A Quantitative Investment Strategy Based On Long Short-Term Memory And Drawdown Control

In many cases,the relationship between financial variables is nonlinear,and the true global optimal solution is beyond the traditional quantitative models.In the existing machine learning framework,the use of neural networks in quantitative investment can help us to research the complex nonlinear regularity between variables.Relying on the computer,the global optimal solution can be found through the iterative algorithm.The purpose of this paper is to build a new quantitative investment strategy,which combines the deep learning with the traditional investment theory,so that it can be applied to the domestic financial investment market.In the framework of the long and short term memory(LSTM)network,this paper constructs a recurrent neural network which can process a long term sequence by designing neural nodes,activation functions and output functions.Then,on the basis of the existing stochastic control asset allocation theory,the maximum drawdown constraints are adjusted from the whole time series to the dynamic time series.And,the drawdown constraints are used only for risk assets.Finally,the investment is considered under the condition that the domestic market is not allowed to be short.According to the empirical results of this paper,the accuracy and AUC value in the LSTM model of risk assets are high,while the two indicators of the risk-free asset are low.After the modification of the drawdown constraints model,the winning rate of the revised portfolio returns is obviously improved,and the actual drawdown will rise slightly.In the actual investment,the forecasting ability of the model is still good because the choice of riskless assets is not much,and the difference is small and the volatility is not strong.Although the cost of the improved asset allocation model is to increase in actual retracement,it is still within the maximum acceptable drawdown range,so the overall application capacity of the model is effective.