In forecasting, it is common to encounter multiple related time series rather than a single series. These time series often follow a hierarchical structure, where data is organized at various levels of aggregation. The primary goal of Hierarchical Time Series Forecasting (HTS) is to achieve coherence, ensuring that forecasts align across all hierarchical levels, which is crucial for decision-making. This thesis explores the application of HTS for electricity demand in the United States, which consists of a large hierarchical system of 67 seasonal time series with hourly granularity.
The effectiveness of HTS methods can vary significantly based on the specific use case, leading to the question of how intuitive approaches, such as the Bottom-Up method, compare to more sophisticated reconciliation techniques in this context. By evaluating various reconciliation methods alongside classical base forecasts like ARIMA and ETS, this work aims to provide insights into their performance in improving forecast accuracy while ensuring coherence in electricity demand forecasting.