Impact of Seasonal Level Shift (SLS) on Time Series Forecasting

The effect of not treating Seasonal Level Shift (SLS) outliers on forecast accuracy, and prediction intervals is the focus of this study. We examine the impact of SLS on point and interval forecasts using simulation experiment for time series models including SAR (1) and SMA (1) for different parameter values, sample sizes and time of occurrences. We extend the strategy suggested by Asghar and Urooj (2017) to forecasting in the presence of SLS by looking at forecast accuracy and prediction interval. We demonstrate that SLS significantly increases the inaccuracy of the SARIMA models, increases the bias in the SARIMA estimates, and significantly affects the prediction intervals. However, after detection and adjustment of SLS, SARIMA estimates become less biased, and forecast accuracy measure and prediction interval significantly improve. The difference of location of SLS from forecast origin has similar effect on bias and forecast accuracy in SAR (1) model. While, in SMA (1) model, the SLS occurring at the beginning of the series has greater adverse effect than that occurring at the middle or end of the series. Three monthly time series data from Pakistan are used to explore the issue.