Research On The Methods For Stock Prediction Based On Representation Learning Of Continuous Trend Of Time Series

Accurate prediction of stock trends can better improve returns and reduce risks.In recent years,with the development of machine learning,especially deep learning technology,its ability of automatically extracting features and solving nonlinear problems has been widely used in the task of stock forecasting.However,the fluctuation of stock price is featured by high randomness,chaos and high noise.With the fluctuation of stock price,the effective information of stock time series,such as transaction price and trading volume,is insufficient,and thus it is difficult to obtain accurate prediction results using the existing prediction methods based on stock time series.How to design solutions according to these characteristics,improve the accuracy of stock trend prediction,and finally build a good portfolio based on the prediction results to reduce risks and improve returns is challenging for scientific research.Given high randomness of the labeling results of traditional stock time series labeling methods and insufficient effective information of stock time series,this paper proposes several stock prediction methods based on the continuous trend representation of stock time series.To solve the problem that the traditional portfolio theory does not consider the dynamic position adjustment according to the model prediction results to achieve risk aversion,a portfolio optimization method based on continuous trend representation return revision and clustering is proposed.The current research mainly includes four aspects as follows.Firstly,given the strong randomness of stock time series and the high randomness of the labeling results of traditional stock time series labeling methods,a labeling method based on the continuous trend of stock time series is proposed.Data labeling is an important way of data representation learning.Different data labeling methods will get differentiated representation information,which will lead to the difference of model prediction results.In essence,the existing labeling methods focus on short-term price fluctuations to label the data.The labeling process is very vulnerable to randomness,which leads to the labeling results with high randomness.This paper proposes a labeling method based on the continuous trend representation of stock time series,which can reduce the impact of short-term random fluctuations of stock prices on the labeling method.Meanwhile,it can also be better used to train machine learning models and improve the accuracy of stock trend prediction.The experimental results show that the labeling method proposed in this paper is a better labeling method in terms of prediction accuracy and other evaluation metrics.Secondly,in view of the insufficient effective information of stock time series and the insufficient decision-making basis provided to the model,this paper proposes a stock prediction method based on industrial chain transfer learning with the continuous trend representation information,which improves the accuracy of prediction by combining transfer learning of continuous trend representation information with deep learning models.By constructing paired upstream and downstream industry index,after fully training multiple deep learning models on the time series of the industry stock price index corresponding to the upstream industry index,the whole deep learning network structure and network parameters are transferred to the time series of the industry stock price index corresponding to the downstream industry for secondary training.Based on the facts above,some external information related to the industrial chain of upstream industry index is introduced to enhance the prediction ability of the model.The experimental results show that the model trained by the learning process of industrial chain transfer learning of continuous trend representation information improves the accuracy of the model in predicting the future trend of the stock industry index,and the results are supported by statistical tests.Thirdly,aiming at the decline of prediction accuracy caused by the high noise characteristics of stock time series,this paper proposes a wavelet transform and autoencoder stock prediction framework based on continuous trend representation.Firstly,discrete wavelet transform is used to denoise the noise of the raw stock time series,which reduces the impact of noise on the trend continuity of the stock time series.The denoised data show a more stable trend continuity characteristic.At the same time,in order to improve the model’s ability of judging the importance of features,an autoencoder with continuous trend representation penalty is designed to judge the importance of data features after noise denoising,so that the model can give greater weight to the features that contribute more effective information to improve the prediction accuracy.The better performance of the proposed hybrid framework is confirmed by experimental results.Finally,on the basis of the previous research,in view of the problems that it is difficult to determine the pool stocks in the traditional portfolio construction process,the weight calculation only considers the covariance information,and does not fully consider avoiding the downside risk,this paper proposes a method based on the revised return of the continuous trend representation and distance clustering to optimize the portfolio.According to the continuous trend representation of the stock,a classifier model that can predict the continuous trend changes is trained,and the traditional calculation of return is revised,which realizes the dynamic position adjustment and weight update of the portfolio,and effectively avoids the risk of portfolio decline.Through clustering analysis of different stocks based on distance clustering algorithm,the construction of stock pool is achieved according to certain strategies,and the problem of expected stocks entering the pool is solved.By introducing distance information,the problem that the traditional portfolio theory only considers the covariance between different stocks is solved.The experimental results verify the calculation method of the optimal number of stocks and the optimal weight after balancing the risk and return in the portfolio.Meanwhile,it is also shown that the portfolio optimization method proposed in this paper significantly improves the return and reduces the risk,which proves the better performance of this method.