Opposition-Based GA Learning of Artificial Neural Networks for Financial Time Series Forecasting

Abstract

Artificial neural network (ANN) based forecasting models have been established their efficiencies with improved accuracies over conventional models. Evolutionary algorithms (EA) are used most frequently by the researchers to train ANN models. Population initialization of EA can affect the convergence rate as well as the quality of optimal solution. Random population initialization of EAs is the most commonly used technique to generate candidate solutions. This paper presents an opposition-based genetic algorithm (OBGA) learning to generate initial candidate solutions for ANN based forecasting models. The present approach is based on the concept that, it is better to begin with some fitter candidate solutions when no a priori information about the solution is available. In this study both GA and OBGA optimizations are used to optimize the parameters of a multilayer perceptron (MLP) separately. The efficiencies of these methods are evaluated on forecasting the daily closing prices of some fast growing stock indices. Extensive simulation studies reveal that OBGA method outperforms other with better accuracies and convergence speed.