The Use of Parsimonious Neural Networks for Forecasting Financial Time Series

When attempting to determine the optimal complexity of a neural network the correct decision is dependent upon obtaining the global solution at each stage of the decision process. Failure to ensure that each optimization being considered is near a global solution may lead to misleading and often conflicting results jeopardizing any decision rule for choosing an optimal complexity. Here, a genetic algorithm is used for global search and, by modifying the objective function, is used to simultaneously select a parsimonious structure. The chosen structure often eliminates all connections to unnecessary variables and thus identifies irrelevant variables. Several models are examined to forecast the five day relative difference of the S&P 500 index. The first series of models uses only past S&P 500 values while the second series of models uses other explanatory variables. Models with the complete architecture are compared to those with the reduced structure. Based on the preliminary model analysis a composite model is constructed. Department of Economics, Conner Hall, University of Mississippi, 38677, (601)232-7575 Department of Management, Ball State University, Muncie, IN, 47306,(317) 285-5320