Research On Digital Currency Portfolio Strategy

The portfolio problem is to continuously redistribute funds to different financial products.It minimizes the risk of investment on the premise of guaranteeing the expected return,or maximizes the return of investment on the premise of controlling the risk.Traditional investment portfolio ideas focus on risk dispersion,while modern ideas pay attention to the selection of the optimal investment proportion or the optimal portfolio size.Aiming at the optimal investment proportion of portfolio in digital money market,this paper proposes a solution based on deep reinforcement learning framework.The framework mainly includes digital currency historical data preprocessing,deep neural network of output portfolio strategy,experience pool of preserving digital currency historical data and distribution strategy,incentive mechanism of reinforcement learning and online random batch learning scheme.The main work of this paper is as follows:1.In order to filter out the data attributes with high correlation,this paper first ranks the data attributes of digital currency by XGBoost algorithm,and then sorts out three currency attributes with the highest importance ranking.On one hand,it can remove the attributes with weak correlation.On the other hand,it can reduce the computation and improve the training effect.At the same time,the input data is constructed into three-dimensional input data according to currency attribute,currency quantity and historical data period.2.At present,deep learning often tries to predict price changes when dealing with financial problems.Trading agents act on the forecast results,but it is difficult to predict future market prices.In this paper,the deep reinforcement learning algorithm is used to make the decision of trading action directly.Convolutional Neural Network(CNN),Convolutional Long Short-Term Memory Network(ConvLSTM)and improved Depthwise Separable Convolution are used as the output network of portfolio strategy respectively.Through the Policy-based reinforcement learning method,the digital money market is fully explored and studied,and a reasonable portfolio strategy is given.Compared with traditional portfolio strategy methods,in the comparison of the test results of the same currencies in different periods and different currencies in the same period,several network frameworks based on deep reinforcement learning methods achieve higher final asset value.Moreover,the improved depthwise separable convolution network can achieve a good effect,while also effectively reducing the computational parameters to speed up the training of the network.3.On the basis of the above research,in order to further improve the investment returns of the model,a separable gating network is proposed.The network can give different weights to the data of each dimension of the three-dimensional feature map output by convolution network,and fuse the weights of each dimension,so as to enhance and suppress the feature data better.The network is integrated into CNN,ConvLSTM and the improved depthwise separable convolution network,and the experimental results are further improved.