As R.A. Fisher understood, causal discovery is feasible when controlled experiments on the system of interest can be conducted. Even in the absence of controlled human interventions, reality often provides natural experiments, offering observations of the same system under slightly shifted conditions. Rather than merely pooling such information, we propose an approach that minimizes future prediction error by optimizing a functional that accounts not only for pooled risk but also for risk differentials across settings. We demonstrate how this method extends beyond the linear regression framework into functional data analysis (FDA). This is joint work with Lucas Kania (Carnegie Mellon), Philip Kennerberg, Melania Lembo (USI) and Veronica Vinciotti (Trento).