News and Noise in G-7 GDP Announcements

Board of Governors of the Federal Reserve System International Finance Discussion Papers Number 690 December 2000 NEWS AND NOISE IN G-7 GDP ANNOUNCEMENTS Jon Faust, John H. Rogers, and Jonathan H. Wright NOTE:InternationalFinanceDiscussionPapersarepreliminarymaterialscirculated to stimulate discussion and critical comment. References to International Finance Discussion Papers (other than an acknowledgment that the writer has had access to unpublished material) should be cleared with the author or authors. Recent IFDPs are available on the Web at www.federalreserve.gov.

NEWS AND NOISE IN G-7 GDP ANNOUNCEMENTS (cid:3) Jon Faust, John H. Rogers, and Jonathan H. Wright Abstract: Revisions to GDP announcements are known to be quite large in all G-7 countries: many revisions in quarterly GDP growth are over a full percentage point at an annualized rate. In this paper, we examine the predictability of these data revisions. PreviousworksuggeststhatU.S.GDPrevisionsarelargelyunpredictable, as would be the case if the revisions reflect news not available at the time that the preliminary number is produced. We (cid:12)nd that the degree of predictability varies throughoutthe G-7. For the U.S., the revisions arevery slightly predictable, butfor Italy, Japan and the UK, about half the variability of subsequent revisions can be accountedforbyinformationavailableatthetimeofthepreliminaryannouncement. For these countries, it appears that revisions reflect, to a signi(cid:12)cant degree, the removal of noise from the preliminary numbers, rather than the arrival of news. Keywords: Vintage data, preliminary data, (cid:12)nal data, revision, GDP. (cid:3) Division of International Finance, Board of Governors of the Federal Reserve System. We thank for useful comments Dean Croushore, Joe Gagnon, Jane Haltmaier, David Howard, Jane Ihrig, Andrew Levin, Athanasios Orphanides, Matthew Shapiro, and Charles Thomas. Thanks to Anthony Leegwater, David Kobuszewki, and Lisa Workman for excellent research assistance. The views in this paper are solelytheresponsibilityoftheauthorsandshouldnotbeinterpretedasreflectingthe views of the Board of Governors of the Federal Reserve System or of other members of its sta(cid:11).

1 Introduction Macroeconomic data are often subject to large revisions after initial release and for many data series, the data revision process continues essentially inde(cid:12)nitely. The inaccuracy of initial data obviously complicates decision making by policymakers andother agents whoseoptimal choices dependon thestate of theeconomy. Several authors have recently emphasized that the revision process also complicates the ex post analysis of macro data. Macroeconometric work is generally based on the most fully revised available data (ignoring any earlier data vintages). The conclusions of such work rest on the implicit assumption that at each point in time, agents perfectlypredictfuturedatarevisions. Thisassumptionisparticularlycuriousgiven the fact that revisions should be completely unpredictable if the statistical agency is e(cid:14)ciently processing available information. Studies of the revision process in U.S. GDP data show that unpredictability or weak predictability of revisions is a better assumption than perfect predictability (e.g., Mankiw and Shapiro, 1986). Further, macroeconometric work based on real-time data often yields substantively di(cid:11)erent conclusions from work ignoring revisions (see, for example, Diebold and Rudebusch (1991), Robertson and Tallman (1998), Evans (1998), Orphanides (1998, 2000) and Croushore and Stark (1999)). In this paper, we study the predictability of revisions to GDP announcements for the G-7 countries. We extend earlier work both by including a broader range of economies andbytakingalongersampleofdata|atlongest1965-1997. Aswiththe earlierwork, theprimarytools weusearestatistical tests ofwhetherthepreliminary announcement is a rational forecast of the subsequently revised data. We (cid:12)nd some evidence of predictability of revisions for each country we consider. For several countries|the UK, Italy, and Japan|the revisions are highly predictable: about half the variance of revisions can be accounted for by information available at the time of the initial announcement. Preliminary announcements in some countries are biased. The variable with the most power to predict future revisions to GDP is the preliminary GDP announcement itself|extreme prelimi- 1

