Year: 2026
Type: article
Abstract: In many situations, a relative (percentage) error is considered a more appropriate loss function than least squares or least absolute deviation. However, when working with zero-censored or semicontinuous data, a dependent variable value of zero makes a relative error impractical. We develop a novel loss function to emulate the relative error that maintains a finite value at zero and does not approach zero for large values. This is then incorporated into a regression model that is solved using a mixed-integer program. Least absolute deviation and least squares are shown to be special cases. Numeric analyses indicate the proposed loss function closely estimates the relative error, and its application as an estimator is demonstrated using datasets from the literature.
Source: Applied Operations and Analytics
Institution: University of Tulsa
Language: English
Subfield: Statistics and Probability
Field: Mathematics
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W7234989496
