1. 1 MEETING OF THE WORKING PARTY ON NATIONAL ACCOUNTS Paris, 11-14 october 2005 REVISIONS IN QUARTERLY GDP OF OECD COUNTRIES Tommaso Di Fonzo OECD and University of Padova

2. 2 Aims of the paper Building, maintaining and using revisions databases: the QNA-MEI experience Some remarks on strategy and tools for revisions analysis Revisions histories of GDP (real, sa) q-o-q growth rates in 18 OECD countries Size of revisions to the first published estimates (P) on MEI (May 1995 – Aug 2005)

3. 3 Main results Average number of months between the date of publication of GDP on MEI and the third month of the reference quarter

4. 4 Main results Revisions to the first published estimates of q-o-q GDP growth rates (%) on MEI. Mean absolute revision (1997.2-2002.1)

5. 5 Main results Revisions to successive estimates of q-o-q GDP growth rates (%). Mean absolute revision (1997.2-2002.1)

6. 6 The QNA-MEI revisions dataset OECD-ONS workshop on “Assessing and improving statistical quality. Revision analysis for the national accounts”, 7-8 October 2004 “The OECD also agreed to explore the possibility to expand and maintain the revisions database that was set up for purposes discussed in the workshop. An international dataset and the resulting summary measures of revisions should help to put countries’ individual revision analyses into an international context.” First elements presented also at the STESEG meeting, 27-28 June 2005

7. 7 OECD Main Economic Indicators Published in paper format since the early 1960’s Long time series for key short-term economic statistics for most OECD countries –Quarterly GDP and expenditure components –Indices of industrial production & Composite Leading Indicators –Retail Trade, Consumer and Producer Prices –Wages, employment and unemployment –Interest rates, exchange rates, monetary aggregates –Business tendency and consumer opinion surveys –International trade and international finance –Balance of payments

8. 8 MEI as a revisions database source MEI monthly snapshots in principle provide historical data on subsequent releases on a wide range of key variables –Provides a coherent set of variables representing the ‘information set’ available at successive monthly intervals MEI has been published as a CD Rom for many years

9. 9 OECD interest in a revisions database Analysis – Are revisions random (i.e. centred around 0)? Are they in general becoming smaller? Support internationally coordinated research work to improve the quality of early estimation methods Encourage transparency in the statistical process

10. 10 The QNA-MEI revisions dataset Initial OECD work (Ahmad, Bournot & Koechlin, 2004) was for G7 countries, period 1996 – 2000 Current project (Di Fonzo, STESEG 2005) –Expand coverage to all OECD countries –Expand time period covered –Devise a scheme which can support ongoing production of real time database & summary revision analysis indicators –Quarterly GDP and its components

11. 11 Data availability for GDP (constant prices, sa) in the QNA-MEI revisions database MEI editionsVintages Countries FirstLastFirstLast# May-95Aug-0595q105q21241: USA May-95Aug-0594q405q1124 13: AUS, CAN, CHE, DNK, ESP, FIN, FRA, GBR, ITA, JPN, NLD, NOR, NZL Nov-95Aug-0595q205q11181: DEU Nov-96Aug-0595q405q1 106 1: PRT Mar-97Aug-0596q305q2 102 1: BEL May-97Aug-0596q405q2 100 1: KOR Oct-03Aug-0503q205q1231: CZE* Nov-03Aug-0503q205q1221: AUT* Apr-04Aug-0503q405q1171: GRC Nov-04Aug-0504q204q4101: IRL Aug-05 05q1 21: HUN May-95Sep-0094q400q2652: SWE, TUR May-95Sep-0094q400q1651: MEX --- 04: ISL, LUX, POL, SVK

12. 12 QNA-MEI revisions dataset Caveats for international comparisons P are not necessarily flash estimates Timeliness of P differs between countries No attempt is made to isolate the different types of revisions, (including errors) Differences may exist between estimates and national estimates No investigation into influence of economic cycle Period covers SNA93 introduction and other changes, fixed base, chain-linking

13. 13 Release time of GDP (P) on MEI Further cautionary notes MEI is commonly published during the first week of the month Different release timing across countries There is no guarantee that the first number published in MEI is really the first number ever released ‘QNA-MEI view’ on timeliness in the release of GDP The true timeliness is ‘filtered’ through the publication practices adopted by OECD in the last decade

