Testing for news and noise in non-stationary time series subject to multiple historical revisions Citation for published version (APA): Hecq, A. W., Jacobs, J. P. A. M., & Stamatogiannis, M. (2016). Testing for news and noise in non- stationary time series subject to multiple historical revisions. Maastricht University, Graduate School of Business and Economics. GSBE Research Memoranda No. 004 https://doi.org/10.26481/umagsb.2016004 Document status and date: Published: 01/01/2016 DOI: 10.26481/umagsb.2016004 Document Version: Publisher's PDF, also known as Version of record Please check the document version of this publication: • A submitted manuscript is the version of the article upon submission and before peer-review. There can be important differences between the submitted version and the official published version of record. People interested in the research are advised to contact the author for the final version of the publication, or visit the DOI to the publisher's website. • The final author version and the galley proof are versions of the publication after peer review. • The final published version features the final layout of the paper including the volume, issue and page numbers. Link to publication General rights Copyright and moral rights for the publications made accessible in the public portal are retained by the authors and/or other copyright owners and it is a condition of accessing publications that users recognise and abide by the legal requirements associated with these rights. • Users may download and print one copy of any publication from the public portal for the purpose of private study or research. • You may not further distribute the material or use it for any profit-making activity or commercial gain • You may freely distribute the URL identifying the publication in the public portal. If the publication is distributed under the terms of Article 25fa of the Dutch Copyright Act, indicated by the “Taverne” license above, please follow below link for the End User Agreement: www.umlib.nl/taverne-license Take down policy If you believe that this document breaches copyright please contact us at: [email protected] providing details and we will investigate your claim. Download date: 30 Dec. 2022 Alain Hecq, Jan P.A.M. Jacobs, Michalis P. Stamatogiannis Testing for news and noise in non-stationary time series subject to multiple historical revisions RM/16/004 Testing for news and noise in non-stationary time series subject to multiple historical revisions Alain Hecq Maastricht University [[email protected]] Jan P.A.M.Jacobs University of Groningen, University of Tasmania, CAMA and CIRANO [[email protected]] Michalis P. Stamatogiannis University of Liverpool [[email protected]] This version: January 2016 Abstract Before being considered definitive, data currently produced by statistical agencies undergo a recurrent revision process resulting in different releases of the same phenomenon. The collection of all these vintages is referred to as a real-time data set. Economists and econometricians have realized the importanceofthistypeofinformationforeconomicmodelingandforecasting. This paper focuses on testing non-stationary data for forecastability, i.e., whether revisions reduce noise or are news. To deal with historical revisions which affect the whole vintage of time series due to redefinitions, method- ological innovations etc., we employ the recently developed impulse indicator saturation approach, which involves potentially adding an indicator dummy for each observation to the model. We illustrate our procedures with the U.S. Real Gross National Product series from ALFRED and find that revisions to this series neither reduce noise nor can be considered as news. Keywords: Data revision, Non-Stationary Data, News-Noise Tests, Structural Breaks JEL-code: C32, C82, E01 1 Introduction Before being considered definitive, data currently produced by statistical agencies typically undergo a recurrent revision process resulting in different releases of the same phenomenon. The collection of all these vintages is referred to as a real-time data set. In the recent past, economists and econometricians have come to realize the importance of this type of information for economic modeling, forecasting and policy formulation. Consequently there exists a growing interest for investigating this type of data (see inter alia Croushore and Stark, 2001, Orphanides and van Norden, 2002, and Croushore, 2011a, 2011b). Several aspects of real-time data can be investigated: (i) structural or trend breaks(seeJacobsandvanNorden(2015)forasummaryofthereliabilityofproduc- tivity growth rate trends); (ii) forecastability, i.e., whether revisions reduce noise or are news (the literature is briefly reviewed in Section 3.1); (iii) historical revisions, which affect the whole vintage of time series due to redefinitions, methodological innovations, etc., make testing difficult. The standard approach to dealing with his- torical revisions is either to employ growth rates to mitigate the effects of historical revisions, or to ‘clean’ the series in an attempt to get rid of the effects of historical revisions. The former approach has been criticized by Knetsch and Reimers (2009). Go¨tz, Hecq and Urbain (2016) illustrate that growth rates can also be affected by large revisions. Whereas the tests and the procedures to deal with historical revisions are well- documented for stationary time series (e.g., using