systems. While these approaches benefit from expert knowledge and are transparent, they struggle to scale with growing datasets and cannot easily adapt to new or complex anomaly patterns. Recent advances in machine learning, particularly unsupervised methods, offer promising alternatives by learning the underlying structure of data and identifying deviations without requiring extensive labeled examples.

This paper investigates the application of ensemble-based unsupervised learning models for anomaly detection in insurance data, focusing on autoencoder and variational autoencoder ensembles. By training multiple models with different architectures and combining their outputs, the aim is to improve robustness, scalability, and detection accuracy.

Key discussion points:

• Business rationale: Areas in the insurance industry where unsupervised anomaly detection adds value.
• Benchmarking: Performance, strengths, and limitations of unsupervised methods in a controlled environment.
• Robust decision making within autoencoder ensembles: Criteria for classifying anomalies using 10 autoencoder models featuring different architectures.
• Data migration: Before transferring complete datasets from legacy systems to modern platforms, how insurers can employ automated outlier detection methods prior to migration.
• Fraud detection: Evaluation of our models’ performance on realworld data.