[D] A variable with Pearson r ≈ 0.12 turned out to be useless for prediction — how do you formally distinguish state variables from predictive signals? # Methodological Issue Inquiry I ran into a methodological issue I'd love input on. I had a variable showing Pearson correlation of roughly 0.12 with my outcome variable modest but consistent across the sample. Based on that alone, it looked like a potentially useful predictor. The problem appeared when I introduced a one-step time delay: using the value at t-1 to predict at t, the relationship essentially disappeared. The correlation was contemporaneous , it described the current state of the system well, but carried no forward-looking information once you respected the temporal ordering of the data. This got me thinking about a distinction I'm not sure how to formalize: the difference between a variable that's correlated with the current state of a system versus one that's genuinely predictive of future state transitions.…