Stock Index Forecasting using Recurrent Neural Networks

S.-I. Wu and H. Zheng (USA)


Recurrent Neural Network, Financial Forecasting and Prediction, ARIMA, Time Series, Stock Market


We have developed two different architectures of recurrent neural networks (RNN) to forecast the daily closing prices of stock indexes (S&P500, Dow Jones, and NASDAQ) five days ahead. Experiments showed that the price is predictable and much better than the random guess. According to a simple short-term investment strategy, a good annual profit rate can be obtained. Different activation functions were tested in order to dynamically choose the best neural network topology. We explored more than 600 network structures for each neural network. The same data sets were analyzed by the ARIMA model using commercial statistics package MINITAB. The RNNs outperform a better result than the ARIMA model in terms of profit rates. The autoregressive parameter helped us determine the recurrent neural network structure.

Important Links:

Go Back