nary announcements tend to be revised toward the mean. We consider seasonal dummy variables and 5 other predictors that are publicly known at the time of the preliminary announcement: the lagged preliminary announcement, an equity index, short-term interest rates, oil prices, and a dummy variable for national elections. While each of these variables has predictive power in at least one country, in no case do these additional variables make major contributions to the overall predictive power. Most notable among these results for additional predictors is the fact that in the Japanese data, the last GDP growth rate announced beforenational elections is systematically about one percentage point (at an annual rate) more optimistic than in other quarters. This result is robust for Japan, but the election e(cid:11)ect is not found in any other country. We discuss possible sources and implications of all these results in the (cid:12)nal section. The plan of the remainder of this paper is as follows. Section 2 reviews standard models of the revision process, the role of e(cid:14)cient forecasts and forecast tests. Section3describesourdataset. Sections4and5presentbaselineresultsandextensions, and the (cid:12)nal section provides some interpretation of the results. 2 Revisions, News, Noise, and E(cid:14)cient Predictions The revision process is characterized by two polar cases labelled news and noise by Mankiw, Runkle, and Shapiro (1984) and Mankiw and Shapiro (1986). Under the noise view, the preliminary GDP data are polluted with measurement error, noise, that is uncorrelated with the true values. The preliminary GDP number will not be an optimal estimate of GDP in this case, and agents face a (cid:12)ltering problem in forming their optimal estimate. Various approaches to this (cid:12)ltering problem have been proposed (Howrey, 1978, 1984; De Jong, 1987; and Mariano and Tanizaki, 1995). Under the news characterization, the statistical agency optimally uses all availableinformationinformingthepreliminarynumber,andrevisions mustreflectnews that arrives after the announcement. The news view will be appropriate so long as 2

the statistical bureau is choosing the preliminary number to minimize any one of a number of standard loss functions that are symmetric and increasing in the size of revisions. Some national statistical o(cid:14)ces explictly discuss minimizing revisions as a goal of their processes, and most presumably include this among the desiderata 1 in data construction. When revisions are minimized, the preliminary number will be what is known as a rational, e(cid:14)cient, or optimal forecast of the subsequently revised data. The revision will be orthogonal to information available when it is produced. More formally, under both the news and noise views, we can characterize the preliminary data as equal to the (cid:12)nal plus an error term: X t p = X t f +" t : (1) Underthenoiseview," t isorthogonaltoX t f ,whileinthenewsview," t isorthogonal to X t p . Obviously, there are intermediate cases in which " t is correlated with both preliminary and (cid:12)nal data. Predictability of revisions can arise in fairly innocuous ways. Suppose output data are available before income and expenditure data so that the preliminary announcement gives GDP measured on an output basis only. The (cid:12)nal data involve some reconciliation of output-based, expenditure-based, and income-based 2 methods. Even if each of the three sources of information are unbiased and similarly noisy, using the output-based number (or any of the three) as a preliminary will give rise to predictable revisions. An optimal linear estimate of GDP would involve scaling the output-based number by a factor reflecting the signal-to-noise ratio in those data. We follow others in attempting to distinguish the news and noise views using standard forecast e(cid:14)ciency tests. Under the news view, revisions must be mean zero; under noise, they need not be. Thus, we initially test the hypothesis that 1FortheU.S.,seetheeditorsnoteinMankiwandShapiro(1986);fortheU.K.,Barklem(2000), and for Japan, Economic Research Institute(2000). 2 The UK process is roughly like this hypothetical example. These three bases for GDP measurement and thereconciliation are explained in Reed (2000). 3

the revisions are unbiased. If " t in (1) is correlated with X t p as in the noise case, Xp will predict the subsequent revision, R(t) (cid:17) Xf −Xp . To test this we use the t t t classic Mincer-Zarnowitz (1969) forecast e(cid:14)ciency test, which involves running the regression, R t = (cid:11)+(cid:12)X t p +u t : (2) We test the forecast e(cid:14)ciency implication that (cid:11) = (cid:12) = 0. One can also augment thesetofexplanatoryvariablesintheMincer-Zarnowitzregressionwithanyvariable known at time t. Forecast rationality implies that all the coe(cid:14)cients should bezero. Several earlier papers have applied these tests using revisions to money stock 3 and/or output data, typically for the U.S. Mankiw and Shapiro (1986) (cid:12)nd little evidence against the null hypothesis of forecast rationality using a short sample of U.S. GNP data and so characterize the data revision process as incorporating news. Using U.S. consumption data, Croushoreand Stark (1999) (cid:12)nd that revisions up to 1 year after the initial data release are uncorrelated with the preliminary data, but that subsequent revisions are weakly predictable using preliminary data. Barklem(2000)(cid:12)ndsevidenceofbiasinrevisionstovariousUKpreliminaryreleases, including GDP. All of this work, including ours, requires one caveat: because data construction methods are constantly being revised, it is unclear whether past predictability is evidence of future predictability. We partially address this problem by considering both a full sample and a more recent 10-year sample and by using pseudo out-ofsample methods. It remains true that methods in most countries have changed considerably, even in the past decade. Considering even shorter samples is problematic because statistical power falls as the sample size shrinks and because very recent data has not had much time to be revised. 3 Mankiw, Runkleand Shapiro (1984), using U.S. money stock data, reject the null hypothesis offorecastrationality. Otherauthors,includingKavajeczandCollins(1995),alsousingU.S.money stockdata,reject forecast rationality whenusingseasonally adjusteddata, butnotwhenusingthe unadjusted data. Similar results are obtained for Canadian money stock data by Milbourne and Smith (1989). 4

