Financial Data Forecasting Using Optimized Echo State Network

Abstract

The echo state network (ESN) is a dynamic neural network, which simplifies the training process in the conventional neural network. Due to its powerful non-linear computing ability, it has been applied to predict the time series. However, the parameters of the ESN need to be set experimentally, which can lead to instable performance and there is space to further improve its performance. In order to address this challenge, an improved fruit fly optimization algorithm (IFOA) is proposed in this work to optimize four key parameters of the ESN. Compared to the original fruit fly optimization algorithm (FOA), the proposed IFOA improves the optimization efficiency, where two novel particles are proposed in the fruit flies swarm, and the search process of the swarm is transformed from two-dimensional to three-dimensional space. The proposed approach is applied to financial data sets. Experimental results show that the proposed FOA-ESN and IFOA-ESN models are more effective (~50% improvement) than others, and the IFOA-ESN can obtain the best prediction accuracy.