The traditional risk management approaches of capital markets have been challenged by the lack of high-dimensional data processing capacity, nonlinear relationship capturing and dynamic risk factor identification, which makes the portfolios weak during extreme market situations. To overcome these problems, this paper will develop a risk management and optimization system built on machine learning. To extract dynamic market features, first a Long Short-Term Memory (LSTM) network to model stock price time series is employed. Second, a Random Forest (RF) model is used to discover the important risk factors that influence asset returns and top 20 most significant indicators by importance of the features are chosen. This is followed by the creation of a risk warning model that uses a Support Vector Machine (SVM), with the risk threshold of the 95 th percentile of the Value at risk (VaR). Lastly, a Genetic Algorithm (GA) is to optimize the weight allocation of the portfolios, and the objective is multi-objective maximization of Sharpe ratio. Empirical findings depict that the model predicts the direction of returns at least 5 days ahead with an accuracy of 76.3%, the risk type with an accuracy of 91.4%, and the annualised return at 18.7% hence demonstrating that the approach has substantial impacts in risk management and returns optimization.
Siyu Tong (Thu,) studied this question.