Multi-Step-Ahead Exchange Rate Forecasting For South Asian Countries Using Multi-Verse Optimized Multiplicative Functional Link Neural Networks

摘要

The dynamic nonlinearity approach, coupled with the exchange rate data series, makes its future predictions difficult. Sophisticated methods are highly desired for effective prediction of such data. Artificial neural networks (ANNs) have shown their ability to model and predict such data. This article presents a multi-verse optimizer (MVO) based multiplicative functional link neural network (MV-MFLN) model to forecast the exchange rate data. Functional link neural network (FLN) makes use of functional expansion for input data with a fewer number of adjustable neuron weights, which makes it capable of learning the uncertainties accompanying the exchange rate data. In contrast to the summation unit at the output layer of FLN, the proposed model uses a multiplicative unit to enhance the ability to learn the complex correlations within the input data. The MVO is employed to fine-tune the parameters of the MFLN. We validate the MV-MFLN on multi-step-ahead forecasting of six exchange rate series through the mean absolute percentage of error (MAPE) metrics. A comparative study with additional forecasts such as genetic algorithm based MFLN (GA-MFLN), differential evolution based MFLN (DE-MFLN), teaching-learning based optimization trained MFLN (TLB-MFLN), and gradient descent based MFLN (GD-MFLN) developed similarly is carried out. It is found that the proposed forecast produces the lowest MAPE values and quite capable of capturing the uncertainties associated with exchange rate data. Observations from comparative performance analysis suggest the superiority of the MV-MFLN-based forecast.