Distributional robustness of K-class estimators and the PULSE

Martin Emil Jakobsen, Jonas Peters

While causal models are robust in that they are prediction optimal under arbitrarily strong interventions, they may not be optimal when the interventions are bounded. We prove that…