Research On Quantitative Strategy Based On Hierarchical Deep Reinforcement Learning

Traditional quantitative trading strategies are widely used in stock,futures and other financial markets,but the manual feature extraction method lacks the capability of dynamic adjustment.Hierarchical deep reinforcement learning can effectively simulate the complex market environment and solve the problem of dynamic quantitative trading.We study the quantitative trading strategy based on h deep reinforcement learning.The main work is as follows:Based on the asynchrony of trend judgment and quantitative timing decisionmaking in practical trading and the dependence on data correlation,this work analyzes the shortcomings of the existing reinforcement learning model,innovatively applies the hierarchical reinforcement learning method to quantitative transactions,and creates a hierarchical deep learning intelligent trading agent H-Trader based on Semi MDPs.Considering the profit reinvestment in the actual strategy,the reward function Reward of compound method is used to replace the simple sum profit expression,and the double exponential discount factor is introduced into the calculation process;Fuzzy representation technology is used to learn the features of long-term trend,and FCM algorithm is used to assign weights to fuzzy feature representation;The multilayer layered LSTM structure of exponential expansion stack is used to realize the processing of memory in coarse particle state-size for Meta-controller;For the calculation of the strong correlation of data required by the Controller in quantitative timing,the dependency between timing data is extracted by adding the self attention layer and FC hybrid structure.The H-Trader experiment results show that these three reinforcement learning trading strategies H-Trader,Vanilla DQN and DDPG have certain performance on the four test data sets.H-Trader has even better performance in terms of total profit and risk control capacity.