A hybrid deep learning architecture for cross sectional stock ranking. Causal temporal convolutions, overlapping local self attention, and GRU aggregation with Monte Carlo Dropout uncertainty estimation.