Predicting excess return in dividend models under heterogeneous trading and learning: a diffusive approach