The Research Of Stock Price Predition Based On BP Neural Network

Stock market is a significant component of modern financial market. The stock prices are affected not only by the internal factors, but also by the impact of macroeconomic factors and market factors. Due to the influence of many factors, it is hard for the investors to grasp stock changing rules. Therefore, it has actual application value to build a model for predicting future stock prices accurately.The BP neural network memorize variation rules of the stock in the weights of neurons through learning historical data of the stock price, and then predict future stock prices by the trained network. In this thesis, firstly we improve the activation function, use it to build the univariate forecasting model based on BP neural network and test the effectiveness of the activation function. And determine the input data and the target output fitting mode combined with the weekend effect, but the prediction result is less than ideal. Then we improve the fitting mode of input data and target output, the prediction error is reduced significantly, but it can only predict one future stock prices every time. After determining the optimal model parameters through empirical research, we extend the model to consecutive predicting,and the continuously forecast accuracy in short term turns out to be acceptable. At last, we use the fitting mode combined with time series theory, the prediction error presents smaller. By selecting the appropriate number of neurons in the input layer, the error can be controlled in a small range, which indicates this fitting mode is the most excellent in the univariate forecasting model. So we determine the best fitting mode and establish a single-variable dynamic forecasting model,then added in the decision function and state function to make investment decision, the strategy yields is apparently higher than the rate of return of one single stocks or index,which indicates we have achieved excess returns.We extend the univariate model to the multivariate model, add opening price, highest price and the lowest price as new variables, and divided into multiple output model and single output model depending on the different output ways. In multi-output model, we found that when using opening and closing prices of two variables, the prediction error is the smallest, While in the single-output model, using four variables makes the model prediction accuracy higher. Both models are able to achieve better prediction results, and both has its advantages and disadvantages. Then extend the two models to dynamic forecasting model and combined with the investm.ent strategy, they have achieved excess returns, which means the models are effective.