Deep portfolio selection with contrastively aligned cross-modal attention

Open Webpage

Yupeng Fang, Huichou Huang, Ruirui Liu, Chaoyu Chen, Haoxian Liu, Qingyao Wu

Published: 01 Jan 2026, Last Modified: 12 Oct 2025Pattern Recognit. 2026EveryoneRevisionsBibTeXCC BY-SA 4.0
Abstract: Highlights•Inner-Modal Contrastive Learning improves encoders by capturing fine-grained features in financial data.•Inter-Modal Contrastive Fusion aligns numeric and text modalities in a unified spatio-temporal feature space.•Multi-modal relations incorporate industry categories to guide predictive dependencies across modalities.•Portfolio optimization uses Constant Relative Risk Aversion(CRRA) utility loss to learn predictive statistical features from joint return distribution.