Stock Market Trend Prediction By Combination Models

To predicate the future with high accuracy is a holy grail in financial market.However,the volatility of chaotic financial market challenges new technologies from computer science to economic science all the time.In recent years,the development in data mining algorithms has provided many new methods and new ideas for the research work in this field.Financial time series data,as the external manifestation of the stock market,contains a lot of information.It is an essential task to analyze financial time series data and predicate the trend in financial research.This thesis divides existing predication algorithms of time series analysis into two categories,including 1)the algorithms by predicting behavior patterns in time series,such as Motif,self-organizing maps etc;2)the algorithms by minimizing the loss function,i.e.,combining real data distributions to achieve approximate results of real outputs,such as neural network,SVM etc.In this thesis,we take the time series of stock prices as the research objects and explore two kinds of algorithms.First,due to the characteristics of dynamic changes in financial time series,it is difficult to determine the starting point.We define a set of representative patterns which contains the pattern information.Based on that,We present the RPD algorithm to extract representative patterns and the PPEM prediction algorithm based on partial matching and elastic matching.Additionally,the LSTM neural network in the deep learning is also adopted to model the stock price time series and investigate its practical effects in the financial field.As we all know,the price information we obtained is only a low-dimensional projection of all the market information.Single predication method is on the basis of one-sided viewpoint.In this thesis,PPEM and LSTM at first are applied to predication from pattern mining and data fitting respectively.There are a potential complementarity for the predication results from the two algorithms above.Therefore,we present a combinational model by taking advantages of two algorithms,named the LSTM-PPEM model by means of linear combination and residual fitting.And then we verify the effectiveness of the proposed method through experiments.We use real stock trading data to evaluate the performance of the proposed algorithms in the thesis.We find that the PPEM and LSTM methods are both with a good performance on the prediction accuracy rate with an average of 55%,but PPEM has a high RMSE.In our combination models,the prediction model based on linear combination out-performs LSTM with an improvement over 2% on the prediction accuracy rate,and maintains the lowest level on RMSE with a decrement from 3% to 14% over other three methods.However,due to the low correlation of PPEM prediction residuals,the combined model based on residual fitting increases RMSE and the prediction accuracy rate remains only about 52%.