صحت، عدم جانبداری و بهره وری پیش بینی های اقتصاد کلان حرفه ای : مقایسه تجربی برای G7

In this paper, we use survey data to analyze the accuracy, unbiasedness and efficiency of professional macroeconomic forecasts. We analyze a large panel of individual forecasts that has not previously been analyzed in the literature. We provide evidence on the properties of forecasts for all G7-countries and for four different macroeconomic variables. Our results show a high degree of dispersion of forecast accuracy across forecasters. We also find that there are large differences in the performances of forecasters, not only across countries but also across different macroeconomic variables. In general, the forecasts tend to be biased in situations where the forecasters have to learn about large structural shocks or gradual changes in the trend of a variable. Furthermore, while a sizable fraction of forecasters seem to smooth their GDP forecasts significantly, this does not apply to forecasts made for other macroeconomic variables.

In this paper, we use survey data to analyze the accuracy, efficiency, and unbiasedness of professional macroeconomic forecasts in the G7 countries. We analyze individual forecasts from large cross sections of professional forecasters, enabling us to throw light on the heterogeneity across forecasters. Moreover, our results are not affected by problems that arise from the use of average, so-called consensus, forecasts (e.g., aggregation bias). Our large data set has not previously been used exhaustively in the literature. By using this large amount of disaggregate data on individual macroeconomic forecasts, we are able to provide a much broader evidence base on the properties of macroeconomic forecasts than has previously been available in the literature. One weak point of the empirical literature, which uses survey data to assess the efficiency or unbiasedness of macroeconomic forecasts, is that only a limited number of non-US data sets provide information on forecasts. As a consequence, the existing evidence is based predominantly on US data. Notable exceptions are Harvey, Leybourne, and Newbold (2001), who analyze a set of selected individual forecasts for the UK from the survey data set provided by Consensus Economics; Gallo, Granger, and Jeon (2002), who analyze the evolution of macroeconomic forecasts for the US, the UK, and Japan; Bowles et al. (2007), who analyze the performances of forecasts summarized in the Survey of Professional Forecasters, conducted by the European Central Bank; Ager, Kappler, and Osterloh (2009) and Isiklar, Lahiri, and Loungani (2006), who use data from the Consensus Economics data set on forecasts for a set of industrialized countries; Loungani (2001), who additionally examines data for developing countries; Timmermann (2007), who analyzes the performances of IMF forecasts from the World Economic Outlook for various countries; Batchelor (2001), who compares the forecasts made by the IMF and the OECD to private sector forecasts; and Boero et al., 2008a and Boero et al., 2008b, who analyze forecasts from the Bank of England Survey of External Forecasters. However, all existing international studies, with the exception of Boero et al., 2008a and Boero et al., 2008b and Harvey et al. (2001), make exclusive use of consensus forecasts rather than analyzing individual forecasts (note though that these three studies are confined to UK data sets). The purpose of our paper is to fill this gap, covering individual forecasts for all G7 countries and for four macroeconomic variables. Our results are based on an approach which has commonly been used in the literature to model the structure of macroeconomic forecasts, dating back to early contributions by Ball (1962), Figlewski and Wachtel (1981), Mincer and Zarnowitz (1969) and Nordhaus (1987), who introduced the basic modeling framework for analyzing fixed event forecasts.2 A sequence of fixed event forecasts consists of consecutively formed forecasts for the same event (such as an annual figure for a macroeconomic variable). The data we use below are of this type. Some more recent contributions have proposed improving the econometric approach for testing the rationality of such large panels of fixed event forecasts. These include Batchelor and Dua (1990) and Keane and Runkle (1990), who introduce an analysis in a panel framework using the Generalized Methods of Moments (GMM) method, as well as Davies and Lahiri (1995), who develop a framework for analyzing three-dimensional panels of survey data, enabling the use of information along all dimensions. To ensure that our results are comparable to existing studies, we closely follow the approach which was suggested by Davies and Lahiri (1995), and recently used by Ager et al. (2009), Boero et al. (2008a), and Clements, Joutz, and Stekler (2007), and suggest only minor modifications to the econometric framework. Using this framework, we analyze the accuracy and heterogeneity of the forecasts provided by the panelists of the survey and test whether or not they are unbiased and efficient. Assuming that forecast accuracy is the only objective of a forecaster and that her loss function is symmetric and increases with the forecast error, the latter two properties are inevitable features of a rational forecast. Regarding this point, it should be noted, however, that there are also arguments against the assumption that published forecasts reflect true expectations and are meant to minimize a loss function of the described form. Some of these arguments are as follows. First, forecasters might seek to maximize public attention. In this case, an unbiased forecast is no longer optimal, since the utility of the forecaster depends on more than one argument (Laster, Bennett, & Geoum, 1999). Second, forecasters might produce a so-called “intentional” forecast in some situations (Stege, 1989); for example, a forecaster could predict a specific event in order to provoke a policy action that actually prevents the occurrence of the event. Third, forecasters might have asymmetric loss functions (Boero et al., 2008a and Capistrán and Timmermann, 2009). These could have different weights on possible over- or underestimations of an outcome. However, we believe that these arguments are not strong a priori, particularly because the identities of the panelists are revealed in the data set we use. We therefore ignore these issues and base this paper on the null hypothesis that it is in the forecasters’ best interests to provide unbiased and efficient forecasts. Our findings show that the dispersion of forecast accuracies across panelists is surprisingly high for most of the countries and variables examined in this paper. We also find that there are large differences in the performances (in terms of accuracy, unbiasedness, and efficiency) of forecasters, not only across countries, but also across different macroeconomic variables. In general, the forecasts for inflation are mostly consistent with the hypothesis of unbiased and efficient forecasts. Furthermore, the forecasts tend to be biased in situations where the forecasters have to recognize either large structural shocks or gradual changes in the trend component of a variable. The remainder of this paper is structured as follows. Section 2 presents a brief overview of the data set we use and a first visual inspection of the data. Section 3 illustrates the econometric framework we use for modeling the forecast errors, and shows how tests on the unbiasedness and efficiency of forecasts can be derived. Section 4 discusses the heterogeneity in the accuracy of individual forecasts. Section 5 presents the empirical results on the unbiasedness of individual forecasts. Section 6 presents the empirical results on the efficiency of individual forecasts. Finally, Section 7 concludes.

