The Research On Key Techniques Of Time Series Prediction Based On Machine Learning

Time series is an essential kind of time-based data that appeals in a varity of areas, including finance, physics, biology, medicine and meteorology, to name just a few.Therefore, the time series prediction is worth exploring and of critical importance in applications. Nowadays, since time series is always infected by numerous factors, traditional time series prediction methods are limited in some cases. Admittedly, a growing number of researchers focus on introducing machine learning methods into time series prediction.However, as the increasing speed of time series growing and the common appearance of high-dimensional time series data, the machine learning based time series prediction faces a series of problems. One of the key problem is how to process high-dimensional time series data at an extremely high speed. Solving this problem is significance to improve the performance of machine learning based time series prediction and cannot leave feature selection technique and extreme learning technique. The work in this thesis focuses on the key techniques of time series prediction based on machine learning, learning high-dimensional feature data at a high speed. Its main contributions are summarized as follows.(1) We propose a binary feature selection framework in kernel spaces. In this paper, we propose a binary feature selection framework in kernel spaces, where each feature is projected into kernel spaces and a binary classification task is constructed in this space. Subsequently, the features are selected according to the normal vector of the learned classifier, which reflects the importance of each feature. To achieve the effect of feature selection, an ?1-norm regularization is imposed on the normal vector to enforce its sparsity. Also, our framework can be naturally extended to the semi-supervised feature selection scenario via the well-known manifold regularization technique. Furthermore,the issue of eliminating the potential redundancy among the selected features is well discussed. Finally, we provide some theoretical results which guarantee the feasibility of the proposed framework. Comprehensive experiments have been conducted on six benchmark data sets and the results demonstrate the performance of our framework.(2) We propose a fast multiple kernel learning approach based on distance.In this paper, we propose a distance based multiple kernel extreme learning machine(DBMK-ELM), which provides a two-stage multiple kernel learning approach with high efficiency. In specific, DBMK-ELM first projects multiple kernels into a new space, in which new instances are reconstructed based on the distance of different sample labels.Subsequently, an ?2-norm regularization least square, in which the normal vector corresponds to the kernel weights of a new kernel, is trained based on these new instances. After that, the new kernel is utilized to train and test Extreme Learning Machine(ELM). Extensive experimental results demonstrate the superior performance of the proposed DBMKELM in terms of the accuracy and the computational cost.(3) We apply the key techniques of machine learning based time series prediction into commercial intelligence model. In this paper, we apply the key techniques of machine learning based time series prediction into commercial intelligence model to construct a stock decision support system, in which the machine learning based time series prediction is well combined with trading boundary model, an improvement model based on oscillation box theory. Specifically, the system first uses the binary feature selection method in kernel spaces to select key features for time series prediction. Then, it introduces the distance based multiple kernel extreme learning machine to learn the history data of stock market and forecast future stock price using the features extracted in the above step. After that, a robust trading strategy based on trading boundary model is used to give the trading support. Comprehensive experiments in real history data of stock market demonstrate the performance and usefulness of our proposed system. In other word,the system not only can obtain fruitful profit but also can maintain a low investment risk.