Financial Time Series Prediction Using Swarm Intelligence Techniques

Predicting the future value of fnancial time series, due to its potential proft, hasattracted increased attention from investors and researchers. However, due to the in-herent complexity of fnancial time series, which is non-linear, non-stationary, noisy,and deterministic chaotic, fnancial time series prediction is considered as one of themost challenging research topics in modern time series analysis.Recently, support vector regression (SVR), a machine learning algorithm basedon statistical learning theory, has been widely used in fnancial time series predictionand has shown better prediction performance than artifcial neural network based ap-proaches. This is mainly due to the fact that SVR has good nonlinear approximationability, a quick convergence rate, global optimal solution and high generalization abil-ity. Inthisthesis,SVRisusedincombinationwithothermachinelearningandartifcialintelligence techniques to increase the prediction accuracy of fnancial time series.The choice of SVR parameters has decisive efect on the prediction accuracy andgeneralizationabilityofSVR.Particleswarmoptimization(PSO),aswarmintelligencetechnique, is used to fnd the optimal SVR parameters using cross-validation.The selection of features to be used as inputs and outputs of SVR models is alsoinvestigated. Experimental results show that better prediction results can be achievedby choosing appropriate features.Due to the non-stability of fnancial time series, single SVR model sufers fromthe problem of unstable prediction accuracy. Inspired by ensemble learning methods,a prediction algorithm that combines multiple SVR models is proposed. Multiple SVRmodels are trained using diferent subsets of the training data. Prediction is based on the weighted sum of the outputs of these models. Weights of the models are adjustedaccordingto their previousprediction errors. The algorithm reduces the overallpredic-tion error by exploiting the diversity of the models. By dynamically adjusting modelweights, the algorithm is self-adaptive, capable of dealing with non-stationary timeseries.The proposed method is tested using stock price series from fve major fnancialmarkets. The results show signifcant enhancement of prediction accuracy and gener-alization performance in comparison with single SVR model.