Performance Evaluation via Error-based Goodness-of-fit Measures of a Newly Proposed Regression Model in Comparison with Twelve Existing Models

A comprehensive comparative evaluation of twelve existing simple regression
models (Tangential, Sinuoidal, Square Root, Exponential, Power, Hyperbolic,
Logarithmic, Polynomial, Quadratic, Linear, Cosinusoidal, and Cubic Root)
versus a newly hybrid developed regression model using multiple error-based
goodness-of-fit measures (Root Mean Squared Error, Mean Absolute Error,
Mean Absolute Percentage Error, Symmetric Mean Absolute Percentage Error
and Coefficient of Variation of Root Mean Squared Error) was conducted. The
study focuses on scale-independent error behavior and predictive accuracy,
which makes the evaluation more appropriate for modeling practical
applications, unlike numerous past studies which relied much on coefficient of
determination or information criteria measures. A developed model in this study
consistently outperforms the twelve existing models in all chosen goodness-of-fit
measures based on the empirical results from the two different datasets of
programming performance and Physical Health Measurements of final year
Computer Science students used. The non-parametric Friedman test also proves
that the noted differences in the performance of the regression models are
statistically significant, whereas the high value of Kendall coefficient of
concordance shows that there is a high level of agreement between evaluation
metrics. Hence, a combination of these outcomes gives strong statistical data of
the effectiveness of the proposed model.