3 Our preliminary and (cid:12)nal GDP data Our data comprise preliminary estimates of real quarterly GDP growth rates and their subsequent revisions for the G-7 countries. The data come from the OECD’s Main Economic Indicators (MEI). To obtain our preliminary estimate of GDP growth for a given quarter, we (cid:12)nd the (cid:12)rst monthly issue of the (hardcopy) MEI in which GDP is reported for the relevant quarter and we calculate the implied GDP 4 growth rate. Throughout, the growth rate is de(cid:12)ned as the quarter-over-quarter percent change of real seasonally adjusted GDP (not annualized). The (cid:12)nal growth 5 rateis taken fromtheApril1999 CDof MEI. Thesedata arenottruly(cid:12)nalinthat, for example, base-year and de(cid:12)nitional changes continue inde(cid:12)nitely. To insurethat our (cid:12)nal numbersare at least mature we end the sample for all countries in 1997Q4. FortheUnitedStates, UnitedKingdom,andCanada,preliminarydataareavailable beginning in 1965Q1. For Japan, the starting date is 1970Q1; for Italy and Germany, 1979Q4; and for France, 1987Q4. The German data refer to West Germanyuntil1994Q4 andtoallofGermanyinallsubsequentquarters. Unfortunately, West German real GDP data are no longer included in MEI, so for Germany the (cid:12)nal growth rate is taken from the Haver Analytics Germany database, rather than from the April 1999 MEI CD. As such, the results for Germany should be treated with some caution. 4 Although MEI is published monthly, we are not guaranteed that the (cid:12)rst number published in MEI is really the(cid:12)rst numberever released. It is advantageous tocollect all of our data from a consistent source, such as MEI. Data for a given quarteris usually reported in MEI soon after the endofthatquarter(usuallywithin2or3months). FortheUnitedStates,ourpreliminarydatafor GDP growth is very similar, though not always identical, to that in the Croushore-Stark dataset (CroushoreandStark(1999)). Thesedata,inturn,areobtainedfromtheMay,August,November and February issues of theSurveyof Current Business, for quarters1, 2, 3, and 4, respectively. 5 GNPwas in fact reported in the early years of thesample. 5

4 Predictability of GDP growth revisions 4.1 Bias in Revisions and Other Summary Statistics Table 1 reports summary statistics for the preliminary and (cid:12)nal rates of seasonallyadjusted quarterly real GDP growth for each of the G-7 countries. The table gives both the results over the longest span of data available for each country (top panel) and for the more recent 10 years, 1988Q1-1997Q4, for all countries (bottom panel). For all countries the revisions are large, as reflected in the root mean square error. Consistent with these numbers, the (cid:12)nal annualized growth rate is more than a percentage point di(cid:11)erent from the preliminary at least half the time in these 6 data. Revisions are generally smaller for the shorter, recent sample, but even in this period, the root mean square revision ranges from a low of about one- third of a percentage point for the U.S., Canada, and France, to over three-quarters of a 7 percentage point for Germany, the UK and Japan. In comparing the magnitudes of data revisions across countries, one should bear in mind that some countries, including the United Kingdom, issue their preliminary data much more quickly 8 than others. For thefullsample,themeanrevision toGDP growthis positiveforallcountries except Japan, indicating a general tendency toward pessimism in initial numbers. The bias is quite large in the UK at over one-quarter percentage point (a full point at an annual rate). We also report a t-statistic for testing the forecast e(cid:14)ciency hypothesisthatthemeanrevisioniszero. Throughoutthepaper,allstandarderrors areheteroskedasticityandautocorrelationrobust,usingNewey-Weststandarderrors 9 with a lag truncation parameter of 4. The hypothesis that the mean revision is 6 This share is 0.48 for Canada and over0.5 for all the othercountries. 7OurevidencefortheUKisroughlyconsistentwithwhatBarklem(2000)foundusingsomewhat di(cid:11)erent revision measures. 8 The Federal Reserve Board has recorded the dates of (cid:12)rst announcements on GDP growth over the last 5 years, for the G-7 countries. The average time between the end of the quarter and the (cid:12)rst announcement is 56 days across the G-7 as a whole, but is only 26 days for the United Kingdom and 30 days for theUnited States. 9 Newey and West, 1987. The results are not sensitive to the choice of this lag truncation parameter: indeed it would makevery little di(cid:11)erence if it were set to 0. If we knewthat the(cid:12)nal 6

