Robust prediction methods that use causal structure typically rely on heterogeneity across observed environments to learn invariant predictors and to define the perturbations they guard against. This paper shows that, under a structural additive-noise model for the response, both can be recovered from a single observational distribution. The key insight is that the causal prediction function has constant conditional mean squared error given its prediction, an invariance that can be exploited without multiple environments.
The authors generate synthetic environments by reweighting the covariate distribution along the prediction, then use a Cramér–von Mises criterion to penalize departures from this invariance. The resulting penalised risk admits a regularised worst-case representation, linking the criterion to robustness against covariate shifts induced by reweighting observations according to their predicted values. Under an additional orthogonality condition, the causal prediction function is identified as the unique risk minimiser subject to the invariance restriction.
The paper establishes consistency with an explicit rate, derives a goodness-of-fit test for the invariance restriction, and characterises the population target under misspecification. When multiple observed environments are available, the construction applies environment by environment and can be combined with Distributional Robustness via Invariant Gradients (DRIG) to control two complementary classes of shifts.