In this paper, we have analyzed individual macroeconomic forecasts for all G7 countries based on survey data from the Consensus Economics data set. We have shown the degree of heterogeneity across panelists with respect to the forecast accuracy and tested whether or not the forecasts in the sample are unbiased and weakly efficient. The empirical results lead to the following conclusions. First, the dispersion of forecast accuracies is surprisingly high. Second, we observe that forecasters who perform well in terms of forecast accuracy for real GDP growth are likely to also perform well for other variables. Third, we find large differences between the performances of forecasters with respect to unbiasedness and efficiency across both countries and different macroeconomic variables. Fourth, of the four kinds of forecasts analyzed, inflation forecasts perform best in terms of unbiasedness. Fifth, forecasters, on average, seem to smooth their GDP forecasts more heavily than the other macroeconomic forecasts they make. Sixth and lastly, forecasts tend to be biased in situations where the forecasters have to realize either large structural shocks or gradual changes in the trend of a variable. As a consequence, if a sizeable fraction of panelists produce biased forecasts for a variable, then virtually all of them are biased in the same direction, i.e., the biases are not uncorrelated across panelists. There are several dimensions along which this research could be expanded in the future. For simplicity, we have assumed that both the variance of the macroeconomic shocks (λt,hλt,h) and the variance of the forecaster-specific error component (ϵi,t,hϵi,t,h) decay linearly if hh goes to 1. Firstly, more general functional forms could be developed in the future to match the data better. Secondly, as soon as a sufficient number of longer time series become available for individual forecasts, one could implement the estimation of the horizon-specific bias, which would be more attractive from a theoretical point of view. At present, however, the time dimension of the data set is too small; that is, for most of the panelists the estimates would be based on fewer than ten observations. Finally, taking correlations across countries into account — as Isiklar et al. (2006) do in their analysis of consensus forecasts — would clearly be desirable, given the high impact that international shocks can potentially have on the size of forecast errors. However, this would require animmense computational power for the estimation of the covariance matrices.