Abstract:
Cyclical asymmetry has been recognized as a non-linear phenomenon in recent studies examining
unemployment rate time series. The probabilistic structure of such time series is different during
economic upswings as compared with economic downswings. So, with forecasting unemployment
rates in mind, it seems intuitive that models should reflect this change in structure by incorporating
non-linearities. This allows for the switching in optimizing behaviour between different phases of
the business cycle. Accordingly, this paper evaluates the point forecasts from models of the monthly,
Australian unemployment rate series, these models being drawn from both the linear and nonlinear
classes. The non-linear model is based on a standard logistic smooth transition autoregressive
(LSTAR) model of unemployment which includes a lagged leoel term and a seasonal, rather than
first-difference transition variable. Forecasts from this model are evaluated against the best-fitting
linear autoregressive (AR) alternative. Dynamic point forecasts over twenty four months, suggest
that the LSTAR forecasts are more accurate than the linear AR alternative. However, there is no
statistical difference between the forecasts from both models on a one-to-twelve step-ahead basis.
