Empirical comparison of the performance of structural time series methods in forecasting daily temperature: case of Johannesburg, South Africa.

Abstract

TheAimofthisstudywastocomparetheforecastingaccuracyofexponentialsmoothing methods and unobserved component methods in forecasting Johannesburg daily temperature. The other objective of this study was to assess the effect of aggregating daily temperature to monthly temperature has on forecasting accuracy of the two structural time series methods. The Johannesburg daily temperature time series used spanned from 01 March 2007 to 31 March 2017. An extension of Holt-Winters model know as TBATS(Trigonometric Fourier representations, Box-Cox transformations, ARMA errors, Trend, and Seasonal component) by De Livera et al. (2011) was found to be more accurate to forecasts Johannesburg daily temperature. This model had high accuracy in Mean Absolute Error (MAE), Root Mean Square Error (RMSE) and Mean Absolute Percentage Error (MAPE). Thestudyalsofoundthatthetwostructuraltimeseriesmethodsweresensitivetotime series classical components present in the data. The monthly temperature data was much smoother than daily temperature data. The two structural time series models used were much accurate in identifying the classical components of a smoother time series. This resulted in much accurate forecast. Holt-Winters additive seasonal model was found to be more accurate to forecast Johannesburg monthly temperature. This model out-performed local linear trend plus Fourier seasonal unobserved component model with high accuracy in Mean Absolute Error (MAE), Mean Percentage Error (MPE) and Mean Absolute Percentage Error (MAPE).