|
Abstract:
|
Seasonal time series can be separated into seasonal and nonseasonal components. The former accounts for differing variability across seasons. The latter covers more of the underlying dynamics of the series; it can be further separated into trend, cycle, and irregular components (with trend and cycle sometimes combined). Model-based procedures like SEATS use signal extraction with the canonical decomposition of ARIMA models to perform seasonal adjustment, while the X11 procedure uses a series of moving averages. In this paper, series of airline models are constructed by simulating each of the individual canonical ARIMA components and then summing. Using the correct model with known parameters, SEATS will yield seasonal adjustments with minimum mean squared error. We perform seasonal adjustments of the simulated series with the SEATS and X11 procedures to evaluate their mean squared errors under various alternative settings. In particular, relative to the optimal SEATS adjustment, how much worse does one do with a SEATS seasonal adjustment when one has to choose a model and estimate its parameters, and with an X11 adjustment when the seasonal filter is automatically chosen?
|
