Comparison of SARFIMA and LSTM methods to model and to forecast Canadian temperature

An empirical study is performed using seasonal autoregressive fractionally integrated moving average (SARFIMA) time series models and long short-term memory (LSTM) methods from deep learning to evaluate their statistical performance for the fit and prediction of temperature variables. These two methods are applied to monthly minimum and maximum air temperature data collected from four Canadian stations, covering the eastern, western, and central regions of Canada. It is a generally accepted fact that accurate temperature forecasting is important to everyone and stakeholders in several economic activities including the tourism, agriculture, and energy sectors rely on this information. Therefore, it is crucial to select the best methods which provide appropriate statistical modelling and an accurate prediction of temperature. The results of this study show that deep learning LSTM models have better fit and smaller root-mean-square errors compared to time series SARFIMA models.