Regressions aimed at detecting forecast errors predictability are a widespread tool to assess deviations from the full information rational expectations equilibrium benchmark. We show that interpreting these regression coefficients as evidence of over- or under-reaction may be misleading when the object of interest is an endogenous variable. We simulate scenarios where an econometrician would detect short-term under-reaction and long-term over-reaction, when in reality the equilibrium outcome exhibits over-reaction at all horizons when compared to the rational expectations benchmark. We then turn to stock market expectations data and compare regression results on earnings and dividends (exogenous variables) with those on price targets (endogenous variable). Regressions of forecast errors predictability are still instructive of the precise biases agents have, but only when they are interpreted through the lens of the underlying structural model.