Mincer-Zarnowitz type tests), the situation is less clear for non-stationary time series. The paper aims at filling this void, building upon Hecq and Jacobs (2009). We focus on testing forecastability for non-stationary real-time data, putting data releases in vector-error correcting forms (VECMs hereafter). To deal with forecastability under historical revisions at unknown dates, we estimate VECMs using an automatic modelling method for selecting conditional mean parameters (the Autometrics algorithm, see Hendry and 1 Doornik, 2014) together with the Impulse Indicator Saturation approach (IIS here- after, see e.g., Hendry and Santos, 2005). Briefly, IIS involves adding an indicator dummy1 forpotentiallyeachobservationtothemodelandhenceisabletodetermine a parsimonious model that fits model requirements in terms of misspecifications. We illustrateourprocedureswiththeU.S.RealGrossNationalProduct(realGNPhere- after) series from ALFRED and find that, in general, revisions neither reduce noise nor can be considered as noise. Conclusions would have been different without the IIS approach. An alternative strategy to the IIS algorithm consists of introducing dummy vari- ables for each revision because historical revisions are often well documented and recorded. This operation is less obvious than one might think at first glance and can be very tedious and time consuming for an external researcher who does not have complete information on thousands of economic variables for different coun- tries. While one can easily find the description of the modifications for the main aggregates for the U.S. or the European Union for instance, this task is much more demanding when the information about data revisions is for instance not in English or not available online on national statistical agencies websites. Using IIS helps in investigating those time series within a few seconds. Secondly, one can also notice that the date at which vintages are released might differ from the date at which the series has been theoretically modified. As an example, books may describe that there is a new definition of an economic indicator in January but the series pub- lished on, say the 10th of January, still applies the old definition. It might be for this latter example that a second vintage is available at the end of the month such that we observe multiple vintages for one particular month, a situation that adds difficulties for the researcher. Third, IIS can also capture smaller revisions (e.g., annual or seasonal revisions due to e.g., the change of seasonal factors) that would have ignored based on historical revisions only. Finally, many real-time databases 1We leave for further research the use of additional step dummies in the IIS framework. 2 have been build manually, either by merging files or using manpower for scanning or copying figures from statistical reports. Those operations can also introduce errors. The remainder of the paper is structured as follows. After a brief introduction of data revisions and notations in Section 2, Section 3 describes news-noise tests for stationary and non-stationary real-time data as well as the intuition underlying the IIS approach. Section 4 illustrates our procedure with the U.S. Real National Product series. Section 5 concludes. 2 Data revisions and notation Real-time data are typically displayed in the form of a real-time data trapezoid as in Figure 1. We move to later vintages as we move across columns from left to right and we move to later points in time when we move down the rows. Note that the frequency of vintages needs not necessarily correspond to the unit of observation; for example, in our illustration below the statistical agency publishes monthly vintages of quarterly observations. In this paper we investigate the releases, namely the diagonals of the data trapezoid. We use the following notation: superscripts refer to releases i = 1,...,v, while subscripts refer to periods; y1 denotes the first release t for variable y in period t, whereas the sequence {y1}T or simply y1, t = 1,...,T t t=1 t refers to as the whole time series for the first release, namely the first diagonal in Figure 1. Figure 1: The real-time data trapezoid y1 ... yi ... yv  1 1 1  ... ... ... ...     y1 ... yi   t−l t−l  ... ...  y1 t 3 Data revisions may be conveniently categorized into three types: 1. initial revisions in the first few vintages, 2. annual (seasonal) revisions due to updated seasonal factors and the confronta- tion of quarterly with annual information, and 3. historical or comprehensive revisions, related to changes in statistical method- ology, etc. The distinction of revisions into these types requires careful handling of the real- time data and in many cases direct access to the officials of the statistical agency. Initial and seasonal revisions are regular and recurring, and can in principle be modeled and forecast. As an example, Eurostat releases its first estimate of e.g., real GDP 45 days after the end of the corresponding quarter (flash estimate), the next release is 15 days later. Historical revisions are much more difficult to handle. Redefinitions like changes of base years do not cause many difficulties, however changes in definitions changes do. Whatevertheirorigins, datarevisionsimplytheexistenceofmeasurementerrors. The modeling of measurement errors has two main traditions that are surveyed in the next section. 