[1O3-GS-10] AI application:

[1O3-GS-10-02] Portfolio Optimization Using Co-Attention: Capturing Temporal Patterns and Inter-Asset Relationships

〇Mihiro Otaki¹, Yoshimasa Tsuruoka¹ (1. Univ. of Tokyo)

Keywords:portfolio optimization, reinforcement learning, Co-Attention, finance

Portfolio optimization is a fundamental challenge in finance, requiring the extraction of meaningful features from asset price data to achieve an optimal balance between risk and return. Previous methods struggle with limitations such as dimensional constraints stemming from the sequential processing of temporal patterns and inter-asset correlations, as well as reliance on numerical prediction models or manually designed statistical indicators. In this study, we propose a novel model that leverages Co-Attention to independently process temporal patterns and inter-asset relationships and integrate them effectively. This model captures multi-dimensional features of asset price data and enhances adaptability to non-stationary market environments by learning temporal and asset-level features in parallel. Experimental results demonstrate that the proposed approach outperforms classical methods and achieves performance comparable to state-of-the-art attention-based techniques. These findings suggest new possibilities for portfolio optimization in financial markets and contribute to the development of more effective asset management strategies.