Financial Time Series Modeling And Forecasting

With the continuous development of program trading,the trend forecasting algorithms of financial time series have been widely studied.On the one hand,direct prediction of financial time series has been studied in this dissertation,and algorithm optimization has been realized aiming at overcoming such problems that occurs in the application of the BP neural network in prediction as slow convergence rate,easy convergence into local extreme value,unstable training results and limited prediction accuracy;On the other hand,improved pattern recognition algorithm has been proposed based on SAX.Detailed contents are as follows:First of all,due to low prediction accuracy of BP algorithm,a hybrid system of financial time series has been proposed which processes data first before inputing the BP neural network.In this system,historical data are used to select sub-series which matches best with the test sample data,and selected sub-series are used to train BP neural network,which has constructed the based-on-pattern-recognition BP neural network prediction model.Simulation results show that the prediction error of the proposed hybrid forecasting system is smaller and the prediction effect is more accurate than that of trend-based forecasting method or multi-input BP neural network forecasting system by analyzing RMSE,MRE and WDS of different data segments.Secondly,since there are still weakness of long training time and easy convergence to local extreme of BP algorithm,an improved PSO-BP prediction algorithm has been presented,which has better global search ability by training the weights and thresholds of BP neural network with the improved PSO algorithm.Then BP neural network is used to furtherly optimize the weight and threshold due to its local search ability.Such BP neural network parameters have been determined to predict the time series.Simulation results from both aspects of the execution efficiency and accuracy prove the feasibility and effectiveness of the proposed PSO-BP forecasting model algorithm.Finally,a novel algorithm of symbolic aggregation,which uses the minimum value,the average value and the maximum value of each segment of the subsequence to represent the symbols,is presented in this paper in order to effectively identify several common patterns in stock price series.Cosine similarity is used to measure the similarity of the model.Simulation results show that the proposed algorithm is more efficient than traditional SAX algorithm and TD_SAX algorithm in terms of matching time and matching accuracy.Less time is required and fewer false-negative and false-positive sequences are generated compared with SAX and TD-SAX.