Survey data and short-term forecasts of Swedish GDP growth

Abstract

In this article, we evaluate forecasting models for Swedish GDP growth which make use of data from Sweden’s most important business survey, the Economic Tendency Survey. Employing nine years of quarterly real-time data, we conduct an out-of-sample forecast exercise. Results indicate that the survey data have informational value that can be used to improve forecasts, thereby confirming the empirical relevance of survey data for GDP forecasters.

Keywords:

• out-of-sample forecasts
• real-time data

JEL Classification:

• E22
• E27

Notes

¹ Additional studies on the relationship between survey data and the real economy include Carroll et al. (1994), Ludvigson (2004), Dreger and Schumacher (2005), Mitchell et al. (2005), Kwan and Cotsomitis (2006) and Siliverstovs (2013). Alternatively, one could consider using financial data to forecast the real economy; see, for example, Mody and Taylor (2003).

² The questionnaires employed in the survey can be found at http://www.konj.se/1666.html. These show exactly how the question underlying each variable employed in the empirical analysis in this article is phrased.

³ As an alternative, one could consider working with mixed frequencies; see, for example, Armesto et al. (2010), for a discussion.

⁴ For a general description of the survey, see http://www.konj.se/1670.html

⁵ This way of summarizing the data is common practice in the literature; see, for example, Carabenciov et al. (2008).

⁶ For discussions concerning the importance of using real-time data, see, for example, Croushore and Stark (2001).

⁷ For a discussion of the importance of good nowcast/short-term forecasts, see, for example, Banbura et al. (2011).

⁸ The AR(1) model is a commonly used benchmark in the macroeconomic forecasting literature due to its simplicity, flexibility and a forecasting performance which typically tends to be decent; see, for example, Mitchell (2009) and Pesaran et al. (2009).

⁹ The weak serial correlation is visually confirmed in Fig. 1. In real time, standing at the point in time of the first forecast, GDP growth would actually have been judged a white noise process, according to the autocorrelation and partial autocorrelation functions. (These are not reported but are available upon request.)

¹⁰ This means that the earliest point in time at which it could have been made was late February/early March 2004.

¹¹ Some models are estimated on a shorter sample, starting in 1996Q2, since some of the survey data series are not available from 1993.

¹² No tests for whether differences in forecasting performance are statistically significant are conducted. Significance testing – using, for example, tests in the style of Diebold and Mariano (1995) – is, in our opinion, not particularly interesting in the present application. In line with, for example, Beechey and Österholm (2010), we argue that the model which minimizes the loss function of the forecaster (which here is assumed to be quadratic), should be the preferred one. For further criticism of significance testing, see Armstrong (2007).

¹³ Specifically, the best 10 models are those relying on data from questions 107, 108, 205 and 206 for the manufacturing industry, questions 107, 108, 201, 205 and 206, for the investments goods industry and question 205 for the construction industry; see table in the Appendix section for details. 

¹⁴ Other studies that have found that the Economic Tendency Survey has had predictive power for real economic variables in Sweden include Hansson et al. (2005) and Österholm (2010). Neither of these studies used real-time data though.