10 zero is rejected for the U.S., Canada, and the UK. The fact that the results are only signi(cid:12)cant for these 3 countries is in part driven by the fact that these are the countries with the longest sample period. The statistically signi(cid:12)cant bias for the 11 U.S. can also be found in the Croushore-Stark data. During the 1988-97 sub-period, mean revisions are quite small for all countries but the UK and, given the short sample, one cannot reject the hypothesis that the mean is zero. For the UK the mean revision remains large and statistically signi(cid:12)cantlydi(cid:11)erentfromzero. Biasis,ofcourse,thesimplestformofpredictability in revision; we now examine the predictability of the data revision process more comprehensively. 4.2 Forecastable revisions: a preliminary look If the preliminary data are an e(cid:14)cient forecast, it is also the case that they should have lower standard deviations than the (cid:12)nal data: optimal forecasts are less variable than the item forecasted. In Canada, Germany, Japan, Italy, and the UK, however, the ratio of preliminary to (cid:12)nal standard deviations is greater than one. For Japan, Italy and the UK the ratio is above 1.3. Further, if the data revision processonlyincorporatesnews, thenthereshouldbenosystematic relation between the preliminary announcement and the subsequent revision. In fact, one can see a strikingnegative relationshipinthescatter plotsofrevisionsagainstthepreliminary data for several countries, notably Japan, Italy and the UK (Figure 1). For these countries, high preliminary numbers are systematically revised downwards and low data were released m+1 quarters after the preliminary data, then the data revision process would be an m-dependent process, as an implication of forecast rationality. However, our construction of the data does not record the timing of the release of the (cid:12)nal data (which is in any case never truly (cid:12)nal, because of benchmark revisions among other things). Accordingly, in our dataset, forecast rationalitycarriesnospeci(cid:12)cimplicationsfortheautocorrelation ofdatarevisions. Thisis the reason why we use autocorrelation robust standard errors, while noting that this makes little di(cid:11)erence in practice. 10 Unless otherwise stated, we use the5 percent level to judge signi(cid:12)cance and test rejections. 11 In calculations not shown in this paper, we (cid:12)nd that in the Croushore-Stark data, available on thewebsite of thePhiladephia FederalReserveBank, therevision of output growth going from the(cid:12)rst release tothemost recentlyavailable data issigni(cid:12)cantly positive, as istherevision going from the (cid:12)rst release to thesecond release (one quarterlater). 7

preliminary numbers are systematically revised upwards. Thisinformalevidenceagainstforecastrationality isexactlyasonewouldexpect under the noise interpretation of the data revision process. The noise contributes to excessive variance of the preliminary data. Further, unusual observations in the preliminary data tend to be revised toward more normal values as the noise is removed. This informal evidence is con(cid:12)rmed in the next section. 4.3 The Mincer-Zarnowitz test We now turn to the Mincer-Zarnowitz test of forecast rationality{a test of the hypothesis that (cid:11) = (cid:12) = 0 in the regression (2). This regression allows us formally to measure the relation seen in Figure 1. For the full sample, the F-statistic for testing this hypothesis suggests rejection of forecast rationality for every country except France, which has the shortest sample (Table 2, top panel). As one would predict from Figure 1, overwhelming rejections are obtained for Japan, Italy and the United Kingdom. The degree of predictability in data revisions varies substantially across countries. The adjusted R2 ranges from 2 percent in the United States to 62 percent in Italy. For the most recent 40 quarter sample (Table 2, bottom panel), the F-test again rejects forecast rationality for Italy, Germany and Japan, and for the UK at the 10 percent level, but not for the other countries. The explanatory power of the regressionalsoremainsquitelargeforthefourcountrieswherewe(cid:12)ndpredictability. Overall,wetaketheseMincer-Zarnowitzresultsasevidenceofstrongpredictability of GDP revisions in 4 of our countries both in the full and recent samples. For Canada and the U.S., there is weak evidence of predictability in the full sample. In the shorter, recent sample, either the reduced number of observations or, perhaps, improvements in the revision process in these 2 countries leaves little evidence of predictability. We now turn to some robustness checks of and extensions to these basic results. To preview, neither the robustness checks nor the extensions alter 8

much the basic conclusions just illustrated. 4.4 Recursive estimates and outliers Since our estimates are based on a sample that includes very recent data, statistical agencies could not have used our results in the past to generate better preliminary announcements. As a simple check of whether statistical agencies could have done better with data available in the past, we construct recursively adjustedpreliminary th data series. For each quarter, starting in the 20 quarter of the full sample for each country, we ran the Mincer-Zarnowitz regression using only data that was available in that quarter. Next we adjusted the preliminary data for that quarter by adding toitthe(cid:12)ttedrevisionforthatquarterimpliedbytheregression. Franceisexcluded from this exercise due to the short sample. These results require one caveat. In order to run the regressions we needed an assumption about when the (cid:12)nal data become available. We assume that our (cid:12)nal data are released 4 quarters after the preliminary data, which is not literally true. We report the mean square error of the raw preliminary data and the recursively adjusted preliminary data as forecasts of the (cid:12)nal data (Table 3). We also report a Diebold-Mariano (1995) test of the hypothesis that these mean square errors are equal. For all of the countries, except the United States, the real-time recursivelyadjusted preliminary data has smaller mean square error than the original preliminary data. The reduction in mean square error is large for Germany, Italy, Japan, and the UK and is statistically signi(cid:12)cant for Japan and the United Kingdom. The fact that the di(cid:11)erence is not statistically signi(cid:12)cant for Germany and Italy may be due to the relatively short samples for Germany and Italy. Especiallyintheearlyyearsofthesample,manyofthepreliminaryGDPgrowth rates are quite extreme. This is particularly true for the United Kingdom in the 12 1970s. To see if our results are driven by such outliers, we re-ran the regressions 12 Forexample,thepreliminarygrowth ratefortheUnitedKingdomin1971Q1was-4.7percent (not annualized): this was subsequently revised to -1.2 percent. 9

