RDP 2000-02: Forecasting Australian Economic Activity Using Leading Indicators 1. Introduction

The economic indicator approach, pioneered by Burns and Mitchell (1946), is based on the idea that business cycles are driven by repetitive sequences and that certain economic variables or combinations of variables can be found which underlie these sequences. These variables and the constructed lagging, coincident and leading indices can be used to confirm, identify and predict the business cycle. Economic indicators such as these, if they perform well, are potentially useful for policy-makers as a complement to standard modelling approaches used to assess current economic circumstances and predict the likely future course of economic activity.

Our concern is the contribution composite leading indicators, or leading indices, can make to forecasts of economic activity in Australia. While policy-makers may frequently augment their model-based forecasts with a subjective weighting of individual economic indicators, formal composite leading indicators are, we suspect, less commonly used. In part, this may reflect uncertainty about what, if any, contribution these indices can make and hence our objective to systematically investigate this question. The reluctance to use leading indices may also reflect a demand for greater understanding of the underlying developments in the individual components of the index, rather than relying on a weighted composite. While obviously a legitimate concern, leading indices that perform well can still play a useful role as a benchmark, providing forecasts based upon a stable summary of a set of economic indicators.^[1]

We consider three leading indices of activity regularly published for the Australian economy: the Westpac-Melbourne Institute (WM) Leading Index, the ABS Experimental Composite Leading Indicator and the NATSTAT published by the National Institute of Economic and Industry Research (NIEIR).^[2] We use simple time series models to examine the leading indices' predictive performance for real output, employment and unemployment. The first step is to examine the within sample performance of these indices. This provides information about the usefulness of each index as a predictor of activity as well as information about the timing of the relationship between the series of interest and the relative importance of the leading index for forecasting activity. A similar exercise is undertaken in Trevor and Donald (1986) for two of the indices considered here (WM and NATSTAT); our results can in part be viewed as an updating of this previous study.

Within sample results, however, provide very limited information about forecasting performance. To better assess the contribution of leading indices, we next consider out of sample forecasting performance. To do this in a transparent and systematic manner, we consider two variable VAR models consisting of a leading index and a measure of activity. We stress, however, that we are not putting these models forward as our preferred forecasting models. They are just being used as a simple and consistent method of assessing the contribution of the leading indices to forecasting activity, analogous to our within sample evaluation.^[3]

Previous findings in the Australian literature are generally favourable: the leading indicator approach is a relatively quick and easy process that produces reasonable results, complementing those from much more expensive and sophisticated models. Many of these studies, however, focus on the ability of leading indicators to predict turning points in the business cycle. For example, Boehm and Moore (1984) and ABS (1997) detail the consistency with which turning points in the WM and ABS leading indices lead turning points in the WM coincident index and real GDP respectively. Discrete dependent variable models have also been used to provide probability assessments of turning points in activity based upon movements in leading indices. For Australian examples, see Layton (1997) and Summers (1997).

The traditional assessment of leading indices, which focuses only on their ability to predict turning points in the business cycle, has two aspects that are troubling.^[4] First, there is a considerable amount of subjective judgment involved; in particular, the definition of the turning points in the business cycle and the criteria for a successful prediction by the leading index (most importantly, the signal and lead provided).^[5] The limited information available from such assessments, however, is of greater concern. An index may generally predict a turning point in activity but convey little or no information about the duration and extent of the contraction or expansion, either prior to or during the event. As policy-makers, concerned with maintaining price and output stability, it is the latter information that is of crucial importance. Consequently, knowledge that a leading index does or does not provide such information is important. Our assessment, based upon simple time series models, is designed to assess the quantitative relationship between activity series and leading indices over the entire cycle for just such reasons.^[6]

Because leading indices are commonly constructed with the focus on turning points in mind, it is important to be aware that we are assessing the general forecasting performance of indices that were not necessarily constructed for this purpose. For example, the National Institute of Economic and Industry Research and ABS state that their leading indices are designed solely to predict turning points in activity (National Institute of Economic and Industrial Research 1993; Salou and Kim 1993). As our focus is broader than the original purpose of these indices, poor forecast performance based upon our assessment should not then be construed as these indices failing to perform as designed; rather, they do not meet the more general criteria we examine.

The paper is organised as follows. Section 2 describes the leading indicators of Australian activity that we evaluate. In Section 3 we present both within sample and out of sample results for each of the three leading indices. We first focus on the contribution leading indices can make to forecasts of activity. We then consider how well simple two variable time series models, using GDP and the WM index, compare to a structural model of GDP due to Gruen and Shuetrim (1994). This serves to put the previous results in some context. Section 4 concludes.

Footnotes

This is not strictly accurate since the components and weights of any leading index can and do change. In principle, however, the leading index should represent a reasonably stable summary of leading indicators. [1]

The OECD also publishes a leading index for Australia, which moves relatively closely with the WM index. After allowing for publication lags, however, the OECD index is not very timely and consequently we do not examine it in this study. [2]

There are obviously a great number of different time series models that one might consider to assess the usefulness of leading indices for forecasting. For example, Summers (1998) uses the WM index in a large Bayesian VAR model. Our approach, while limited, does have the advantage of explicitly focusing attention on the contribution of the leading index. [3]

See Auerbach (1982) for similar arguments in favour of regression-based assessment of leading indices over the entire cycle. [4]

See Harding and Pagan (1999) for a recent discussion concerning the identification of the business cycle. [5]

Spectral analysis is another method of assessing the correspondence of movements in leading indicators and activity throughout the cycle, although it provides within sample information only. Using these methods, Layton (1987) shows that the WM leading index reliably anticipates the WM coincident index throughout the cycle. [6]