This paper introduces a method for causal inference using invariant prediction. The key idea is that causal models remain invariant under interventions, while non-causal models may not. By identifying models that are invariant across different experimental settings, we can infer causal relationships and construct confidence intervals for causal effects. The method is applicable to various scenarios, including observational and interventional data, and is robust to model misspecification. The approach is based on the assumption that the conditional distribution of the target variable given its direct causal predictors remains invariant across different experimental settings. This invariance is closely related to causality and has been discussed under terms like "autonomy" and "modularity." The method does not require knowledge of the intervention targets or the specific causal structure, making it flexible and applicable to a wide range of settings. The paper discusses the use of structural equation models and provides sufficient conditions for the identifiability of causal predictors. It also examines the robustness of the method under model misspecification and discusses possible extensions. The empirical properties of the method are studied using various datasets, including large-scale gene perturbation experiments. The method is designed for scenarios where data are collected from different experimental settings or regimes. It allows for the construction of confidence intervals for causal predictors and coefficients without prior knowledge of the causal ordering of variables. The method provides confidence intervals without relying on assumptions such as faithfulness or other identifiability assumptions. If a causal effect is not identifiable from the given data, the method automatically detects this and avoids making false causal discoveries. The paper also discusses the computational requirements of the method, noting that the complexity grows super-exponentially with the number of variables. However, practical implementations can be optimized by limiting the size of the set of causal predictors or by using adaptive methods to reduce the number of variables considered. Overall, the method offers a new approach for causal inference that leverages the invariance of causal predictions under different experimental settings, providing valid confidence intervals and robustness to model misspecification.This paper introduces a method for causal inference using invariant prediction. The key idea is that causal models remain invariant under interventions, while non-causal models may not. By identifying models that are invariant across different experimental settings, we can infer causal relationships and construct confidence intervals for causal effects. The method is applicable to various scenarios, including observational and interventional data, and is robust to model misspecification. The approach is based on the assumption that the conditional distribution of the target variable given its direct causal predictors remains invariant across different experimental settings. This invariance is closely related to causality and has been discussed under terms like "autonomy" and "modularity." The method does not require knowledge of the intervention targets or the specific causal structure, making it flexible and applicable to a wide range of settings. The paper discusses the use of structural equation models and provides sufficient conditions for the identifiability of causal predictors. It also examines the robustness of the method under model misspecification and discusses possible extensions. The empirical properties of the method are studied using various datasets, including large-scale gene perturbation experiments. The method is designed for scenarios where data are collected from different experimental settings or regimes. It allows for the construction of confidence intervals for causal predictors and coefficients without prior knowledge of the causal ordering of variables. The method provides confidence intervals without relying on assumptions such as faithfulness or other identifiability assumptions. If a causal effect is not identifiable from the given data, the method automatically detects this and avoids making false causal discoveries. The paper also discusses the computational requirements of the method, noting that the complexity grows super-exponentially with the number of variables. However, practical implementations can be optimized by limiting the size of the set of causal predictors or by using adaptive methods to reduce the number of variables considered. Overall, the method offers a new approach for causal inference that leverages the invariance of causal predictions under different experimental settings, providing valid confidence intervals and robustness to model misspecification.