An Ensemble for Automatic Time Series Forecasting With K-Nearest Neighbors

In this paper a novel approach for automatically configuring a k-nearest neighbors regressor for univariate time series forecasting is presented. The approach uses an ensemble consisting of several k-nearest neighbors models with different configurations for their hyperparameters and model selection...

Descripción completa

Detalles Bibliográficos
Autores: Frías, María Pilar, Martínez-del-Río, Francisco
Tipo de recurso: artículo
Estado:Versión publicada
Fecha de publicación:2025
País:España
Institución:Universidad de Jaén
Repositorio:RUJA. Repositorio Institucional de la Producción Científica de la Universidad de Jaén
OAI Identifier:oai:ruja.ujaen.es:10953/6145
Acceso en línea:https://hdl.handle.net/10953/6145
Access Level:acceso abierto
Palabra clave:Trending time series
Univariate time series forecasting
Model combination
004
311
Descripción
Sumario:In this paper a novel approach for automatically configuring a k-nearest neighbors regressor for univariate time series forecasting is presented. The approach uses an ensemble consisting of several k-nearest neighbors models with different configurations for their hyperparameters and model selection choices. One advantage of this scheme is that the uncertainty associated with choosing a wrong configuration for the model is reduced. This approach is compared with the classical way of selecting a configuration by doing a grid search among several configurations of hyperparameters and model selection choices and choosing the one that performs best on a validation set. The experimental results, using datasets from time series forecasting competitions, show that, in line with previous works, the use of an ensemble produces a robust model, outperforming the approach that uses a grid search for obtaining the best configuration on a validation set and almost any specific configuration. The forecast accuracy of the ensemble is similar to state-of-theart models. Furthermore, this paper also tests the effectiveness of some recent approaches for dealing with trending time series when using the k-nearest neighbors algorithm.