Forecasting business and consumer surveys indicators–a time-series models competition

Abstract

The objective of this article is to compare different time-series methods for the short-run forecasting of Business and Consumer Survey Indicators. We consider all available data taken from the Business and Consumer Survey Indicators for the Euro area between 1985 and 2002. The main results of the forecast competition are offered not only for raw data but we also consider the effects of seasonality and removing outliers on forecast accuracy. In most cases, the univariate autoregressions were not outperformed by the other methods. As for the effect of seasonal adjustment methods and the use of data from which outliers have been removed, we obtain that the use of raw data has little effect on forecast accuracy. The forecasting performance of qualitative indicators is important since enlarging the observed time series of these indicators with forecast intervals may help in interpreting and assessing the implications of the current situation and can be used as an input in quantitative forecast models.

Acknowledgements

We acknowledge financial support from ECFIN/2002/A3-01. We would like to thank the European Commission for their helpful comments and support. The usual disclaimer applies.

Notes

¹ More details on the data set can be found in Table 1. Although the number of observations in any empirical analysis should be ideally as large as possible, the type of analysis performed in this article is considered to be valid over 50 observations.

² We provide the AIC since it is one of the criteria most widely used when comparing time-series models. However, the sensitivity of the results has been analyzed so as to check whether the results differ significantly if other criteria were used. We have also obtained the Hannan–Quinn criterion as well as the Schwartz criterion (or Bayesian information criterion) for all the models and a wide range of variables. The selected model is the same in all the cases, confirming the robustness of our results.

³ The univariate ARIMA models are usually used in the forecasting literature for comparative purposes when other forecasting methods are analyzed (e.g.Debenedictis, 1997;Feng and Liu, 2003).

⁴ We have assumed 1 year in order to minimize the sum of squared errors since variables in Business and Consumer Surveys only ask for agents’ perceptions and expectations of their environment in some next months or as a maximum of 1 year ahead. Thus, since the short term is the one being analyzed in these surveys, this is the one that is taken in this article when comparing the different time-series methods.

⁵ The Hamilton filter is an iterative procedure which provides estimates of the probability that a given state is prevailing at each point in time given its previous history. These estimates are dependent upon the parameter values given to the filter. Running the filter through the entire sample provides a log likelihood value for the particular set of estimates used. This filter is then repeated to optimize the log likelihood of obtaining the MLE estimates of the parameters. With the maximum likelihood parameters, the probability of state 0 at each point in time is calculated and these are the probabilities of recession and expansion.

⁶ An alternative approach would have consisted in imposing the value of P and k instead of estimating them. These models are known as Markov Switching Autoregressive Models (MS-AR) and, in general, the values of P are 0.7 or 0.8 and the values of k, 0 or 1.

⁷ It is also interesting to note that all the tests only coincide in 10 series.

⁸ The seasonality analysis was conducted for variables v1, v2, v3p, v3e, v3m, v3b, v4p, v4e, v4m and v4b.