Forecasting In Stock Market With Neural Networks

Feedforward neural networks can simulate any nonlinear function.Nowadays,it has broaduse in many fields.As one popular learning rule,BP algorithm has figured out the problem ofXOR and T-C matching.But since the training process of the BP network is also a globaloptimization question of a highly nonlinear function,there are still many unsolved issues duringthe training course.For example,BP network suffers from a slow training speed and can easilyget stuck into local minima.So we've analyzed conventional BP algorithm and usual improvingtechnique,found out the reason of those problem. We tried to propose a new algorithm based onformer effort to overcome these, problems in our forecasting work of the stock market.Wemainly altered two aspect of the original algorithnrtransfer function and learning rule.Firstly,weadvanced a new transfer function,and revised the learning rate and momentum accordingly .Thenew transfer function is also monotone and bounded,being a S type function.Besides,the newfunction can avoid the incorrect simulation matter near 0 and 1. Secondly,to solve the efficiencyproblem of the algorithm with fixed learning rate,we proposed a novel global optimizationlearning rule on the basis of the conjugate-gradient method,and analyzed the convergence ofit.After that, we apply the algorithm to the forecasting of the volatility of the option traded at theChicago Board Options Exchange and the integrated index of Shanghai Stock Market.Duringthe application,we compare the results with other existing forecasting technique.From thecomparison,we found the new algorithm has high accuracy and efficiency comparing to othermethods.This paper is organized as follows.Part 1 describes the basic frame and standardalgorithm of BP network,summarizes the existing improving techniques and propose a new fastalgorithm.Part 2 introduces the stock option where we'll apply our new method and analyzesthe importance of volatility in economic field and describes the popular computation andforecasting means of it. Part 3 apply the new algorithm to the foresting of the volatility of theoption in CBOE and index of Shanghai stock.During the forecasting,we got a high valuation tothe new algorithm.Finally,we concluded the new algorithm could accelerate training speedgreatly.settled the local minimum problem,offered a novel way to analyze the neural network.