Efficient outlier detection for large time series databases

P. Galeano San Miguel, D. Peña, R. S. Tsay

We propose a fast and powerful approach for detecting outliers in individual time series within a large database. The approach is highly flexible and can handle databases with diverse characteristics among the series. The proposed method detects outliers by examining the residuals of observed series after a robust model fitting, employing saturated regression models to consider nearly all observations as potential outliers, and using the Orthogonal Greedy Algorithm to identify outlying effects. The method is automatic and has been implemented to run in parallel in the R package outliers.ts.oga, allowing researchers and practitioners to effectively and efficiently detect outliers in large databases. We demonstrate the efficacy of the proposed procedure by various simulations and by the fast cleaning of the FRED-MD macroeconomic database, a well known dataset often used in macroeconomic analysis.

Palabras clave: Additive outlier, Boosting, Forecasting, Level shift, Orthogonal greedy algorithm, Parallel computing.

Series Temporales
13 de junio de 2025  11:00
Sala VIP Jaume Morera i Galícia

R. Morales Arsenal, C. Bergmeir, L. Escot Mangas

J. A. Martín Segura, C. Pérez López

A. Lazcano, P. Hidalgo, J. E. Sandubete Galán

J. Cara