A comparison of accuracy of forecasting models: A study on selected foreign exchange rates

Abstract

Exchange rate, an economic indicator of the country is the relative price of one country's currency in terms of another country's currency. Stability in exchange rate is important for stable economic growth. Exchange rates, a financial time series highly fluctuate and are chaotic in nature. Forecasting exchange rate fluctuations is very important to countries' economy. Many researchers reported that the three classes of models, namely stochastic models, Artificial Neural Network (ANN)models and Support Vector Regression(SVR) models provided good forecasts. The aim of this research was to compare the forecasting accuracy of most widely used classes of models and to identify better model for forecasting daily exchange rates of Sri Lankan Rupees to Euro and Yen. Daily time series data collected from 2nd July, 2012 to 31st August, 2016 (1008 trading days) from the official website of Central Bank of Sri Lanka were analysed using Eviews, MATLAB and R packages. Stochastic models fitted were found to be inefficient in explaining the variations of daily exchange rates. A Nonlinear autoregressive neural network (NAR) model using Scaled Conjugate Gradient (SCG) learning algorithm and aSVR model with Gaussian radial basis kernel function were designed to the exchange rate returns. Mean Squares Errors and directional accuracy measures revealed that both the machine learning models, ANN and SVR models explained the variation in the series well. However, SVR models provided a better directional accuracy than ANN models. These findings could be useful for domestic as well as foreign investors. Further the forecasting ability can be improved by evolutionary neural networks.