Research On The Trend Feature Extraction Of Securities Data Based On Machine Learning

Stock transaction data is one kind of time series and has the characteristics of high complexity and uncertainty.This leads to only few human stock traders have the ability to analyze and deal with the features of stock transaction correctly.Machine learning can imitate or even actualize human's thinking model,through learning algorithm as well as massive historical data,that makes machine learning an effective tool to extract trend features of stock data.In this thesis,two different machine learning models such as convolution neural network and Long Short-Term Memory,along with two feature extraction method of classification and regression,are adopted to extract the trends features of stock data,resulting in trend classification,trend strength and trend endurance.The main research contents are as follows:(1)Obtain and preprocess stock historical transaction data.In this thesis,stocks' historical daily transaction data is gathered automatically from the Sina Finance website.Then daily transaction data is pre-processed for machine learning application,and the data processing includes historical data supplementation,daily data merging into weekly data and daily weekly calculation.Finally,technical indicators of stock data are calculated,such as moving average,moving average convergence divergence and so on.(2)Label tock data with trend features.This thesis adopts a supervised machine learning algorithm which requires input data along with desired result during the training process.Three trend features of daily weekly data are labeled by moving average convergence divergence algorithm,and to calculate the expected result value for future training.(3)Train machine learning model.This thesis adopts a convolutional neural network and a Long Short-Term Memory as the fundamental framework respectively to construct two different machine learning models.Both models adopt sigmoid function or softmax function to obtain trend classification,and calculate trend strength and trend endurance by linear regression function.The cross entropy or minimum mean square error is used as the objective equation for such models,as well as adaptive learning rate algorithm is adopted to optimize the models parameters.Finally,the loss value of two models converged to appropriate value.(4)Performance tests.The performance of the proposed machine learning model for stock trend features extraction was evaluated by analysis of the accuracy or error between the output of the model and desired values.The results show that the proposed learning model can provide excellent classification accuracy close to 70%,along with relative error less than 50% for trend strength,and relative error more than 50% for trend endurance.Finally,this thesis gives the suggestion of optimal machine learning model structure through comparison and analysis of performance test results.This thesis proposed two machine learning models to extract the trend feature of stocks data such as trend classification,trend strength and trend endurance.The performance test results show that machine learning model is outstanding for trend features extraction of stock data.