A Divide-and-Conquer Framework for Attention-based Combination of Multiple Investment Strategies

Abstract

In order to maximize the profit, investors usually examine diverse investment strategies based on various information when constructing their portfolios. However, they can hardly always construct a profitable portfolio due to the dynamic performance of the strategies and the dynamics of the market state along with the time. To address this challenge, we propose a 2D-attention framework to capture the dynamics of the above two factors in this paper. To capture the dynamic of the first factor, we design a strategy-wise attention model to dynamically combine multiple strategies according to their respective effectiveness. To deal with the second factor, we design a divide-and-conquer framework to learn multiple strategy-wise attention models, which categorizes the whole market periods into a few stable states and jointly learn respective models for each state and then build a state-wise attention model to combine them dynamically for the final task. Extensive experiments on real-world data demonstrate that our 2D-attention framework can significantly outperform several widely-used baselines.