Prediction Accuracy of SARIMA-STAR-CNE

With respect to non-linearity that appears in most financial and macroeconomic time series data, linear models have been discovered to have a deficiency of modeling and foresting such data. To astound this delinquent, non–linear models were developed and used in the forecasting framework. So, the objective of this article is to develop a new hybrid model that compares the predictive ability of the STAR model in combination with (S)ARIMA and clustering-based nonlinear ensemble (CNE) estimates. The use of SARIMA-STAR-CNE forecasting coupled with a self-organizing map (SOM) and kernel-based extreme learning machine (KELM) improved the forecasting accuracy and prediction performance as a clustering scheme via SOM neural networks aims at addressing the issue of fixed weights. We then adopt an SOM network to determine the number of clusters of various component forecast data. On average, our model gave a prediction power that is high above 80% in both in-sample and out of sample forecasting for the Brazil, Russia, India, China and South Africa (BRICS) exchange rates. The results of this study shall be useful to policymakers and the financial industry for future use, planning, and forecasting of exchange rates. In addition, a foundation is provided for imminent scholars to conduct studies on un-industrialized markets. These results are also important for risk managers and investors.