Long And Short Term Risk Control For Online Portfolio Selection

Online portfolio selection is to allocate capital among a group of assets to maximize cumulative returns.Most of online portfolio selection algorithms focus on maximizing returns without effectively controlling risk.Further,many risk control algorithms use the maximum drawdown,the Sharpe ratio,and others as risk indicators.However,these risk indicators are not sensitive to the short-term risk.In view of the above problems,this paper studies the risk control of online portfolio as follows:1)This paper proposes the Long and Short Term Risk(LSTR)control algorithm for online portfolio selection.After obtaining the proportion of funds allocated for each trading day,the algorithm adjusts the proportion of risk-free assets through the combined effect of two parameters to achieve high return and low risk to the investment portfolio.The first parameter learns the long term risk of the market.The second parameter reflects the short term risk of the market and responds quickly to asset adjustments.Through the joint action of the two parameters,the risk control ability of the algorithm is effectively improved.2)According to Roy's Safety-First Criterion,a random variable of risk indicator is designed.Long-term risk control parameters learn the long-term risk parameters by counting the total number of times that a random variable of the risk indicator appears and using an inversion strategy.The short-term risk control parameter counts the number of consecutive occurrences of the risk indicator random variable,which changes rapidly in the short-term to ensure a rapid response to the short-term market.3)Reinforcement Learning Deterministic Strategy Gradient(DDPG)is an algorithm of the Actor-Critic framework obtained by transforming the DPG method using the idea of the DQN extended Q learning algorithm.We use the LSTR algorithm to modify the DDPG algorithm.The good performance of the experimental results illustrates the rationality of the change.In order to verify the effectiveness of the LSTR algorithm,this paper conducted a comparative experiment of the algorithm.The six datasets used in the experiment frequently appear in various online portfolio papers.These datasets come from different countries and regions,which not only have a large time span,but also have diverse market trends.The experimental results show that the performance of the LSTR algorithm is not only better than online portfolio selection algorithms with risk control,but also better than existing online portfolio selection algorithms without risk control.