Research On Classical Portfolio Model Improved By Deep Learning

In recent years,the research on portfolio has increasingly become a hot topic.For the unpredictable capital market,it is a problem that needs to be deeply considered to find a way to obtain both higher returns and lower risks.In the early stage,investors invested more through some historical past experience.Such investment depends more on intuition,and may rely too much on the rise and fall of an asset,resulting in high investment risk.After introducing the mean variance model,some scholars consider the past yield data and other information into the subsequent investment process to determine the investment ratio and other information.Later,some scholars improved it and got a more classic Black Litterman portfolio model.However,when the Black Litterman model was first proposed,its income expectation and expectation of future investment uncertainty were subjective choices,so it is inevitable to fall into the trap of "empiricism" to a certain extent.Therefore,a considerable number of scholars try to improve the BL model from the expectation of future income and the expectation of future investment uncertainty,which has also achieved good results,and basically overcome the trap of "empiricism" in the essence of the method.However,in the past,traditional forecasting methods often only deal with linear data relations.When encountering complex nonlinear data such as stocks,it will be difficult to have better performance,which also directly affects the results and performance of the final portfolio model.In recent years,with the improvement of computer computing power,neural network began to heat up gradually.Its excellent performance in nonlinear data and different adaptability to different data also make its prediction on stock,futures and other data become very powerful,which can obtain more accurate prediction ability than the traditional time series model.However,in the prediction process,the number of prediction information dimensions can also determine the prediction results to a certain extent.Putting stock data information into the prediction model as much as possible can also improve the final prediction accuracy and results to a certain extent.However,too many data information dimensions will bring a certain degree of information redundancy,and will slow down the operation speed of the prediction model.In contrast to the black Litterman model,its portfolio in the same asset performs well,but it is difficult to reflect the differences and relationships between different types of assets.In contrast,there will be a large gap in the laws of historical return between different assets,and few scholars have studied this aspect in previous papers.In view of the above problems and defects,this paper puts forward some improvements and innovations,as shown below.(1)Principal component analysis is used to reduce the dimension of information dimensionIn this paper,more than 20 relevant stock information indexes are involved,which will inevitably produce a certain amount of information duplication.In this paper,the principal component analysis method is used to delete the redundant information and maintain the amount of information contained in the original data as the input data of the subsequent prediction model.(2)The long-term and short-term memory model is used to predict the yieldThe historical rate of return of assets involved in this paper has an obvious nonlinear trend,while most scholars used the traditional time series model to predict.However,because the traditional time series model not only has certain requirements for data,but also mainly uses linear model at the bottom of the model,it is difficult to effectively predict the historical benefits of nonlinear changes.Using the long-term and short-term memory model,including the forgetting gate,can effectively select the impact of the historical rate of return on the subsequent rate of return.The closer the data is in time,the greater the impact on the current rate of return.The input gate and output gate contain nonlinear activation function and hidden layer,so that the model can better grasp the nonlinear change and hidden relationship of asset return,and achieve better results than the traditional time series model in the prediction of return.(3)The convolution neural network is used to adjust the investment results of BL on different assetsAfter the Black Litterman model is applied to different assets for portfolio selection,different assets are constructed into corresponding abstract assets according to the results of the portfolio.In this way,the income matrix results of various assets are constructed and brought into the convolution neural network for training.In the process of data training for thousands of days,Convolutional neural network can continuously learn the abstract information contained in the data,so as to have a further in-depth understanding of the subsequent portfolio,so as to obtain a more accurate portfolio proportion,so as to further correct the proportion of asset allocation,so as to obtain higher and more stable income in the investment process.Using the powerful feature extraction performance of convolutional neural network is conducive to further extract the performance of the obtained asset allocation data in the training process,so as to further improve the income on the basis of ensuring the risk.In a nutshell,this paper uses the popular machine learning related methods to make some improvements in the dimensionality reduction of early data information,the selection of prediction methods and the update of the final portfolio proportion,and proposes a bl portfolio model under the improvement of machine learning,in order to obtain higher and more stable returns in the complex capital market.