Forecasting macroeconomic variables using disaggregate survey data

Abstract

We construct factor models based on disaggregate survey data for forecasting national aggregate macroeconomic variables. Our methodology applies regional and sectoral factor models to Norges Bank’s regional survey and to the Swedish Business Tendency Survey. The analysis identifies which of the pieces of information extracted from the individual regions in Norges Bank’s survey and the sectors for the two surveys perform particularly well at forecasting different variables at various horizons. The results show that several factor models beat an autoregressive benchmark in forecasting inflation and the unemployment rate. However, the factor models are most successful at forecasting GDP growth. Forecast combinations using the past performances of regional and sectoral factor models yield the most accurate forecasts in the majority of the cases.

Introduction

Many central banks conduct surveys which yield regional and sectoral information on the general economic outlook. Following the example of the Federal Reserve’s Beige Book, which was implemented in 1970, and the Bank of England’s Agents survey, which begun in 1997, other central banks such as the Bank of Canada, Norges Bank, Sveriges Riksbank, and the Swiss National Bank have also initiated their own surveys. The information provided by these surveys is typically anecdotal and qualitative, unlike the well-known, quantitative Livingston survey, the Michigan survey, or the Survey of Professional Forecasters (see Thomas, 1999 for supplementary information about these surveys). While it is well-documented that the information obtained from quantitative surveys has strong forecasting power for macroeconomic variables (see for example Ang, Bekaert, & Wei, 2007; Fama & Gibbons, 1984; Mehra, 2002; and Thomas, 1999), there is less evidence of the forecasting power of information obtained from qualitative surveys.

This paper investigates the abilities of the Norges Bank’s regional survey and the Swedish Business Tendency Survey to forecast the gross domestic product (GDP) growth, consumer price inflation, and the unemployment rate for Norway and Sweden. Each survey consists of both backward- and forward-looking qualitative information. Studies such as those of  Abberger (2007),  Claveria, Pons, and Ramos (2007) and  Lui et al., 2011a, Lui et al., 2011b focus on examining specific survey questions in order to predict individual macroeconomic variables. Our approach is different, applying a dynamic factor model to the full database in order to construct regional and sectoral factors. These factors should contain the most relevant information for the regions and sectors from which they are extracted.

Our approach is similar to that of Hansson, Jansson, and Löf (2005), who use a dynamic factor model (based on net balance indices, representing differences between the shares of firms that have specified increases and decreases for a particular economic activity) from the Swedish Business Tendency Survey to forecast the Swedish GDP. Hansson et al. (2005) find that their factor model outperforms popular alternatives such as econometric VAR models in most cases. We extend their analysis in at least four directions. First, we consider the Norges Bank’s regional survey, which is more comprehensive in terms of sectors and regions of the economy. Our choice follows the claims made by Beck, Hubrich, and Marcellino (2009) that highly disaggregated regional and sectoral information is important in explaining aggregate Euro area and US inflation rates. Second, we work at a higher level of disaggregation and construct regional and sectoral factor models from the surveys. Out of ten sectors and seven regions for the Norwegian economy, and three sectors for the Swedish economy, our results identify which ones perform particularly well at forecasting different variables at various horizons. Third, we mitigate the uncertainties in the construction of the factors, the numbers of factors, and the relationships with the variable of interest by investigating two different classes of factor models where the number of factors is fixed a priori (denoted as model A) or estimated via a selection criterion (model B). Finally, we use forecast combinations to address the model uncertainty created by the use of several factors constructed by different datasets (regions or sectors). Each factor model is used to extract information and produce forecasts from a given dataset (regions or sectors) for the particular variable of interest.

We find that factor models based on several regions and sectors systematically beat the nowcasts and one-quarter-ahead forecasts of Norwegian inflation and unemployment rate given by the benchmark model. However, the factor models are most successful in nowcasting and forecasting GDP growth. Forecast combinations of the regional and sectoral models based on past performances are more accurate than the best regional or sectoral model in several cases, and provide more accurate forecasts than the benchmark model in almost all cases. Furthermore, we empirically find that aggregating the survey data either by pooling all of the Norwegian regional and sectoral survey information in a single factor model or by aggregating individual question-based forecasts via model combinations, to account for the heterogeneity in individual survey questions, results in less accurate forecasts than our regional and sector factor models. This finding is qualitatively similar when we use the Swedish Business Tendency Survey.

The paper proceeds as follows: Section  2 outlines the methodological aspects of our dynamic factor model, and Section  3 explains the forecasting models. Section  4 describes Norges Bank’s regional survey data, presents the factors and discusses the forecasting results. Section  5 reports results using the Sweden Business Tendency survey. Finally, Section  6 concludes.