Financial Trading Strategy System Based on Machine Learning

The long-term and short-term volatilities of financial market, combined with the complex influence of linear and nonlinear information, make the prediction of stock price extremely difficult. This paper breaks away from the traditional research framework of increasing the number of explanatory variables to improve the explanatory ability of multifactor model and provides a new financial trading strategy system by introducing Light Gradient Boosting Machine (LightGBM) algorithm into stock price prediction and by constructing the minimum variance portfolio of mean-variance model with Conditional Value at Risk (CVaR) constraint. The new system can capture the nonlinear relationship between pricing factors without specific distributions. The system uses Exclusive Feature Bundling to solve the problem of sparse high-dimensional feature matrix in financial data, so as to improve the ability of predicting stock price, and it can also intuitively screen variables with high impact through the factor importance score. Furthermore, the risk assessment based on CVaR in the system is more sufficient and consistent than the traditional portfolio theory. The experiments on China’s stock market from 2008 to 2018 show that the trading strategy system provides a strong logical basis and practical effect for China’s financial market decision.