trabajo final de especialización.page.titleprefix Optimización de modelos predictivos para Series de Tiempo Jerárquicas (HTS)
Loading...
Date
2021-10-21
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
"Las series de tiempo jerárquicas (HTS) son un subconjunto de series de tiempo que están contenidas dentro de una estructura o jerarquía. Esta estructura puede estar dada por características propias del producto/servicio, como su categoría, departamento, etc., o por una división geográfica, como por ejemplo ciudad, estado o país. Las técnicas clásicas como ARIMA o ETS no son óptimas para este tipo de conjunto de series, ya que la estructura puede aportar información valiosa que pasaría inadvertida o al menos no tan fácilmente identificable si modelamos las series de tiempo por separado. Debido a esto, se comparó el desempeño de Redes Neuronales LSTM, las cuales son ampliamente utilizadas para series de tiempo, y de técnicas específicas para series de tiempo jerárquicas como Bottom-Up y Reconciliación Óptima para predecir las ventas de departamentos dentro de tiendas de Walmart en EE.UU. Se analizaron en detalle 70 series de tiempo a nivel tienda-departamento y se demostró la eficacia de estas."
"Hierarchical Time Series (HTS) are a subset of time series that are contained inside of a structure or hierarchy. This kind of structure may be defined by important characteristics of the product/service, like its category or department, or it may be defined by some geographic division, for instance, city, state, country and so on. Classic approaches for time series, like ARIMA or ETS are suboptimal for this kind of series due to the fact that this hierarchy could give us interesting information for forecasting that may be hidden or not easily identifiable if we model all the series separately. Because of that, we compared the performance of LSTM Neural Networks, which are widely used for time series, and some specific approaches for hierarchical time series as Bottom-up and Optimal Reconciliation in order to forecast department sales for Walmart stores in the USA. We performed an in-depth analysis for 70 store-departamental time series and showed the efficacy of these techniques for each one."
"Hierarchical Time Series (HTS) are a subset of time series that are contained inside of a structure or hierarchy. This kind of structure may be defined by important characteristics of the product/service, like its category or department, or it may be defined by some geographic division, for instance, city, state, country and so on. Classic approaches for time series, like ARIMA or ETS are suboptimal for this kind of series due to the fact that this hierarchy could give us interesting information for forecasting that may be hidden or not easily identifiable if we model all the series separately. Because of that, we compared the performance of LSTM Neural Networks, which are widely used for time series, and some specific approaches for hierarchical time series as Bottom-up and Optimal Reconciliation in order to forecast department sales for Walmart stores in the USA. We performed an in-depth analysis for 70 store-departamental time series and showed the efficacy of these techniques for each one."
Description
Keywords
REDES NEURONALES, APRENDIZAJE AUTOMATICO, PREDICCIONES