in Table 2, deleting all observations for which the preliminary growth rate deviated 13 from the country-mean by more than 3 percentage points. The results (Table 4) are entirely consistent with the earlier results (Table 2): these outliers are not driving the results. 5 Augmented forecast e(cid:14)ciency regressions Forecastrationality requiresthattherevisionsbeunforecastableusinganydatathat was known at the time the preliminary data were released. To see if added explanatory variables strengthen the evidence against forecast rationality, we augment the basic regression with seasonal dummies and (cid:12)ve variables known at the time that the preliminary data were released: lagged preliminary data, the growth rate of equity prices, a 3-month interest rate, oil price inflation, and a dummy variable for 14 national elections (Canada, Japan, UK, U.S. only). We includethe seasonal dummies and lagged preliminarydata to check for additional simpleforms of dynamics in the original speci(cid:12)cation. We follow Mankiw and Shapiro (1986) in using equity prices and short-term interest rates as business cycle indicators. We add oil prices for the same reason. These business cycle variables may have predictive power if the systematic ine(cid:14)ciencies in the data construction process are a(cid:11)ected by the state of the business cycle. Nordhaus(1975) suggested thatincumbentgovernments might attempttoboost the economy before elections to enhance election prospects. Of course, incumbents might prefer to arti(cid:12)cially boost the economic data, rather than actually stimulating the economy. This would avoid some of the e(cid:14)ciency costs of deviating from (otherwise) optimal policy, and after data revisions the evidence would disappear from the historical record. There have been accusations in the press of this sort of 13 There were 9 such outliers in the United Kingdom, 3 in Japan, 2 each in Germany, Italy and the United States, 1 in Canada and none in France. All but 2 of these outliers occurred prior to 1987. 14 Equity prices are measured by the stock price indices for each country reported in MEI. The 3-monthinterestrateforeachcountryisalsotakenfromMEI.TheoilpriceisthespotWestTexas intermediate crudeprice. 10

manipulation in Japan (New York Times, 2000). We create an election dummy variable that is one if the preliminary number is thelastoneannouncedbeforeanationalelectionandzerootherwise. Weincludethis dummyvariableonlyforCanada,Japan,theUnitedKingdomandtheUnitedStates because for the other countries there were not enough elections within our sample 15 period to obtain meaningful results. For the 4 countries with enough elections, we construct the dummy variable by obtaining the exact dates of elections and of 16 the preceding initial GDP announcements from various national newspapers. If the data for the election quarters are optimistic, then we expect revisions for these quarters to be systematically more negative than at other times. Weemphasizeattheoutsetthattherearetworeasonsotherthanpoliticalmanipulation that the election dummy could besigni(cid:12)cant. First, the prospective election could lead to breaks in economic behavior that are not captured by the methods used in constructing preliminary GDP. For example, suppose that businesses delay marginal investments around the time of elections to wait for the resolution of electoral uncertainty. Methods for estimating a preliminary investment number that do not reflect this break in behavior would lead to optimistic investment estimates. Second, since election timing is endogenous in each country except the U.S., the election dummy could also come in if data-construction biases are correlated with the same variables driving election timing. For example, elections might be called when the economy is robust and the GDP construction methods might be most optimistic at those times. We partially control for this by including our business cycle variables, but these controls are obviously imperfect. 15Withinoursampleperiodtherewere8electionsintheUnitedStates,9intheUnitedKingdom, 10 in Canada and 16 in Japan, but only 3 in France and 5 each in Germany and Italy (although governments collapse notoriously frequently in Italy, these collapses often do not lead to fresh elections). Weare grateful to Deepak Mishra for providing uswith some election data. 16IntheUnitedStates,theelectionswerefertoarepresidentialelections(midtermelectionswere disregarded). For Canada, Japan and the United Kingdom, they are all national parliamentary elections (in either house, in thecase of Japan). 11

