Analysis Method And Its Application Of Financial Time Series Based On Machine Learning

The financial market is an open and complex system with intricate relationships among various economic variables.Mastering the volatility law and structure of the financial market is of great significance to the avoidance and prevention of financial risks.However,the financial market is affected by many factors.Various data contains a lot of noise and has strong uncertainty.Financial time series analysis is an effective tool for mastering and forecasting financial market fluctuations.Financial time series analysis refers to the use of time series data of financial assets to establish an analysis model to predict the volatility of financial assets.The study of the financial time series of Chinese A-shares provides quantifiable decision data for the investment of all kinds of investors,helps investors to predict the risks and potential opportunities in A-shares market,reduce investment losses and improve the rate of return.On the other hand,the trading behavior of investors will bring liquidity to the market,and liquidity will promote the healthy development of the market.In recent years,with the development of machine learning and deep learning theory and the maturity of technology,scholars have applied machine learning and deep learning to financial time series analysis and made many significant achievements.But these achievements still cannot meet the practical application requirements.This paper takes financial time series as the research object and uses machine learning and deep learning methods to study the feature selection algorithm,prediction algorithm and trading algorithm of financial time series.The main research contents of this paper are as follows:(1)Aiming at the existing problems of feature selection algorithm based on forest optimization algorithm in the initialization stage,population limitation stage and updating stage,a feature selection algorithm based on improved forest optimization algorithm is proposed.Firstly,in the initialization stage,the new algorithm uses Pearson correlation coefficient and L₁regularization method instead of the original algorithm's random initialization method.Secondly,in the population limitation stage,the new algorithm adopts the difference complement method of good tree and bad tree to improve the unbalance of candidate forest categories of the original algorithm.Finally,in the updating stage,the new algorithm reserves the trees with the the same fitness of the optimal trees rather than removing them as the original algorithm does.Experiments show that the new feature selection algorithm has better feature selection capability than other algorithms.(2)Aiming at the low prediction accuracy of artificial neural network and single long and short memory neural network,a financial time series prediction algorithm based on ensemble long and short memory neural network is proposed.The novelty of this algorithm is that an 11-layer long and short memory neural network is used as the base learner of Bagging algorithm to realize the combination with ensemble learning.Experiments on financial time series data show that compared with other commonly used algorithms,the predictive performance of the proposed algorithm can be improved by 5%-10%.(3)An 8-layer convolutional neural network trading algorithm for financial time series is proposed.This algorithm combines multiple one-dimensional financial time series into two-dimensional financial images and adopts a new tag generation algorithm to mark financial images into three states:buy,sell or hold.This new tag generation algorithm solves the problem of category imbalance.The algorithm uses the convolutional neural network to train and recognize financial images and uses the recognition results to make investment decisions.Experimental results show that compared with other algorithms,the prediction accuracy of the proposed algorithm is improved by 10%-15%,and the investment return is increased by 5%-15%.