14. 14 Release time of GDP on MEI

15. 15 Average number of months between the date of publication of GDP on MEI and the third month of the reference quarter

16. 16 Some notation L: latest available estimate of GDP published on MEI at least 36 months after the first published estimate P The revision process is evaluated at 1, 2 and 3 years after the first published estimate: Y1: estimate published 12 months after P Y2: estimate published 24 months after P Y3: estimate published 36 months after P In Ahmad et al. (2004) the successive revisions between P and L are considered as the values published on MEI of December for the years t+1, t+2 and t+3 (D1, D2 and D3, respectively), where t denotes the year to which the reference quarter belongs In what follows, only Y1, Y2 and Y3 will be considered, while in the statistical annex also D1, D2 and D3 are analyzed

17. 17 Revisions database From levels … P Y1

18. 18 … to QoQ rates of change …

19. 19 … to revisions spreadsheet …

20. 20 … to charts …

21. 21 … to revisions analysis tables

22. 22 Revisions analysis: Summary statistics Mean revision Mean absolute revision Mean squared revision % L > P % sign(L) = sign(P)

23. 23 Revisions analysis Assessing the significance of the mean revisions Standard t-test Modified t-test using AR(1)-based corrections (ONS) Modified t-test using HAC (Heteroskedasticity, Autocorrelation Consistent) estimate of the variance of the mean revision (technical details in Di Fonzo, 2005) HAC procedure has been implemented in the OECD spreadsheet for revisions analysis

24. 24 Revisions analysis Mean squared revision and its decomposition Summary measure based on a quadratic loss function

25. 25 Revisions analysis Mean squared revision and a useful decomposition UM: proportion of MSR due to systematic differences between the preliminary and the latest estimates UR: proportion of MSR due to the slope coefficient in the regression UD: disturbance proportion of MSR, part of the revision which is not explained by the mean or slope error ‘Good’ preliminary estimates will have low values of UM and UR and a high value of UD

26. 26 Revisions to the first published estimates of q-o-q GDP growth rates (%) on MEI. Mean absolute revision (1997.2-2002.1)

27. 27 Revisions to successive estimates of q-o-q GDP growth rates (%). Mean absolute revision (1997.2-2002.1)

28. 28 Comments The revisions to the first published estimates have been of similar magnitude for most countries In Japan and Korea and, to a minor extent, Belgium, Denmark, Finland, New Zealand, Norway and Portugal, they have tended to be larger The size of the revisions over the period 1997.2 to 2002.1 becomes smaller: on average |Y1-P| > |Y2-Y1| > |Y3-Y2| for 10 countries over 18 (Australia, Belgium, Canada, Denmark, Finland, Italy, Korea, Netherlands, Switzerland and USA). The opposite is registered for Spain and, less markedly, for France

29. 29 Is the mean revision to the first published estimates significant?

30. 30 Is the mean revision from successive estimates significant?

31. 31 Comments The test conducted to check whether mean revisions are statistically different from zero gives the following results: –Revisions to the first published estimates of GDP: evidences confirming this fact emerge for Canada, Denmark, France, Italy, Norway and United Kingdom. –Process of revision: significant mean revision of Y2 to Y1 only for Canada (last common 20 quarters); the mean revision of Y3 to Y2 resulted significant for Canada, Japan, New Zealand (last 20 quarters), Switzerland and United Kingdom (whole sample); Germany is the only country showing significant mean revision in the Latest vs. Y3 comparison.

32. 32 Obstacles to extending the work Possible gaps in our electronic sources pre-1995, only resort may be paper copies.... –How far back should the revisions database go? Operational problems (e.g. series code changes) Methodological issues –Constancy of national series over time (e.g. various methodological changes causing breaks & revisions) –OECD procedures (basically, for QNA series, seasonal adjustment) Data management (produces big databases)

33. 33 Elements for discussion Presentation of revisions, evaluating the first revision registered on MEI, effect of different time periods on the summary statistics, stability of the results, use of other (simple) descriptive measures? Possible use of a QNA-MEI revisions database by member countries? Priority for analyzing other QNA series and how to use the results? - e.g. as an input to internationally coordinated methodological work? - To reconsider policies for communicating information on revisions to users?

34. 34 THANK YOU!