Multi-task Learning With Mixture Density Networks For Stock Auto Trading

The financial industry is an important part of China’s modern economy.With the vigorous development of artificial intelligence technology,artificial intelligence technology has gradually penetrated into all walks of life and empowered all walks of life.The financial industry has also ushered in a new financial technology era of product innovation and technological change.Under the tide of artificial intelligence,major financial institutions iterated rapidly and scrambled to launch various artificial intelligence financial services and financial products.For example,intelligent investment adviser of ToC financial products.An important ability of intelligent investment advisers is to use a large number of existing trading data to predict the fluctuation trend of stocks,and make trading decisions and trading operations according to the prediction results.According to the efficient market hypothesis theory,the stock market price is random and unpredictable.Through the study of Brownian motion,price random walk and bell curve(normal distribution),Edward Thorpe found that although the price change is unpredictable,the probability of price change can be quantified.This research lays a foundation for securities quantification and fully automated trading.Nowadays,the quantification and fully automated trading of securities have entered an intelligent era.From pure mathematical model to simple machine learning model LR,xgboost to neural network model and deep neural network model,securities trading is constantly exploring and advancing from simple strategy to intelligent strategy.The automatic and intelligent trading of securities promotes the optimization of human cost,the improvement of risk control ability and the stability of securities trading income.Combined with the characteristics of securities trading decision-making perspective prediction and the current cutting-edge deep learning technology,this paper designs a new deep learning network-multi task mixed density network.For intelligent decision-making in securities trading,from the perspective of prediction objectives,at present,most of them are prediction of price,rate of return,risk probability,etc;In terms of deep learning technology,at present,more research is focused on time series model LSTM sequence and deep reinforcement learning model,such as dqn.This paper holds that the decision-making of securities trading needs to consider many factors,such as price,rate of return and risk probability,and the current research technology is a single task learning system that only models a single goal,which does not make good use of the correlation between tasks to improve the learning effect.This paper holds that a multi task learning system should be established for the prediction of decision-making in securities trading.At the same time,because the efficient market hypothesis of the securities market believes that the price itself is random,this paper believes that the estimated price itself has great uncertainty.However,the probability of price change follows a certain distribution.Therefore,this paper believes that the idea of mixed density network can be introduced.When estimating,outputting a mixed distribution and making the mixed distribution approximate to the distribution of real value will be much better than directly predicting a value,such as price stability and accuracy.According to the above situation,a new deep neural network of multi task mixed density network is proposed in this paper.In the experiment of this paper,the multi task learning part applies the MMOE(Multi-Gate with Multi-task of Expert)network that is not sensitive to task relevance and suitable for all tasks proposed by Google on KDD in 2018,and accesses the mixed density network MDN on the upper layer of the task tower to output the probability density distribution.This paper verifies that multi task mixed density network can bring additional benefits on UCI’s population survey income data set used in MMOE paper;On this basis,the multi task mixed density network is applied to the transaction by transaction data set of Shanghai and Shenzhen in February to predict the stock trading decision,and the expected results are also obtained.It is concluded that in the stock trading decision-making problem,the multi task mixed density network proposed in this paper can obtain better prediction effect,and the multi task mixed density network is applicable to other similar problems.