基于多元函数型聚类和深度强化学习的最优资产配置策略OA

This paper proposes a novel optimal asset allocation method by comprehensively using multivariate functional clustering,deep learning,and deep reinforcement learning techniques.Firstly,the multivariate functional clustering method is used to cluster the available stock assets in the trading pool.Secondly,the LSTM-Non-stationary Transformer model with time series feature extraction capabilities is applied to predict the return of each stock.Next,based on the return prediction information,high-quality assets are selected from each clustered stock pool.Finally,dynamic optimal asset allocation models are constructed using three deep reinforcement learning methods:advantage actor-critic,proximal policy optimization,and soft actor-critic,based on the representative assets selected from each stock pool and their prediction information.This asset allocation strategy is implemented using the constituent stocks of the Shanghai Stock Exchange 50 Index and the Nasdaq 100 Index.The empirical results show that asset screening through functional clustering can effectively reduce the impact of non-systematic risk on asset allocation performance.By incorporating return prediction information into the deep reinforcement learning models,the profitability of asset allocation can be significantly improved,and robust,sustainable returns can be achieved.Furthermore,compared with the maximum Sharpe ratio model,the equal-weight model,and the market index,the proposed model demonstrates superior investment performance.