Abstract Granger causality is an established, contentious method that seeks causal temporal connections via association and precedence. While not true causal inference, it assists in mapping networks of information flow that may warrant further study. Adaptation to nonlinear data remains challenging, and no single method is widely accepted like the vector autoregressive form for linear Granger causality. The universal approximator ability of artificial neural networks circumvents function specification, but many works pair these tools with in‐sample inferential procedures. These schemes fall short of conducting reliable inference. This article defines permuted Granger causality, a method that lends itself to out‐of‐sample testing, and advocates for its use as a simple fix to correct the imbalance in existing literature. The methodology is simple, shows promise in identification of causal connections in simulated and real‐world data, and vastly improves control for false positive Granger causal connections.