Mean-CVaR Portfolio Optimization Based On Machine Learning

Since the Reform and Opening-up in 1978,China’s economy has entered the fast lane of growth and the stock market has come into being.People have more and more money and are no longer satisfied with keeping their funds in the bank,thus the demand for investment keeps rising.However,the asymmetry of information and the lack of professional knowledge make it difficult for individual investors to obtain the desired investment returns.Therefore,it is important to find a good investment method for individual investors to reduce the risk of people’s investment and increase their returns.In order to reduce investment risks and bring more returns to investors,we plan to combine machine learning algorithms and portfolio optimization to construct a quantitative investment strategy.We select the fundamental and technical indicators of 50 listed companies of SSE 50 index as input variables and the closing prices of individual stocks as output variables,and use a rolling window to forecast,dividing the data into 15 cycles,each cycle having a training set and a test set.LSTM,SVR and XGboost models with different cycles are constructed first.The best prediction model is found by comparing the prediction results of the three machine learning models.Then the returns are calculated from the predicted closing prices,and the 50 stocks are ranked according to the magnitude of the returns.Once the different predicted return rankings of the 50 stocks of the SSE 50 index by the three models are obtained,the top 5 to 10 stocks are taken for portfolio optimization.The Mean-CVaR model is used to find out the optimal portfolio proportions for the returns predicted by the three machine learning models respectively,while the Mean-variance model and the Equal-proportion model are introduced for comparative analysis.Finally,the back-tested annualized return,standard deviation,Sharpe ratio and cumulative return of the corresponding portfolio models are obtained by calculating the optimal investment ratio with the actual data.The following conclusions are obtained from the comparative experiments:First,the XGboost algorithm has better prediction effect on stock return prediction compared with SVR algorithm and LSTM algorithm,followed by LSTM model;Second,when the optimal number of portfolios is 5,the three indices of portfolio model yearly return,volatility and Sharpe ratio perform better overall in each model than n taking other values;Third,the Mean-CVaR models and Mean-variance models are more sensitive to the input variables than the Equal-proportion models,thus the stock selection of different algorithms has a greater impact on them;Fourth,the LSTM-Mean-CVaR model performs better than the other models,and the results illustrate that the stock selection of LSTM is more suitable for the Mean-CVaR model.