Towards Designing and Performance Analysis of Evolving Higher Order Neural Networks for Modeling and Forecasting Exchange Rate Time Series Data

Abstract

Achieving improved prediction accuracy with minimal input data and computationally less complex model is a challenge in financial time series forecasting research. Constructing the model from training data and evaluate it on test data is a common methodology which requires lots of human interventions. This paper developed three evolving higher order neural networks (EHONN) and evaluated their performances on modeling and forecasting five exchange rate time series. A Pi-Sigma neural network (PSNN) is used as base model. The optimal architectures of PSNN are evolved on fly with three evolutionary learning methods, i.e. genetic algorithm (GA), differential evolution (DE), and particle swarm optimization (PSO), therefore forming three hybrid models. The single layer tunable weight and biases of PSNN contributed less complexity to the hybrid models. The models are evaluated on five real exchange rate datasets, compared with other state-of-art models trained similarly and found better. Further, statistical test are conducted to justify the significance of the proposed models.