What is distribution shift robustness evaluation using worst-case risk?

Worst-case risk evaluates model performance under adversarial or shifted distributions.

Instead of evaluating average performance, robust evaluation considers the worst-case loss over a set of plausible distributions. This is central in Distributionally Robust Optimization (DRO). It ensures that even under adversarial shifts, the model maintains acceptable performance. It is particularly important in high-stakes domains like finance and healthcare.