Forecasting and nowcasting real GDP: Comparing statistical models and subjective forecasts

Abstract

We conduct a systematic comparison of the short-term forecasting abilities of twelve statistical models and professional analysts in a pseudo-real-time setting, using a large set of monthly indicators. Our analysis covers the euro area and its five largest countries over the years 1996–2011. We find summarizing the available monthly information in a few factors to be a more promising forecasting strategy than averaging a large number of single-indicator-based forecasts. Moreover, it is important to make use of all available monthly observations. The dynamic factor model is the best model overall, particularly for nowcasting and backcasting, due to its ability to incorporate more information (factors). Judgmental forecasts by professional analysts often embody valuable information that could be used to enhance the forecasts derived from purely mechanical procedures.

Introduction

Information on economic activity and its short-term prospects is very important for decision makers in governments, central banks, financial markets and non-financial firms. Monetary and economic policy makers and economic agents have to make decisions in real time based on incomplete and inaccurate information on current economic conditions. A key indicator of the state of the economy is the growth rate of real GDP, which is available on a quarterly basis only, and is also subject to substantial publication lags. In many countries, an initial estimate of quarterly real GDP is published around six weeks after the end of the quarter. Moreover, real GDP data are subject to revisions that can be substantial, as more data become available to statistical offices over time.

Fortunately, though, a lot of statistical information related to economic activity is published on a more frequent and timely basis. This information includes data on industrial production, prices of goods and services, expenditures, unemployment, financial market prices, loans, and consumer and business confidence. Recently, the forecasting literature has developed several statistical approaches for exploiting this potentially very large information set in order to improve the assessment of both real GDP growth in the current quarter (nowcast) and its development in the near future. Examples of such approaches include bridge models, factor models, mixed-data sampling models (MIDAS) and mixed-frequency vector-autoregressive (MFVAR) models. These models differ in their solutions to the practical problems of dealing with large information sets and the fact that the auxiliary variables are observed at different frequencies and with different publication lags.

Practitioners now have a wealth of statistical models to choose from; but which one should they use? As each model has its own strengths and weaknesses, it is difficult to make a decision on purely theoretical grounds. The ranking of the models in terms of forecasting abilities, and the extent to which this varies with the prediction horizon or the economic circumstances, has to be determined by empirical analysis. On these issues the jury is still out, however, as large-scale comparative studies are scarce. In many papers, the empirical work refers only to a single country, and usually only limited numbers of models are included. Furthermore, studies differ in the size of the information set and the sample period used.¹

This paper is motivated by this gap in the empirical literature. We undertake a systematic comparison of a broad range of linear statistical models–twelve models in all–that have been applied in the recent literature. For the sake of comparability and robustness, we include five countries (Germany, France, Italy, Spain and the Netherlands) and the euro area in our analysis, and utilize an information set that is as homogeneous as possible across geographical entities. Moreover, our sample includes the volatile episode of the financial crisis of 2008 and its aftermath, which may make it easier to distinguish between the various models. We contrast the models’ forecasting abilities before 2008 with those during the crisis period. This may be of great interest for policy makers, financial analysts and economic agents alike, as information on where the economy stands and where it is headed in the immediate short run is particularly valuable at times of great uncertainty.

The provision of cross-country evidence on the relative performances of twelve different statistical forecasting models is our first contribution to the literature. Model forecasts are the result of purely mechanical recipes, and do not incorporate subjective elements. Our second contribution concerns the potential usefulness of forecasts made by professional analysts (published by Consensus Forecasts on a quarterly basis). From a practical point of view, such forecasts are very cheap and easy to use. Moreover, as an expression of the “wisdom of crowds”, they may reflect much more information than the statistical information set, which is inevitably limited. A questionnaire conducted by the European Central Bank (ECB) among the participants of the ECB Survey of Professional Forecasters found that the panelists regard 40% of their short-term GDP forecasts as being judgment-based (Meyler & Rubene, 2009). We investigate the extent to which the subjective forecasts by analysts in our sample contain information beyond that generated by the best mechanical statistical models.

The remainder of the paper is structured as follows. Section  2 describes the various statistical models and discusses how they deal with the challenges posed by large and irregularly shaped datasets. Section  3 describes the data, our pseudo real-time forecast design, and other specification issues. Sections  4 Empirical results for statistical models, 5 Analysis of forecasts by professional analysts present the results for the mechanical models and the professional forecasts, respectively. Section  6 summarizes our findings and concludes.