• The Korean Journal of Applied Statistics
• /
• v.30 no.1
• /
• pp.181-193
• /
• 2017

The paper introduces bottom-up and optimal combination methods that can analyze and forecast hierarchical time series. These methods allow forecasts at lower levels to be summed consistently to upper levels without any ad-hoc adjustment. They can also potentially improve forecast performance in comparison to independent forecasts. We forecast regional traffic accident counts as time series data in order to identify efficiency gains from hierarchical forecasting. We observe that bottom-up or optimal combination methods are superior to independent methods in terms of forecast accuracy.

• The Korean Journal of Applied Statistics
• /
• v.31 no.2
• /
• pp.229-239
• /
• 2018

This paper introduces how to adopt the concept of temporal hierarchies to forecast time series data. Similarly as in hierarchical cross-sectional data, temporal hierarchies can be constructed for any time series data by means of non-overlapping temporal aggregation. Reconciliation forecasts with temporal hierarchies result in more accurate and robust forecasts when compared with the independent base and bottom-up forecasts. As an empirical example, we forecast traffic accident counts with temporal hierarchies and observe that reconciliation forecasts are superior to the base and bottom-up forecasts in terms of forecast accuracy.

• The Korean Journal of Applied Statistics
• /
• v.35 no.4
• /
• pp.553-568
• /
• 2022

The Household Income and Expenditure Survey is a representative survey of Statistics Korea, which aims to measure and analyze national income and consumption levels and their changes by understanding the current state of household balances. Recently, the disconnection problem in these time series caused by the large-scale reorganization of the survey methods in 2017 and 2019 has become an issue. In this study, we model the characteristics of the time series in the Household Income and Expenditure Survey up to 2016, and use the modeling to compute forecasts for linking the expenditures in 2017 and 2018. In order to evenly reflect the characteristics across all expenditure item series and to reduce the impact of a specific forecast model, we synthesize a total of 8 models such as regression models, time series models, and machine learning techniques. In particular, the noteworthy aspect of this study is that it improves the forecast by using the optimal combination technique that can exactly reflect the hierarchical structure of the Household Income and Expenditure Survey without loss of information as in the top-down or bottom-up methods. As a result of applying the proposed method to forecast expenditure series from 2017 to 2019, it contributed to the recovery of time series linkage and improved the forecast. In addition, it was confirmed that the hierarchical time series forecasts by the optimal combination method make linkage results closer to the actual survey series.

• The Korean Journal of Applied Statistics
• /
• v.37 no.4
• /
• pp.499 -508
• /
• 2024

In this paper, we investigate hierarchical time series forecasting that adhere to a hierarchical structure when deriving predicted values by analyzing segmented data as well as aggregated datasets. The occurrences of food poisoning by a specific pathogen are analyzed using zero-inflated Poisson regression models and negative binomial regression models. The occurrences of major, miscellaneous, and overall food poisoning are analyzed using Poisson regression models and negative binomial regression models. For hierarchical time series forecasting, the MinT estimation proposed by Wickramasuriya et al. (2019) is employed. Negative predicted values resulting from hierarchical adjustments are adjusted to zero, and weights are multiplied to the remaining lowest-level variables to satisfy the hierarchical structure. Empirical analysis revealed that there is little difference between hierarchical and non-hierarchical adjustments in predictions based on pathogens. However, hierarchical adjustments generally yield superior results for predictions concerning major, miscellaneous, and overall occurrences. Without hierarchical adjustment, instances may occur where the predicted frequencies of the lowest-level variables exceed that of major or miscellaneous occurrences. However, the proposed method enables the acquisition of predictions that adhere to the hierarchical structure.