5.1 Results for the augmented regressions When we include all the additional regressors (Table 5), forecast rationality is rejected for all 7 countries (as opposed to the 6 rejections reported in Table 2). A number of the added regressors are statistically signi(cid:12)cant for multiple countries, and each regressor is statistically signi(cid:12)cant, at least at the 10 percent level, for at least one country. For the UK and France, oil price inflation is signi(cid:12)cant; for the UK and Japan, the interest rate is signi(cid:12)cant; for Japan and the U.S., equity price growth is signi(cid:12)cant at the 10 percent level. Only in the case of France, however, is there an appreciable increase in explanatorypoweroftheregressionasmeasuredbytheadjustedR2 . ForFrancetheincrease is 0.2; the next largest increase is the U.S. at 0.06. Thus, the vast majority of the predictive power comes from the constant and lagged preliminary number. We also ran the regressions in Tables 5 using only the subsample of data from 1988Q1 on, but omitting the election dummy (because there are too few elections in this subsample). Further, we ran these regressions deleting outliers (as de(cid:12)ned in the construction of Table 4). The results were similar to those in Tables 5, however, and are not shown so as to conserve space. Last, we used the regressions in Table 5 to construct a recursively adjusted preliminary data series (as de(cid:12)ned in theconstruction of Table3), andfoundthatthis hassmaller meansquareerrorthan thatoftherawpreliminarydata. Thisreductioninmeansquareerrorisstatistically signi(cid:12)cant for Japan and the United Kingdom. These results are omitted to save space but are available from the authors. 5.2 Elections For Japan alone, the election dummy is marginally signi(cid:12)cant (signi(cid:12)cant at the 10 percent level). The estimated election e(cid:11)ect is large, as the (cid:12)nal GDP growth rate announced before an election is subsequently revised downwards by about 0.3 percentage points more than for other quarters (thus, the revision is over a full percentage point greater at an annualized rate). 12

Because elections are quite seasonal in many countries|most notably the U.S. th where they all fall in the 4 quarter|we are concerned with possible collinearity 17 between the election dummy variable and the seasonal dummy variables. For example, in the U.S. none of the seasonal dummy variables enters signi(cid:12)cantly, but their presence could make it di(cid:14)cult to detect an independent election e(cid:11)ect. More generally, a joint test of the hypothesis that the seasonal dummies are zero fails to reject for every country but France and the UK (Table 5, (cid:12)nal row). Thus,inTable6,wereporttheregressionfromTable5withthequarterlydummy variables removed. Very few coe(cid:14)cients change greatly in value or in sign(cid:12)cance. As for the election dummy, removing the quarterly dummies does not a(cid:11)ect the conclusion that the election dummy is insigni(cid:12)cant in Canada, the UK, or the U.S. For Japan, the magnitude of the election e(cid:11)ect remains largely unchanged and the election e(cid:11)ect is now signi(cid:12)cant at the 5 percent level. As weemphasizedabove, thereisnopresumptionthatthemarginally signi(cid:12)cant e(cid:11)ect in Japan is due to political manipulation. We found some evidence that this e(cid:11)ect is most prominent in the investment component of GDP, suggesting that it could be due to the sort of time-shifting of investment around elections discussed 18 above. It should also be noted that there are considerably more elections in our Japanese sample than in any other country. Thus, our tests have much more power to detect an electoral e(cid:11)ect. 17 In Canada, there were 1, 3, 2 and 4 elections in quarters 1-4, respectively. In the United Kingdom, there were 2, 6, 0 and 1 elections, and in Japan, there were 1, 3, 7 and 5 elections, in quarters1-4, respectively. 18 We re-ran the regression in Table 5, replacing the preliminary and revised output growth data bythe analogous investment growth data (private(cid:12)xed investment),collected from thesame sources. The coe(cid:14)cient on the election dummy was signi(cid:12)cantly negative in this regression. We also ran these regressions for consumption, government spending, export and import growth data, but the coe(cid:14)cients on the election dummy were not signi(cid:12)cant at any conventional signi(cid:12)cance level in any of these regressions. 13

6 Conclusions Revisions to GDP announcements are quite large in all the G-7 countries. The magnitude has fallen some in recent years, but remains large. In Canada, the UK, and the U.S., the preliminary announcements have been signi(cid:12)cantly pessimistic. In several countries|the UK, Italy, Japan|the revisions are highly predictable. In these countries, about half the variability of revisions can be accounted for by data available at the time of the preliminary announcement. For the other countries, there is some evidence of predictability, but the measured degree of predictability is rather modest. Thus, for these countries it seems that revisions primarily reflect news not available at the time of the preliminary announcement. When we do (cid:12)nd predictability of revisions, it is mostly due to the predictive power of the preliminary number: extreme values, large or small, in the preliminary growth rate tend to be revised toward the mean. This is exactly what one would expectunderthenoiseviewofrevisionsinwhichrevsionsremovemeasurementerror from earlier announcements. This paper has potentially important implications for macroeconometric work. Several recent papers have pointed out the problems with using fully revised data; such work rests on the assumption that agents perfectly anticipate revisions. A corrective thathasbeenusedistore-dotheworkusingpreliminarydata. Ofcourse, this may rest on the assumption that revisions are completely unpredictable. This assumption is a reasonable approximation for theU.S., butfor Italy, Japan, and the UK, revisions are quite predictable and neither extreme assumption is appropriate. Analysts should probably use some sort of recursively adjusted data, such as those constructed here. 14

