The Application Of Recurrent Neural Networks In Forecasting The Stock Price Index

Time series forecasting of financial asset price is an important area of researchin economics. Traditional quantitative methods have dominated the sector offorecasting research and practice. Although the random walk theory claims that pricechanges is continuous and independent, some scholars have observed that market isnot always efficient. Change in the market price is not random but partiallypredictable. Price index time series is nonlinear, non-stationary and chaos, where thestructural changes of the price index is affected by many factors. For classicaleconometric methods to predict changes in the price index, due to the need of manyrestrictive conditions and unrealistic assumptions, and gradually becomes impossible.Thus, by using intelligent methods in applied mathematics and recent achievements,people try to use artificial neural networks to solve these problems.BP neural network is very popular in the prediction of time series, particularlypopular in the stock price forecasting applications, due to the structure relativelysimple and sophisticated algorithm. However, because it is a static neural network,self-similarity exists in the stock market, and the movement of the stock market priceoften shows a bit long--term memory, making the stock price performance of certaindynamic features. Through the use of dynamic feedback neural network, we candescribe the long--term memory of the time series, and may be able to predictchanges in stock prices better.Therefore, by using the daily closing prices of CSI300Index, we conduct anormality and correlation test, and introduce independent and identically distributed BDS test, while calculating a sequence of values by using Hurst rescaled rangemethod. To discover its dynamic characteristics, we introduce a dynamic feedbackneural network named Elman neural network to predict, and use the relativeprediction error to evaluate the model.