Risk adjustment is an important tool in health insurance markets that aims to redistribute payments based on enrollee health. In practice, risk adjustment formulas underestimate the costs of some groups of enrollees, making these groups unprofitable for insurers. This leaves groups vulnerable to discrimination in markets where insurers have the ability to redesign plan benefits, such as provider networks or drug formulary tiers. Changes in plan design targeted at the group level make the associated insurance plan less attractive to unprofitable enrollees. In this research, we explore tree-based methods for automatically detecting previously unidentified groups underpredicted by the risk adjustment formula as well as methods for reducing this underestimation. These techniques improve on the arbitrary nature of existing evaluations of risk adjustment performance on groups and may ultimately lead to deployable methods that can prevent group-level discrimination in insurance markets.