Multi-objective Portfolio Selection Based on Machine Learning

Firstly, random forest, RBF neural network and BP neural network are used to predict the closing price of stocks in this paper. Historical data and predicted closing prices are used to calculate the mean, downside semi-variance, and skewness of investment portfolio returns. Secondly, considering transaction costs, upper and lower bound, and borrowing and lending constraints, a multi-objective portfolio selection model(M-SV-S) based on machine learning is proposed. The optimization problem corresponding to this model belongs to non-convex optimization problems and is difficult to solve so that it is transformed into a single objective optimization model, which is solved by differential evolution algorithm. Finally, the component stocks of the SSE 50 Index is chosen as samples for empirical analysis. The investment performance of the M-SV-S is compared with the equal weight portfolio mo-del in terms of returns and Sortino ratio. The empirical results show that the daily net return of M-SV-S model between 1% and 4%, the cumulative abnormal return of 30 days of M-SV-S over 50%, and the Sortino ratio of M-SV-S greater than 0 within the out of sample window. It means that the investment performance of M-SV-S is significantly better than that of the equal weight portfolio model.