References Barklem, A. (2000): Revisions Analysis of Initial Estimates of Key Economic Indicators and GDP Components, Economic Trends, No. 556, pp.31{52. Croushore, D. and T. Stark (1999): A Real-Time Dataset for Macroeconomists, Journal of Econometrics, forthcoming. De Jong, P. (1987): Rational Economic Data Revisions, Journal of Business and Economic Statistics, 5, pp.539-548. Diebold, F.X. and R.S. Mariano (1995): Comparing Predictive Accuracy, Journal of Business and Economic Statistics, 13, pp.253-263. Diebold, F.X. and G. Rudebusch (1991): Forecasting Output with the Composite Leading Index: A Real-Time Analysis, Journal of the American Statistical Association, 86, pp.603-610. EconomicResearchInstitute(2000): TheRecentOpinionsonJapan’sGDPFigures and Our Approach, Economic Planning Institute, mimeo (June 9). Evans,C.L.(1998): Real-TimeTaylorRulesandtheFederalFundsFuturesMarket, Federal Reserve Bank of Chicago Economic Perspectives, 22, pp.44-55. Howrey, E.P. (1978): The Use of Preliminary Data in Econometric Forecasting, Review of Economics and Statistics, 60, pp.193-200. Howrey, E.P. (1984): Data Revision, Reconstruction and Prediction: An Application to Inventory Investment, Review of Economics and Statistics, 66, pp.386- 393. Kavajecz, K. and S. Collins (1995): Rationality of Preliminary Money Stock Estimates, Review of Economics and Statistics, 77, pp.32-41. Mankiw, N.G., D.E.RunkleandM.D. Shapiro(1984): ArePreliminaryAnnouncements of the Money Stock Rational Forecasts?, Journal of Monetary Economics, 14, pp.15-27. Mankiw, N.G. and M.D. Shapiro (1986): News or Noise: An Analysis of GNP Revisions, Survey of Current Business, May 1986, pp.20-25. Mariano, R.S. and H. Tanizaki (1995): Prediction of Final Data with Use of Preliminary and/or Revised Data, Journal of Forecasting, 14, pp.351-380. Milbourne, R.D. and G.W. Smith (1989): How Informative are Preliminary AnnouncementsoftheMoneyStockinCanada?,CanadianJournalof Economics, 22, pp.595-606. Mincer, J. and V. Zarnowitz (1969): The Evaluation of Economic Forecasts in J. Mincer (ed.), Economic Forecasts and Expectations, NBER, New York. 15

NewYorkTimes(2000): JapanAssailedforOmittingDatainGrowthCalculations, May 24, p. C1. Newey, W.K.andK.D.West(1987): ASimplePositiveSemide(cid:12)nite,Heteroskedasticity and Autocorrelation Consistent Covariance Matrix, Econometrica, 55, pp703-708. Nordhaus, W. (1975): The Political Business Cycle, Review of Economic Studies 42, pp.169{190. Orphanides,A.(1998): MonetaryPolicyRulesBasedonReal-TimeData, Boardof Governors of the Federal Reserve System, Finance and Economics Discussion Paper, forthcoming American Economic Review. Orphanides, A. (2000): The Quest for Prosperity Without Inflation, European Central Bank Working Paper. Robertson, J.C. and E.W. Tallman (1998): Data Vintages and Measuring Forecast Model Performance, Federal Reserve Bank of Atlanta, Working Paper. Reed, G. (2000): How the Preliminary Estimate of GDP is Produced, Economic Trends, No. 556, pp.53{61. 16

