支持向量机选择及其在股票走势预测中的应用

Application of Online Selection Support Vector Classification in the Prediction of Ups and Downs in Stock Market

郭辉

支持向量机（SVM）是数据挖掘中的一项新技术，是借助于最优化方法解决机器学习问题的新工具。在研究了股票数据的特点以及对股票预测的研究结果后，本文根据传统的SVM算法原理，提出一种在线选择训练样本的在线增量训练的方式完成模型更新的动态预测模型（DMDI），使得仅增加较小工作量为代价而获得更高的预测精度成为可能。应用DMDI对股市的大盘和个股的走势分别进行中短期预测，并跟神经网络的预测结果进行了比较。大量数值实验表明，DMDI模型比不进行选择的静态模型和神经网络模型对股票走势的预测更为有效，具有明显的比越性。[著者文摘]

Support Vector Machine （ SVM ） which is a new technology used in Data Mining. It is a new tool that accounts for the problems of the Machine Learning by the method of the optimization, Applying the support vector machine method in the research on the non-linear time series economic prediction problem is underway. It is more feasible and predominant than the Neural Networks algorithm in the extending ability and the tallying precision. After we studied the characteristics of the stock data and the rules of the stock market people, we put forward to a dynamic model which bases on the traditional support vector machine arithmetic. The model selects the training data online when we get the new data and then we modify the model each time base on the increased data in the aggregate. It is a dynamic model, so it can catch the real time change of the market. It make the prediction precision be improved comes to truth with the small workload as the cost. In this paper we use the support vector machine and the Time series dynamic model （DMDI） to predict the short-time and the medium-terrn ups and downs in the single stock and the holistic Shanghai stock market. We perform a large numbers of numerical experiments and compared with the results being got based on the methods of the BP neural networks and the static models which is not changed when the new data is got with the time going, and the prediction rightness probability is higher, and it is more feasible in the extending ability and the tallying precision through the actual application. In addition, It can also avoid the difficult problem study of the training data excessively. The results show that the DMDI is more suitable for the forecasting the index time series of the stock market than the BP neural networks and the static models. The model we have proposed in this paper has more advantages in the prediction of the trends of the stock market than the conventional methods.[著者文摘]

一种新的计算方法：粒度进化计算

A New Computing：Granular Evolutionary Computing

蒙祖强[1] 蔡自兴[2]

从分析进化计算的起源入手，总结了进化计算产生的根源，然后探讨了人类进化——文化进化的基本特征，模仿进化计算的来源机制，提出了粒度进化计算。如果说进化计算是模拟达尔文生物进化机制而发展起来的一种计算方法，那么粒度进化计算则是在模仿文化进化机制的基础上，综合了Agent技术以及粒度计算、进化计算的理论和方法而提出的一种计算方法。文中，从群进化和超群进化两方面来介绍粒度进化计算的基本原理和方法，并给出了基于粒度Agent系统的粒度进化递归模型。[著者文摘]

Beginning with analysis of Evolutionary Computing,Granular Evolutionary Computing（GEC） is put forward by means of summing up foundation stone of Evolutionary Computing and basic feature of Culture Evolutionary.GEC comes into beings by integrating agent technology and theory of Granular Computing and Evolutionary Computing on the basis of simulating mechanism of Culture Evolutionary,while EC grew by means of simulating biological evolution developed by Charles Darwin.In this paper,general principles and methods of GEC are introduced with a view to group-evolutionary and super-group-evolutionary,and reeursive model of granular evolutionary based on multi-granular-agent system is given.[著者文摘]

BP算法和对称ARCH类模型对股市波动性预测的实证比较

Comparison： the volatility forecasting of BP algorithm and symmetric ARCH model to stock market

庞素琳[1] 徐建闽[2] 黎荣舟[2]

利用我国深圳股票市场的实际数据，建立了相应的BP算法网络预测模型和ARCH（1），GARCH（1，1）预测模型，分别用来对深成指数每个周末收盘价的波动性进行预测．研究表明，BP算法对样本外观测值的上凸曲线拟合得较好，对下凸曲线的拟合效果较差；ARCH（1）和GARCH（1，1）则反之，其预测曲线对样本外观测值的下凸曲线拟合效果都较好，但对上凸曲线的拟合效果都较差．通过采用6种常用的预测误差统计量：平均误差、平均绝对误差、均方根误差、平均绝对比率误差、Akaike信息准则、Bayes信息准则对样本外数据的预测结果进行检验，BP算法的预测效果最好，ARCH（1）模型次之，GARCH（1，1）模型偏差．[著者文摘]

Three forecasting models, called BP algorithm, ARCH（1） and GARCH（1,1）, are established based on the actual data of Shenzhen stock market, China. The proposed three models are respectively used to predict the volatility of the weekly closing price of the composition indexes in Shenzhen Stock Exchange. Furthermore, six common statistical methods of the forecasting error, i.e., mean error （ME）, mean absolute error （MAE）, root mean squared error （RMSE）, mean absolute percentage error （MAPE）, Akaike＇s information criterion （AIC） and Bayesian information criterion （BIC） are used to test the forecasting results of the out-of-sample data. The results show that the forecasting result of BP algorithm is the best, the ARCH（1） model takes the second place and the GARCH（1,1） model is the worst.[著者文摘]