3 Method 3.1 News-noise tests Stationary data The older tradition, which is still widespread among statisticians, is that measure- ment errors should be thought of as noise. Data are measured with errors which are orthogonal to true values (y ). This implies for a stationary time series y that for (cid:101)t t 4 all releases i yi = y +ζi, cov(y ,ζi) = 0. (1) t (cid:101)t t (cid:101)t t One implication of this is that revisions will generally be forecastable by taking weighted averages of previous releases. To test whether measurement errors reduce noise, the Mincer-Zarnowitz (1969) test can be used, which regresses the revision yCV −yi on a constant and the most recent, i.e., the current (last observed column) t t vintage yCV, taken as measure of the unobserved true value y t (cid:101)t yCV −yi = δ +β yCV +ζi. (2) t t 1 1 t t The null hypothesis that measurement errors are independent of true values (δ = 1 0, β = 0) may be tested with a Wald test; since the errors may suffer from het- 1 eroskedasticity and autocorrelation, robust HAC standard errors are typically used. The“newer”tradition,motivatedbyMankiw,RunkleandShapiro(1984),Mankiw and Shapiro (1986) and de Jong (1987), describes measurement errors as news.2 News errors imply that published data are optimal forecasts, so revisions are or- thogonal to earlier releases and are not forecastable. More precisely, yCV = yi +νi, cov(yi,νi) = 0. (3) t t t t t The analogous test of the “news” model regresses the revision (yCV − yi) on a t t constant and the ith-release yCV −yi = δ +β yi +νi. (4) t t 2 2 t t A similar null hypothesis (δ = 0, β = 0) now tests whether data revisions are 2 2 predictable. The two null hypotheses are mutually exclusive but they are not collec- 2See also the more recent contributions of Faust, Rogers and Wright (2005) , Swanson and van Dijk (2006) and Aruoba (2008). More references are in Jacobs and van Norden (2011). 5 tively exhaustive, i.e., we may be able to reject both hypotheses, particularly when the constant in both test equations differs from zero (see Aruoba, 2008, Appendix A.2). The main conclusion of the empirical literature on characteristics of real-time data is that macroeconomic time series are in principle not well-behaved. Revisions can be substantial and reduce noise or add news at different horizons. An alternative way to test for news and noise is to estimate the Jacobs and van Norden (2011) data revision model, a state-space form in which measurement errors are decomposed into news and noise with the possibility of spillovers. Recently, Clements and Galv˜ao (2013) extended the Jacobs and van Norden (2011) framework by allowing for revision bias. The alternative state-space forms of Cunningham et al. (2012) and Kishor and Koenig (2012) should in principle be able to do the same. FixlerandNalewaik(2009)proposeanalternativetest, whosepropertiesstillhaveto be explored. Finally, the multi-period survey approach of Patton and Timmermann belongs in this category too. Non-stationary data Testing measurement errors in case of non-stationary variables is more complicated even when a single time series, like gross national product, is considered. Indeed, the existence of cointegration between different releases hampers the application of Mincer-Zarnowitz tests explained above for two reasons. First, the presence of coin- tegration implies that there exists a long-run relationship between different releases and hence news/noise tests would be subject to the usual omitted variable problem if we estimate (4) or (2) on the growth rates of time series only. Second, assuming that we correctly account for cointegrated I(1) series in VECM systems, the issue still remains that we cannot establish the direction of causality, i.e., whether the first release is explained by the last release, or the other way around. However, weak exogeneity tests in cointegrated systems (see Urbain, 1992, 1995) can be helpful here. 6 Cointegration between time series of different releases, or intra-variable cointe- gration, can be modeled in two ways.3 The approach most frequently adopted in the literature looks at releases on an observation basis, for example first and second releases of variable y observed on T+1 data points. The Observation Balanced Sys- t tem (OBS hereafter) tests for cointegration between series y1 and y2, t = 1,...,T. t t Superscripts denote respectively the first and the second released diagonals as vi- sualized in Figure 2. The red box emphasizes that the most recent observation in period T + 1 is excluded from the analysis. Note also that in this description we explicitly consider that publication lags do not exist. We also prefer to denote the diagonal releases while making an explicit reference of the release number instead of adding another t component in the superscript of y. It must be understood though that we take the first two releases as a convenient explanatory example but that we investigate the relationships between several releases in this paper.4 Figure 2: Observation Balanced System (OBS) The alternative approach compares the releases on a vintage basis, i.e., the two most recent observations of vintages. In the Vintage Balanced System (VBS here- after) cointegration between y1 and y2, t = 1,...,T, is considered, see Figure 3. t+1 t 3The remainder of this section draws upon Hecq and Jacobs (2009). 4We leave the multivariate investigation of the whole set of releases for further investigations. In this paper we only look at pairwise tests. 7