A0671
Title: Modelling COVID and crime in the US as hierarchical time series
Authors: Thomas Fung - Macquarie University (Australia) [presenting]
Joanna Wang - University of Technology Sydney, Australia (Australia)
Abstract: Crime time series data can often be naturally disaggregated by various attributes of interest, either by their crime type or geographical location. When modelling this type of data, the current recommended practice in crime science is to model each series at the most disaggregated level as it helps to identify more subtle changes. However, authorities and stakeholders are often only interested in the big picture, requiring researchers to either simply sum up the fitted value series or model the aggregated series independently. This leads to poor forecasting performance at the higher levels of aggregation in practice, as the most disaggregated series often have a high degree of volatility, while the most aggregated time series is usually smooth and less noisy. Intuition also requires the forecasts to add up in the same way as the data, but one can't guarantee that would be the case when series are modelled independently. The aim is to explain why a previous hierarchical and grouped time series method should be considered as the default technique for modelling this kind of data. US COVID and crime data will be used as an example.