SURE-Autometrics algorithm for model selection in multiple equations
The ambiguous process of model building can be explained by expert modellers due to their tacit knowledge acquired through research experiences. Meanwhile, practitioners who are usually non-experts and lack of statistical knowledge will face difficulties during the modelling process. Hence, algorit...
محفوظ في:
| المؤلف الرئيسي: | |
|---|---|
| التنسيق: | أطروحة |
| اللغة: | eng eng |
| منشور في: |
2016
|
| الموضوعات: | |
| الوصول للمادة أونلاين: | https://etd.uum.edu.my/6060/1/s92279_01.pdf https://etd.uum.edu.my/6060/2/s92279_02.pdf |
| الوسوم: |
إضافة وسم
لا توجد وسوم, كن أول من يضع وسما على هذه التسجيلة!
|
| الملخص: | The ambiguous process of model building can be explained by expert modellers due to their tacit knowledge acquired through research experiences. Meanwhile, practitioners who are usually non-experts and lack of statistical knowledge will face
difficulties during the modelling process. Hence, algorithm with a step by step guidance is beneficial in model building, testing and selection. However, most model selection algorithms such as Autometrics only concentrate on single equation modelling which has limited application. Thus, this study aims to develop an algorithm for model selection in multiple equations focusing on seemingly unrelated regression equations (SURE) model. The algorithm is developed by integrating the SURE model with the Autometrics search strategy; hence, it is named as SURE-Autometrics.
Its performance is assessed using Monte Carlo simulation experiments based on five specification models, three strengths of correlation disturbances and two sample sizes. Two sets of general unrestricted models (GUMS) are then formulated
by adding a number of irrelevant variables to the specification models. The performance is measured by the percentages of SURE-Autometrics algorithm that are able to eliminate the irrelevant variables from the initial GUMS of two, four and six
equations. The SURE-Autometrics is also validated using two sets of real data by comparing the forecast error measures with five model selection algorithms and three non-algorithm procedures. The findings from simulation experiments suggested that
SURE-Autometrics performed well when the number of equations and number of relevant variables in the true specification model were minimal. Its application on real data indicated that several models are able to forecast accurately if the data has no quality problem. This automatic model selection algorithm is better than non-algorithm
procedure which requires knowledge and extra time. In conclusion, the performance of model selection in multiple equations using SURE-Autometrics is
dependent upon data quality and complexities of the SURE model. |
|---|