Stock Price Prediction Based On Neural Networks

Stock market is an important component in an economic system. The problems of stock market analysis and prediction have been widely studied. Researchers have proposed many theories and technologies about building computer systems to analyze and predict the behavior of a stock market. Artificial Neural Network, because of its abilities to deal with indeterministic system with many correlative factors, has been used in many stock market prediction systems.In this thesis, we study issues of building a neural-network-based stock price prediction system. Stock price prediction is a typical time series problem: the current price of a stock is affected by the previous stock market factors. To build a neural network model to predict the stock price, the following issues must be considered: data preprocessing, structures and learning algorithm for the neural networks, and selection of regression time frame. In our study, we find out that the traditional solutions for these issues can be improved. The contributions of our work are:First, we proposed a new data preprocessing method, correlative principal component analysis. The traditional data preprocessing method, principal component analysis is inefficient to select the most important factors because it does not consider the relation between the original data and the objective. Correlative principal analysis takes the objective into consideration. Our experiments show that correlative principal component analysis is more effective than principal component analysis in stock price prediction system.Second, we improved the traditional BP algorithm by adding a new parameter. So, we only need to tune up one parameter instead of two. In our study, we also introduce a time sensitive coefficient in error computing formula in...