Comparative Study of Wavelet-SARIMA and EMD-SARIMA for Forecasting Daily Temperature Series

This paper aims to find a forecasting model based on the comparative study of wavelet- ARIMA and EMD-ARIMA models and residual-based model selection technique for temperature time series. Time series analysis is essential in studying temperature data for investigating the variation and predicting the future trend, in which we can control the changes and make good decisions. And most important is to understand the trend in the series with time. This study applied hybridized models of wavelet transform and empirical mode decomposition with seasonal autoregressive integrated moving average (SARIMA), which combines two models to get better accuracy, for forecasting daily average temperature time series data in the central region of Eritrea, Asmara. Daily data was collected for 30 years, from January 1, 1991, to December 31, 2020. The study compares WT-SARIMA and EMD-SARIMA models to find a well fit and better forecasting model. Model selection techniques are essential for time series analysis to determine which model best fits our data. AIC and BIC are the most used methods in model selection. This paper uses an additional method based on the residual series. In estimating accurate parameters, the structure of the residual sequence had a lot of connection, in which a stationary residual depict an accurate estimation. From this perspective, a nonstationarity measurement of the residual series is used for model selection. The relative performance is based on the predictive capability of sample forecasts assessed. The results indicate that the hybridized wavelet-SARIMA model is more effective than the other models, and MATLAB soft-wire is used for this analysis.