Impact of Structural Break Location on Forecasting Accuracy: Traditional Methods Versus Artificial Neural Network

Abstract

Since forecasting future values is fundamental for researchers, investors, practitioners, etc., obtaining accurate predictions is critical in time series analysis. The accuracy is reliant on good modelling and good-quality data. The latter is affected by unusual observations, changes over time, missing data, and structural breaks among others. Economic crises are the major cause of data instability and therefore, this paper focuses on how structural breaks in conditional heteroscedastic financial and macroeconomic data affect forecasting accuracy on short and long-term horizons. More specifically, we are interested in the impact of the location of the structural break and break size on the predictive performance of two linear (ARIMA and Exponential Smoothing) forecasting models and two nonlinear (ARIMA – ARCH and Artificial Neural Network) models. We conducted Monte Carlo simulations and showed that the forecasting accuracy decreases as the structural break location approaches the end of the sample. In addition, break size and length of the horizon show the same impact on the forecasting accuracy as the forecasting error increases with the increase of break magnitude and length of the horizon. We also showed that ARIMA – ARCH model is the best performing in the absence of a structural break while the artificial neural network model outperforms all the competing models in the presence of structural break, especially in large break sizes and long horizons. Last, we applied the above techniques to forecasting daily close prices of brent oil and Turkish Lira – USD exchange rates out–of–sample, and similar results were found.