This is What Leading Indicators Lead

The purpose of this paper is two-fold. First, we compare the accuracy of previous studies that analyze the ability of the Composite Index of Leading Indicators (CLI) for predicting turning points. Alternative filters are also proposed. For these comparisons, we adapt the test developed by Diebold and Mariano (1995) to the business cycles framework. Second, we combine different approaches to produce a filter that transforms the monthly CLI growth figures into a more intuitive measure of the probability of recession. We examine the predictive power of the CLI for movements in GDP. For the first objective, we analyze the accuracy of the following models: First, we generalize the analysis of Hamilton and Perez-Quiros (1996) describing how linear univariate and bivariate models can be used to forecast nonlinear phenomena such as turning points. We update their study of multivariate Markov switching models. Second, we extend the Smooth Transition Regression methodology to a VAR context. We identify the transition function as the filter that shows the probability of locating the economy between the different states. Third, we analyze an expansion of the probit model suggested in Estrella and Mishkin (1998). Finally, we propose a new methodology based upon adaptive kernel estimation for predicting recessions nonparametrically. Despite the good in-sample performance of the switching regimes model, we conclude that a simple linear univariate model for GDP is more accurate than any bivariate specification in real-time. For the second objective, we suggest that a combination of the forecasts may exploit more leading information from the CLI than any of the individual forecasting models. Combining forecasts of growth, we apply the rule proposed by Granger and Ramanathan (1984). Combining forecasts of recessions, we use a method in the spirit of Li and Dorfman (1996). We prove that a combination of the switching regimes (the best within recessions) and the nonparametic (the best within expansions) is as good as a combination of all the models. The out-of-sample results indicate that the real-time combination presents the most accurate statistical forecast of both GDP growth and recessions. Thus, we conclude that the CLI is useful in anticipating both turning points and output growth. In addition, in contrast to Hess and Iwata (1997), we find that nonlinear specifications perform better than simpler linear models at reproducing the business cycles features of real GDP. An illustration of the operation of this filter shows that the same CLI growth rate contains very different information about the probability of an imminent recession depending on the period considered.