Forecasting Models/ Time Series
Sepideh Etemadi; Mehdi Khashei
Abstract
Purpose: The purpose of this paper is to present a new methodology for statistical modeling, which, unlike all commonly developed models and algorithms, maximizes the reliability of the results instead of the resulting accuracy. Accordingly, a new class of statistical modeling approaches has been developed ...
Read More
Purpose: The purpose of this paper is to present a new methodology for statistical modeling, which, unlike all commonly developed models and algorithms, maximizes the reliability of the results instead of the resulting accuracy. Accordingly, a new class of statistical modeling approaches has been developed by replacing conventional processes with the proposed process.Methodology: The multiple linear regression method has been selected to implement the proposed methodology in this paper. To comprehensively evaluate the performance of the proposed regression model, 10 standard datasets from the literature on statistical modeling have been considered.Findings: Overall, the results show that in 65% of the studied data sets, the proposed model can generalize more than the usual multiple linear regression. The proposed regression model, on average, has been able to improve the accuracy of the modeling by 5.571% and 6.466% in mean absolute error and mean square error, respectively, compared to its classic version. These results clearly show the significant effect of reliability of the results on the degree of generalizability, which is basically not considered in the usual statistical modeling processes.Originality/Value: Statistical modeling is one of the most important tools for simulating real-world systems and data sets that are often used to make decisions in a wide range of applications. Several different approaches have been developed in the literature with different features to cover real-world issues with the desired accuracy. However, such methods follow a similar concept and idea in the modeling process. The performance basis in all conventional statistical modeling approaches is based on the assumption that maximum accuracy in experimental and inaccessible data will be obtained from models with minimization of error in training data. Although this is a logical and standard procedure in traditional statistical modeling spaces, it is not the unique way to achieve maximum generalizability. In other words, the generalizability of the model simultaneously depends on the model's accuracy and the level of results' reliability. In this paper, a new methodology for statistical modeling is presented, which, unlike all commonly developed models and algorithms, maximizes the reliability of the results instead of the resulting accuracy.