Timely detection of turning points: Should I use the seasonally adjusted or trend estimates?

Timely and accurate detection of turning points is an important issue in analysing time series data. Different time series estimates, such as the original estimates and the derived seasonally adjusted and trend-cycle estimates, are available to help assess turning points. This paper focuses on detection of turning points from time series estimates derived using a univariate approach. We investigate the difference between using seasonally adjusted and trend estimates for timely detection of time points. The Australian Bureau of Statistics (ABS) regularly publishes original, seasonally adjusted and trend-cycle (referred as short-term trend or trend hereafter) estimates to enable users a choice of complimentary time series estimates. Users may choose to use any, or all of, the time series estimates as provided or as input into sophisticated modelling approaches which can then assist with informed judgement, decision and policy making. The ABS recommends the use of trend estimates to provide the most appropriate estimate of the underlying direction of the original time series (Linacre and Zarb, 1991). Knowles (1997) and Compton (2000) surveyed trend estimation practice of a range of National Statistical Institutes and found that timely detection of turning points and minimisation of the number of false turning points were desirable characteristics of short-term trends. Knowles and Kenny (1997) considered issues with turning points for trend estimates. If a turning point is incorrectly identified or not identified soon enough, it may lead to inaccurate assessments of economic activity, which may in turn impact on important economic decisions. Seasonally adjusted estimates for official government statistics are typically derived using a univariate approach for individual time series. Alternative multivariate approaches which use relationships between time series can improve the detection of turning points (for example, see Zhang and McLaren, 2005). This paper focuses on detection of turning points from time series estimates derived using a univariate approach. We consider issues in detecting turning points for monthly time series in using either the published seasonally adjusted and trend estimates available from a filter based seasonal adjustment process. The trend estimate is often perceived to be a signal extraction or data transformation process and the use of filters, both for the seasonal adjustment and trending process, are known to introduce distortion in the final estimates.