Neural Network Model With Two Hidden Layers To Optimize Technical Indicators In Securities Market

Recent years,the market's trading behavior of domestic and foreign securities market tends to be standardized and trended.Trading behaviors oriented by market price rules gradually occupy a more important position.In those trading behaviors,the analytical methods is from transaction data,making technical indicators and then predicting future price trends.The pros and cons of the method of technical indicator analysis depends on the accuracy of the analysis and prediction.Data from xueqiu.com shows that in the case of a large number of multiple investments,technical indicators with more than 50% accuracy can turn a small accuracy advantage into profit.The current mainstream technical indicators are to select some eligible stocks by setting a number of screening conditions and predict future price trends.For example,"mean line long side by side" technical indicators,"similar K line" technical indicators and so on.However,the screening conditions established by these technical indicators are not precise enough,resulting in the "characteristics" of the selected stocks is not obvious enough,and low prediction accuracy.Machine learning neural network technology can train these stock data,extract hidden "features",generate models,achieve optimization technical indicators,and improve the accuracy of prediction.Therefore,this paper chooses a two-layer hidden layer feedforward neural network with better training speed and stability to optimize the technical indicators.The specific work is as follows.First of all,in order to prove that for different technical indicators,the targeted neural network training method can be used to achieve the optimization effect.In this paper,two different technical indicators are selected for experiment.The first technical indicator uses the data pattern of “horizontal long-term side-by-side” as a screening condition to obtain eligible stocks and predicts that these stocks will follow up.Due to the fixed screening conditions,the neural network training method for continuous training is selected to optimize it.The second is to use the target K line as the screening condition to obtain similar K-line,and use the similar K-line follow-up trend to predict the follow-up trend of the target K line.Since the screening conditions change with time and stocks,select a rapid training method that can form a model at one time to optimize it.Then,in order to complete the optimization experiments of the above two technical indicators,it is necessary to implement software tools and prepare experimental data.This paper builds a model library based on Python language and mongo database.The data and tools needed for the experiment are divided into three layers in the model library.The layers are: a data storage layer that stores basic data,indicator data,and model data;a data processing layer that analysis basic data to generate indicator data and a data analysis layer that implements a neural network model.Among these layer the data analysis layer implements two training forms of the two hidden layer neural network model: a fast training model that uses only one batch of training integration,and a continuous train model that uses batches of training integration from different days.The fast training model corresponds to the daily changing technical indicators,for example,a similar K-line indicator predicts stock's trend according to the historical similar K-line of other stocks.The continuous training model corresponds to the technical indicators without changing,its training sets can be generated from daily market transaction data,such as average long-term parallel technical indicators.At last,according to the above two types of training methods and the listed indicators,this paper do an experiment on neural network training and testing for two types of training methods and corresponding indicators.There are two repeated experiments for each type of training method.The average long-term parallel technical indicators,the traditional method to obtain the prediction accuracy rate of 50.4%,and through the continuous neural network training method,the prediction accuracy of the number of consecutive trainings n=1 and n=5 less than 50%,and the prediction accuracy of n=3 are 52.9% and 51.3%,and the accuracy is 2.5% and 0.9% higher than the traditional method.Through the historical stock similar K-line to predict its trend,the prediction accuracy rate is 51.31%,and the prediction accuracy of the two models obtained by fast neural network training is 54.11% and 56.3%,respectively,accuracy is 2.8% and 5.8% higher than traditional methods.Since the above two types of technical indicators cover most of the technical indicators on the market,the two training methods selected in this paper can be used for more technical indicators.The use of neural network methods to optimize technical indicators has certain application value and a wide range of applications.