|
Abstract:
|
Daily time series have increasingly appeared on the radar of official statistics in recent years, mostly as a consequence of the exploration of new digital data sources for information that could be used to augment established forecasting models for headline indicators, such as quarterly GDP. Many of these daily series are seasonal and thus in need for seasonal adjustment. However, traditional methods in official statistics often fail to model and seasonally adjust them appropriately. The main reason is that granular daily data typically exhibit features that are not observable in monthly and quarterly data. Prime examples include irregular spacing, coexistence of multiple seasonal patterns with integer versus non-integer seasonal periods and potential cross-dependencies as well as small sample issues, such as missing data. We provide an overview of recent modeling and seasonal adjustment approaches that are capable of handling these distinctive feature, or at least some of them, and illustrate selected methods using daily realized electricity consumption in Germany. Special attention is paid to the extended X-11 and ARIMA model-based approaches and to structural time series models as implemented in a preliminary version of JDemetra+ 3.0 that is accessible via R.
|
