Online Portfolios Selection Strategy Based On LSTM

In this article,we builds an LSTM model first to predict the stock price.we just use one hidden layer in the model,adjusting the input layer's,output layer's and hidden layer's parameters to achieve the best prediction result.Also,we try different hyperparameters: use the Adam algorithm to optimize the model,try the unreasonable learning rate,iteration's number,batch size,etc.Also,we use the Mean Square Error(MSE)to evaluate the prediction performance.The experiments show that the accuracy of prediction is high.In the second part,the LSTM stock price prediction model will be combined with two OLPS strategies: A peak price tracking-based learning system(PPT)and Radial basis functions with adaptive input and composite trend representation(AICTR),and form two new strategies,PPT-LSTM and AICTR-LSTM.PPT-LSTM changed the method of using the price peak value of each stock's historical time window to represent the future price in the original strategy,and replaced it with the results of the LSTM stock price prediction model,while AICTR-LSTM added LSTM trend into the original strategy.Experiments show that PPT-LSTM and AICTR-LSTM have a certain improvement over the original strategy on most data sets,but the performance of PPT-LSTM is not as expected,and AICTR-LSTM has a significant improvement effect compared to the original AICTR strategy.It shows that simply using the LSTM forecast results to determine the next period of investment portfolio can't produce the desired effect,and the compound trend strategy AICTR-LSTM can better play the role of LSTM forecast trends and generate more cumulative income.