Information Fusion And Computational Intelligence Models For Financial Time Series Prediction

Financial market is a huge system with a complex reaction pattern which can be affected by comprehensive information from various aspects.Financial time series,as the most important and largest number of data in financial market,is an external expression of the comprehensive internal of financial market.Through the analysis and prediction of financial time series,we can discover underlying rules and extract useful information to provide important basis for financial activities and decision-making,which has very important practical significance.Along with the integration and development of database,parallel technology and artificial intelligence,data mining technology is produced by cross disciplines.It can extract the hidden useful knowledge and rules from a large number of historical data through data integration,protocol,cleaning,transformation,mining,pattern evaluation and knowledge representation.It provides effective theoretical and technical support for us to analyze massive financial time series.Therefore,this dissertation researches on the information fusion and computational intelligence model for financial time series prediction.Specifically,focus is on the feature extraction process,the basic prediction model,and the information fusion of financial time series prediction.As a whole,the dissertation innovatively proposes a class of financial information fusion and computational intelligence models for generating daily prediction of stock index and individual stocks.The main research contents and results can be divided into five parts as follows.First,this paper takes the feature extraction process as the foremost key step of the computational intelligence model for financial time series prediction.The financial time series with the high noise,chaotic,nonlinear and non-stationary characteristics requires real time analysis and prediction.Thus this paper constructs a complex nonlinear feature extraction process integrating empirical mode decomposition for financial time series(FtsEMD)and principal components analysis(PCA),where the FtsEMD uses a sliding window technology with empirical mode decomposition(EMD).EMD is a suitable method for dealing with nonlinear and non-stationary time series,and PCA is a linear dimensionality reduction method which can maintain the most information of the original data.The combination of FtsEMD and PCA is equivalent to a nonlinear PCA.Therefore,the nonlinear feature extraction process integrating FtsEMD and PCA has the adaptability,comprehensiveness and orthogonality of feature extraction.Second,this paper takes the optimization and improvement of the basic prediction model as the basic task of the computational intelligence model for financial time series prediction.We proposes a nested k-nearest neighbor(NKNN)based on adaptive affinity propagation clustering(AAP).In this new algorithm,AAP is used to transform the feature set into clusters and then input to NKNN for regression and prediction.NKNN is a nested reformulation of KNN that can tackle the two main deficiencies of KNN: 1)the large amount of computations;2)the large volume sample account for the majority of k nearest neighbors in the situation of disequilibrium samples.NKNN consists of three functions: NKNN0,NKNN1,and NKNN2.It works in two steps: 1)calculate the similarities between the most recent time point X(t)'s price behavior pattern and any cluster center to find the nearest one;2)calculate the similarities between X(t)and any sample in the same cluster which is the nearest one to find the k nearest neighbors of X(t).The first step can reduce the amount of computation and the second step can avoid disequilibrium samples since it works in one cluster.Third,this paper optimizes the feature extraction process and the basic prediction model gradually,and constructs a series of computational intelligence models for financial time series prediction,including PK(PCA-KNN)model,FEPK(FtsEMD-PCA-KNN)model,PANK(PCA-AAP-NKNN)model and EPAK(FtsEMD-PCA-AAP-NKNN)model.The structure of the EPAK model as a whole is original.Moreover,the EPAK model integrating a nonlinear feature extraction process and a NKNN regression based on AAP has better prediction performance than others.Fourth,by the idea of financial information fusion,this paper proposes a complex prediction model for stock market index synthesizing all the sector index predictions using EPAK as a kernel.In this complex prediction model,EPAK is first used to predict each of all the sector indices of the stock market,and the sector indices predictions are then synthesized via weighted average to generate the prediction of the stock market index.This multivariable information fusion prediction is theoretical and innovative for the analysis and prediction of Chinese stock index..Fifth,for verifying the effectiveness of the proposed information fusion and computational intelligence models for financial time series prediction,this paper makes an empirical analysis and comparison on the real historical data of a single stock-yonyou which is selected random and the CSI 300 index(the Chinese benchmark index)during 10 years.The empirical results shows: 1)the probability prediction accuracy of the PK model is higher than that of the KNN model;2)the prediction accuracies of the FEPK and PANK models are higher than that of the PK model,since the feature extraction process of FEPK and the basic prediction model of PANK are improved base on the PK model;3)the prediction accuracy of FEPK model is higher than the PANK model.This result implies that the improvement of the feature extraction process can rise the prediction accuracy of the whole model more than the improvement of the basic prediction model;4)the EPAK model integrating the nonlinear feature extraction process of FtsEMD+PCA and the NKNN regression based on AAP performs best with a highest hit rate;5)the complex prediction model for stock index is also tested on real historical data of CSI 300 index,with results confirming the effectiveness and better performance than EAPK predicting CSI 300.