Improving Forecast Accuracy with Combination of Forecasts under Structural Breaks and ARCH Innovations

Firuzan E., Aser D. A.

Ekoist journal of econometrics and statistics (Online), vol.37, pp.1-26, 2022 (Peer-Reviewed Journal)

Abstract

An accurate forecast about the future is vital in time series analysis, but it is always challenging to accurately model complex structures in the data. Two major sources of complexity are structural breaks and ARCH effects in the data, which affect the quality of the data and hence reduce forecast accuracy. In this regard, combining the forecasts has been a helpful strategy to improve forecast accuracy for more than five decades since the original paper of Bates and Granger in 1969. Hence, this paper aims to examine if the gains of combined forecasts are sustained in the cases of structural breaks and ARCH innovations. Moreover, we explored which forecast combination schemes are optimal in those cases by combining the forecasts of Exponential Smoothing (ETS), ARIMA and Artificial Neural Networks (ANN) models in simple and regression-based combination procedures. These methods are implemented in both simulated series and empirical data of two popular Turkish stocks namely BIST30 and BIST100. The study found that regression-based forecast combination methods are significantly improving forecast accuracy in the cases of structural breaks and conditional heteroscedasticity. Dynamically weighted combinations show a better improvement in accuracy than their static counterparts if the data contain a trend. Simple combination schemes, including simple average, just perform better than single methods of ETS and ARIMA but hardly beat ANN. Lastly, ANN is the best-performing individual forecasting method in all cases and designs.