Machine learning approaches to outlier detection

A brute-force nested simulations approach to capture risk interdependencies is infeasible with current resources available to insurers. It would require 1,000 or more risk-neutral simulations to produce a single stochastic scenario of the full risk distribution. This paper delves into finding an automated way to identify and remove outliers. We discuss the following: 

• A simple outlier deletion technique and its limitations
• A powerful method known as Cook’s distance
• A few alternative machine learning approaches
• Conclusions and areas for further research