Enhancing the predictive performance of non-stationary time series data through various transformations
Young-Eun Jeon, Suk-Bok Kang, Jung In Seo · Journal of the Korean Data and Information Science Society · 2024
Time series datasets observed in the real world frequently have trends and seasonality, and these datasets are called non-stationary time series datasets. For non-stationary time series datasets, a transformation is an indispensable task for enhancing the accuracy of analysis. A Box-Cox transformation is one of the most widely used transformation methods, but it has burden of estimating a power parameter from the observed data. In response to this predicament, this paper provides a scaled logit transformation and substantiates its superiority in comparison with the Box-Cox transformation. A notable advantage of the provided transformation is that it has no stress associated with parameter estimation, unlike the Box-Cox transformation. For illustration purposes, a bike sharing dataset with features related to weather, temporal, and seasonality is used as a case study. The superiority and applicability of the provided transformation are meticulously examined by applying various modeling techniques including statistical and machine learning techniques to this dataset.