A Comparative Analysis of Robust Penalized Estimators for Periodic Time Series

Panos Evangelidakis, Yiannis Kamarianakis

Keywords: MTE, LSA, MEB, Robust Variable Selection, Adaptive LASSO, AIC/BIC Minimization

Abstract: This study evaluates a novel **Maximum Entropy Bootstrap-based (MEB, H.Vinod, J.Lopez-de-Lacalle; 2017)** adaptive penalized estimator, against two cutting-edge statistical procedures for feature selection and coefficient estimation, namely **Least Squares Adaptive Lasso Approximation (LSA, H. Wang et al; 2007)** and **Maximum Tangent Likelihood Estimation (MTE, Y. Qin et al; 2017)**. The motivating application is vehicular traffic forecasting: specifically our work extends Rodrigues' (2023) decomposition, which identifies significant seasonal components to construct a stable, baseline online forecasting procedure. The analyzed dataset focuses on measurements from Alexandras Avenue in Athens; it was used in a previous traffic forecasting competition (M.Giacomazzo, Y.Kamarianakis; 2020}. Model performance is evaluated using metrics that are robust to outliers (i.e. MAE, Huber Loss) and popular among practitioners (i.e. MAPE, MdAPE and their symmetric counterparts). MTE outperforms the alternative penalized scheme; specifically Huber Loss is reduced by upto 7\% across the examined measurement locations.

Submission Number: 132