| Nowadays,there are more and more works applying deep reinforcement learning to the field of portfolio management.Part of it starts from the algorithm itself and improves the performance of portfolio management by improving the reinforcement learning algorithm.The other part is through the innovation of network among extract diffident stocks features to improve the performance of portfolio management.However,there are still two problems.First,the original stock price information contains a lot of noise,when the acquired stock data set is put into the model,the influence of noise is often not considered,which limits the expression of the deep reinforcement learning;Second,the current extraction of features between stocks especially the extraction of the temporal and spatial relationship between stocks may not be sufficient,and they are all aimed at the original financial sequence containing noise to extract,resulting in poor performance of the model in the face of certain market conditions.Based on this,the main works like this:Frist,in view of the problem of a large amount of noise in the original stock sequence data,this paper will use a signal processing algorithm,EWT(empirical wavelet transform)to denoise the stock sequence data.However,the traditional EWT algorithm has a relatively simple spectrum division method and cannot decompose the noise data well.Therefore,the EWT algorithm is improved and the IEWT algorithm is proposed.For the stock time series data,adaptive spectrum division will be carried out during spectrum division to ensure a small number of IMFs.At the same time,the IEWT algorithm decomposition is optimized by referring to the idea of EEMD algorithm.Second,use the IEWT algorithm to denoise the financial sequence,separate the noise according to a certain correlation coefficient threshold,and then use the denoised sequence to construct technical indicators representing different aspects of the stock,and input them into the model together with the separated noise,using different network structures for feature extraction,so that the model can learn high-level abstract representations of features and noise data at the same time.Third,build the IEWT-PG model,use the TCN network to extract the time series relationship between stocks for technical indicators constructed after denoising,and use the Multi-Head Attention Network to extract the spatial relationship between stocks.For noise data,use LSTM network to extract time series features,and use time attention network to strengthen feature expression,and finally input the obtained feature data to the fully connected layer,after the sigmoid function,use the BWSL strategy to obtain the weight of the final portfolio.The network parameters are updated using the strategy gradient method,and different reward functions,input data and model structures are used to explore the performance of the IEWT-PG model proposed in this paper.In the empirical analysis stage,experiments were carried out on markers representing three different levels of development,and DJIA constituent stocks,HSI constituent stocks and DAX constituent stocks were selected as portfolio datasets,and compared with traditional strategies and classic deep reinforcement learning model.Experimental results show that the IEWT-PG model proposed in this paper is superior to traditional strategies and classic deep reinforcement learning models in terms of return,risk and return-risk,which proves that the model proposed in this paper is effective in controlling risks and obtaining benefits. |
