Intelligent Quantitative Investment Based On Multi-Scale Self-Attention Mechanism

With the rapid development of computer technology and the arrival of big data era,quantitative investment as an investment method has received more and more attention.Also,more researchers use machine learning and deep learning methods to explore the information of financial market,and begin to realize intelligent quantitative investment.However,the current research work has not been able to fully capture the multi-scale features and the correlation between time series features from model.Based on the background above,this paper proposes a multi-scale self attention mechanism to solve the above problems.We take commodity futures and digital currency as the research object,and verify the effectiveness of the multi-scale self attention mechanism from three issues: high-frequency time series prediction,trading strategy and portfolio optimization.The main contributions of this paper are listed as follows:1.For high frequency time series prediction,most of the previous work in highfrequency trading is based on the combination of feature-based engineering and non deep learning models,and the work based on deep learning methods can not capture the multi-scale information in time series and the mutual importance of different features.In this paper,we propose a model called Multi-Scale Self-Attenion Convolutional Network(MSSACN)to predict the price change in the future.The model uses one-dimensional Inception structure to capture multi-scale time series features,and uses self-attention structure to capture the mutual importance of time series.Compared with other published works,MSSACN has the highest prediction accuracy in three futures markets which are the largest trading volume futures in Dalian futures exchange.2.For trading strategy,in order to solve the problem of transaction cost in actual transaction and make better use of long-term high-frequency information,we propose atrade model named Multi-Scale Self-Attention based Recurrent Reinforcement Learning(MSSA-RRL).In this model,MSSA extracts features from high frequency data and reduces dimensions,and RRL is used to output trade decision.The experiment results show that MSSA-RRL has achieved the best results in many indicators and different assets.In addition,MSSA has better performance than other deep learning models in feature extraction,which further verifies the effectiveness of the proposed multi-scale self attention mechanism.3.For portfolio optimization,the previous work in the portfolio management problem has failed to capture the multi-scale features in financial time series and the interaction between different assets.In this paper,we use the multi-scale self-attention structure which is mentioned above and propose an investment model called MultiScale Self-Attention based Portfolio Management(MSSA-PM).The model can learn the information of each asset on different time scales and the interaction of different assets.At the same time,the reinforcement learning mudule is used for optimization and decision-making.The experimental results show that MSSA-PM has achieved better results than those of the same type method in several periods of digital currency market.