Longitudinal Dense-to-Sparse Forecasting: Individual Variability Predicts Conversion Better than Mean Change

Clinical forecasters trained on dense imaging trajectories are often re-purposed to rank sparse downstream events. What survives this hand-off under cohort shift is rarely tested. From a cortical-thickness forecaster trained on one Alzheimer's cohort, we compare two label-free risk signals: predicted mean structural change, and predicted individual variability. In-domain, variability ranks eventual converters about as well as a fully-supervised clinical baseline. Under transfer, variability remains predictive while mean change degrades. The split is not the usual aleatoric-vs-epistemic: variability from input-residual structure transfers; from posterior weight perturbations, it does not. A controlled covariate-shift experiment reproduces the pattern. Forecasting error alone does not predict which uncertainty channel transfers.