Evaluation of bottom-up and top-down strategies for aggregated forecasts: state space models and arima applications

Abstract. In this research, we consider monthly series from the M4 competition to study the relative performance of top-down and bottom-up strategies by means of implementing forecast automation of state space and ARIMA models. For the bottomup strategy, the forecast for each series is developed individually and then these are combined to produce a cumulative forecast of the aggregated series. For the top-down strategy, the series or components values are first combined and then a single forecast is determined for the aggregated series. Based on our implementation, state space models showed a higher forecast performance when a top-down strategy is applied. ARIMA models had a higher forecast performance for the bottom-up strategy. For state space models the top-down strategy reduced the overall error significantly. ARIMA models showed to be more accurate when forecasts are first determined individually. As part of the development we also proposed an approach to improve the forecasting procedure of aggregation strategies.

• Autores: Soto-Ferrari, Milton; Chams-Anturi, Odette; Escorcia-Caballero, Juan P.; Hussain, Namra; Khan, Muhammad
• Tipo de recurso: http://purl.org/coar/resource_type/c_816b
• Fecha de publicación: 2019
• Institución: Corporación Universidad de la Costa
• Repositorio: REDICUC - Repositorio CUC
• Idioma: eng
• OAI Identifier: oai:repositorio.cuc.edu.co:11323/5809
• Acceso en línea: https://hdl.handle.net/11323/5809; https://repositorio.cuc.edu.co/
• Palabra clave: Top-down; Bottom-up; Forecast automation; Forecast performance; State space models; ARIMA
• Rights: openAccess
• License: CC0 1.0 Universal