| Summary: | Finding a good model can be a hefty task, especially when there are many predictors, thus providing many possible interactions. Effects and interactions in the model need to be looked into too. Therefore, model selection is one way to make this task simpler. Different strategies of selecting the right model had been proposed throughout the years. In this study, 13 selection procedures are compared in terms of their forecasting performances based on root mean square error (RMSE) and geometric root mean square error (GRMSE). Water quality index (WQI) data of a river in Malaysia has been analysed for two-equation and four-equation models of seemingly unrelated regression equations (SURE) model. The procedures were conducted either through manual or automated selection with ordinary least squares (OLS), feasible general least squares (FGLS) or maximum likelihood estimation (MLE) method for the final model. All automated manner procedures showed favourable results over manual selections. This proves that one person’s knowledge only may not be sufficient to choose the best model. Out of the 13 procedures, SUREMLE-Autometrics has outperformed for both two- and fourequation models with achievement at rank 1 or 2 only. Therefore, MLE is considered as the best estimation method in this model setting
|
|---|