USING DATA PROCESSING ALGORITHMS AND NEURAL NETWORKS TO FORECAST ONE-MONTH PRICE MOVES OF THE SP 500 INDEX

摘要

This paper presents the use of data processing algorithms to modify the input and output data sets that are supplied to a set of neural network models that are used to predict one-month price changes for the S&P 500 Index. Each time a new neural network is initialized it produces different weights, which in turn leads to different simulated outputs for the model. This paper attempts to deal with this problem by summing the absolute magnitude of the forecasted values. These values can then be sorted so that it is possible to determine the paths that represent the 25th and 75~* percentiles for the absolute magnitude values. The model then takes the average of the 25th and 75th percentiles to determine for each particular time step the neural network output, which for this model is the monthly change in the S&P 500 Index value. For a benchmark analysis, trading using the forecasting results of the neural network model with the new data processing algorithms is compared against trading when using a buy-and-hold investment strategy.