Financial forecasting using the neural networks

The Efficient Market Hypothesis (EMH) says that there is no better forecast of stock price possible. If one believes in the EMH, then the stock price movements should follow a random walk. Stock price changes should be random and unpredictable. Despite this hypothesis, we decided to build three back-propagation neural networks and three recurrent neural networks to forecast the daily closing price of stock indexes (S&P500, Dow Jones, and NASDAQ). Our experiments showed that the price is predictable and much better than the random guess. Using a simple shortterm investment strategy, a good annual profit rate can be obtained. Different activation functions and different data preprocessing techniques were tested in order to dynamically determine the best neural network topology. We explored more than six hundred network structures for each neural network in our experiments. The same data sets were also analyzed by statistical models using a commercial statistics package. We find that the parameters of the statistical models can help us in determining the recurrent neural network structure. Comparing the forecasting results, our recurrent neural networks outperform the statistical models in terms of the annual profit rates.