Font Size: a A A

Stock Market Tendency Forecasting Based On Optimized SVM Using GA

Posted on:2011-12-24Degree:MasterType:Thesis
Country:ChinaCandidate:W ZhangFull Text:PDF
GTID:2189360305455242Subject:Computer application technology
Abstract/Summary:PDF Full Text Request
As an important part of the financial markets, the stock market plays an important role for the stable development of China's economy. As a means of capital accumulation, the stock attracts millions of investors from its birth. Stock price forecasting is a hot issue of financial forecasting. A large number of financiers, computer experts and scholars generated great interest, and they attempted a lot of theories and methods to predict the trend of stock market.Currently, there are many stock prediction methods mainly including traditional methods, artificial neural networks(ANNs) and support vector machine(SVM).However,due to the complexity of the internal structure of the stock price system and the changeability of the external factors, the results of traditional forecasting methods, which simulate and predict the object pattern in linear model, are not satisfactory. ANNs imitating the human brain mechanism is a good parameter estimation method and has good nonlinear approximation ability. But ANNs is prone to leading local extremum and overfitting, and it's forecasting results are not very stable.SVM method is a new widely-used machine learning method especially for limited sample datasets within the framework of statistical learning theory. SVM achieves the structural risk minimization and it has solved local extremum and overfitting problems.SVM has a good generalization ability and is an excellent prediction method. In this paper, SVM was applied in forecasting the tendency of stock. The stock tendency can be labeled as"up"or"down"in modeling. So the stock tendency forecasting is converted to two-classification problems, and SVM is initially studied in pattern recognition for linear separable two-classification problem. In modeling by SVM, we need to select the kernel function, determine the parameter's value, extract the features and pretreat the sample datasets.SVM is regarded as the best classifier for the limited sample datasets, however, a large number of features would affect not only the convergence speed, but also the accuracy of classification algorithms if there were some redundant features. The kernel parameter ? and penalty factor C also have great impact to the accuracy of forecasting by SVM. Therefore, in order to forecast accurately and fast, correctly setting the kernel function parameter ? and penalty factorC , finding the smallest feature sets are critical issues. This process is called the optimization of SVM.For the SVM optimization, genetic algorithms (GA) is undoubtedly a good choice. GA is an adaptive global optimization random search algorithm which is developed from natural selection and genetic mechanism and has a strong global search ability, which does not depend on the specific solution model. In this paper, the optimized SVM using GA method (GA-SVM method) was proposed which combined the basic idea of GA and SVM. The method use GA as a searching algorithm for parameters and features and use SVM to evaluate the searching results of parameters and features. It was proved in our experiment that GA-SVM method could find the best parameters and filter out redundant and irrelevant features. The GA-SVM method has certain advantages not only in forecasting accuracy but also in forecasting speed.This main work is as follows:1. Explain the significance of the stock forecasting. Introduce the status quo of theories research in stock forecasting domestic or abroad. Analyze and summarize the basic theories of stock forecasting.2. Study the statistical learning theory and SVM systematically. First, explain why the experience risk minimization could be replaced by structural risk minimization and introduce the most important conception of Vapnik-Chervonenkis dimension and generalization bound in statistical learning theory. Then, introduce modeling and theory of SVM which is a representation of the statistical learning theory. SVM has good generalization ability for limited training sample datasets, and it solutes the dimension disaster in the high-dimensional feature space by turning the inner product operation in high-dimensional feature space into the kernel function computation in the low-dimensional space function.3. Establish the stock prediction model by SVM. First, determine the kernel function and the parameters in SVM modeling; Then select the feature by experience; Finally, normalize the sample data to improve the speed.4. Establish the stock forecasting model by GA-SVM method. First, the features was coded in binary mode. Each feature was taken as one gene corresponding to one binary bit.Then, generate population randomly and calculate its fitness value. We adopted the concept of sensitivity and specificity in diagnosis and adjusted their weight to strengthen the target of prediction. Finally, compare the results with the termination condition, if not satisfied then generate the next generation of population, calculate the fitness value of the new population, and cycle until satisfied.5. Do experiment to verify the effectiveness of GA-SVM method in stock tendency forecasting. The results were obtained using GASVMMining software the author developed with VC6.0, access database and LibSVM library in Windows XP. The sample datasets were Shanghai composite index data of totally 360 trading days from October 6, 2008 to March 24, 2010 and select 20 features for each sample data. We compared the results of forecasting by GA-SVM with by SVM method, with by traditional method and with by BP method. It was proved by the experiment that GA-SVM method reduced computational complexity and improved forecasting accuracy by optimizing the parameters and filtering out the redundant features.6. Summarize the work in this study. Summarize the work and find the inadequate part and give some reasonable prospect in later research.Above all, this paper has completed the whole process for stock tendency forecasting of modeling - optimization - evidence, and designed and developed a smart stock forecasting software. This research confirms that GA-SVM method is feasible in China's stock market forecast, and offers certain technology reserve to future research.
Keywords/Search Tags:Stock market tendency forecasting, Support vector machines, Genetic algorithm
PDF Full Text Request
Related items