A Data-driven Research On The Model Of Financial Time Series Forecasting

Nobel Laureate in Economics, Robert Merton considers that the core issue of modern financial theory is how to configure resources optimally under uncertain en-vironment. According to the knowledge of Nonlinear Dynamics, the uncertainty of the modern financial theory just originates from the fact that financial system itself is an Open Complex Giant Systems affected by many complex factors. Accordingly, the financial time series data treated as the observation of the system can be looked as the expression of the complex motion law of the system. Literature studies of financial time series show that both single model including traditional statistical methods of linear paradigm and nonlinear computational intelligence methods and hybrid methodologies combining several different models improve the stability and accuracy of the financial time series forecasting problem. But most of them are lack of effective learning of experience knowledge including the internal time cor-relation, the trend information of the price series and mutual information between different markets and so on, which restricts the further improving of the prediction performance of the traditional models.Therefore this study creatively develops the bottom-up program of model-ing and application through machine learning method incorporating the experience knowledge from financial time series data. For the different types of financial time series data this study initiates a reconstruction of forecasting model based on the analysis of the complex relationship among the financial time series data in conjunc-tion with the latest research on intelligent computing, computational experimental finance, data mining, and control theory and so on. The main research results and innovations are summarized as follows:1. For the univariate financial time series data, considering the importance of the data trend and the overnight gap-opening information, this study develops the differential information based and the gradient information based single forecasting model by means of tracking differentiator to extract the approximate differential. From the hybrid model construction perspective, this study proposes a prediction model based on ARIMA and Taylor expansion method to improve the learning efficiency of the experience knowledge among the series data.2. For the multivariate financial time series data with high dimensional com-plexity, this paper proposes a SVM-GARCH forecasting model based on nearest neighbor interaction information, which aims to integrate the linkage information between the different financial markets and improve the traditional model's process-ing capability for nonlinear component of the series data. In order to improve the stability and precision of the prediction result, a novel hybrid model based on time geodesic distance is proposed to improve the leaning efficiency and generalization ability to deal with the time correlation knowledge of the financial time series data.Data-driven modeling method form bottom to top of complex system is to analyze the observed data and start reverse study of the financial system, which is advantageous to overcome the weakness of the strict hypothesis in normative analysis and sensitivity test dilemma in empirical analysis. And this modeling solution is highly regarded as the unity of financial theory and its practice. This study develops a series of single and hybrid models based on financial time series data with different dimensions, which reflects the view of complex system modeling applications in the financial sector. And all of these reconstruction and innovation of the models are expected to find out the regulars dominating the complicated financial phenomena, guide financial practice, improve market efficiency and provide new impetus and direction for the relevant financial theory innovation research.