Table 1: Summary of announcements and revisions for two samples Canada France Germany Italy Japan UK US start-1997Q4 Revision Mean 0.18 0.01 0.11 0.11 -0.07 0.3 0.1 t-stat 2.97 0.11 1.36 1.45 -1.02 4.53 2.57 root mean square 0.81 0.36 0.97 0.85 0.83 1.14 0.53 mean absolute 0.58 0.28 0.69 0.59 0.64 0.84 0.4 (cid:12)nal mean growth 0.86 0.49 0.48 0.48 0.91 0.57 0.76 prelim mean growth 0.69 0.48 0.37 0.37 0.98 0.27 0.65 1988Q1-1997Q4 Revision Mean 0.01 -0.02 0.2 -0.01 0.05 0.22 0.02 t-stat 0.21 -0.28 1.61 -0.19 0.61 2.08 0.41 root mean square 0.34 0.34 1.08 0.52 0.8 0.73 0.33 mean absolute 0.26 0.27 0.74 0.36 0.63 0.4 0.26 (cid:12)nal mean growth 0.53 0.47 0.57 0.41 0.66 0.52 0.62 prelim mean growth 0.52 0.49 0.37 0.42 0.61 0.3 0.6 Notes: The(cid:12)nalandpreliminarynumberarebothquarter-over-quartergrowthrates in percent. The revision is (cid:12)nal minus preliminary. The t-statistics are based on autocorrelation and heteroskedasticity consistent standard errors and are for the hypothesis that the mean is zero. The sample ends in 1997Q4 for all countries. The full sample begins in 65Q1, 87Q4, 79Q4, 79Q4, 70Q1, 65Q1, 65Q1, for the seven countries, respectively. The results for Germany are tentative, because di(cid:11)erent databases were used to calculate preliminary and (cid:12)nal growth rates. 17

Table 2: Mincer-Zarnowitz Regression in two samples Canada France Germany Italy Japan UK US Full sample constant 0.44 0.12 0.28 0.34 0.33 0.44 0.17 4.94 1.32 2.9 4.88 4.25 6.11 2.28 prelim -0.39 -0.24 -0.48 -0.64 -0.41 -0.52 -0.1 -4.8 -1.78 -4.29 -6.54 -7.18 -8.55 -1.16 F 26.7 3.3 18.7 49.4 57.4 83.3 7.6 p-val 0 0.2 0 0 0 0 0.02 R(cid:22)2 0.27 0.07 0.4 0.62 0.42 0.52 0.02 1988Q1{1997Q4 constant 0.06 0.1 0.42 0.17 0.28 0.41 0.09 1.01 1.15 2.81 1.65 2.4 2.18 0.82 prelim -0.09 -0.23 -0.61 -0.44 -0.38 -0.65 -0.12 -1.24 -1.71 -4.18 -2.62 -3.59 -2.24 -1.01 F 1.8 3 18.6 7.3 12.9 5.2 1.1 p-val 0.41 0.23 0 0.03 0 0.07 0.57 R(cid:22)2 0 0.08 0.56 0.28 0.32 0.46 0 Notes: See the notes to Table 1. In each speci(cid:12)cation the dependent variable is the revision in percent; all independent variables are listed in column 1. The number in smaller type under the coe(cid:14)cient estimate is a t-statistic for the hypothesis that the coe(cid:14)cient is zero. The row labelled F is an F test of the hypothesis that all coe(cid:14)cients are zero; the p-value for this test is given in the next row. 18

Table 3: MSE of preliminary data and adjusted preliminary data as forecasts of (cid:12)nal data Canada Germany Italy Japan UK US MSE (Raw) 0.6413 1.1641 0.338 0.5879 1.2973 0.2877 MSE (Adjusted) 0.4656 0.7389 0.2321 0.3816 0.6203 0.3091 Diebold-Mariano 1.14 1.55 1.03 2.31 2.62 -1.56 Notes: This table shows the mean square error of the raw preliminary data and of the recursively adjusted preliminary data (adjusted using the Mincer-Zarnowitz regression, as discussed in the text). The Diebold-Mariano statistic measures the signi(cid:12)cance of the di(cid:11)erence between the MSE (Raw) and MSE (Adjusted). It is asymptotically normally distributed. Table 4: Mincer-Zarnowitz regression, full sample, observations with extreme preliminary growth rates deleted Canada France Germany Italy Japan UK US constant 0.35 0.12 0.21 0.34 0.44 0.43 0.09 4.47 1.32 2.71 4.41 5.28 5.48 2.23 prelim -0.28 -0.24 -0.35 -0.7 -0.5 -0.55 -0.01 -4.51 -1.78 -5.13 -7.35 -7.79 -7.14 -0.24 F 22.7 3.3 27.5 55.5 62.7 53.7 7 p-val 0 0.2 0 0 0 0 0.03 R(cid:22)2 0.14 0.07 0.24 0.62 0.43 0.45 -0.01 Notes: see the notes to Table 2. The Table reports the same results as Table 2, except that observations with outlier preliminary data (as de(cid:12)ned in the text) are deleted. 19

Table 5: Forecast E(cid:14)ciency Regression, full sample Canada France Germany Italy Japan UK US constant 0.51 0.51 0.08 -0.16 0.86 0.53 0.10 2.59 1.79 0.18 -0.29 3.42 2.67 0.51 prelim -0.46 -0.12 -0.45 -0.71 -0.44 -0.51 -0.08 -6.27 -1.70 -3.68 -8.91 -7.48 -9.08 -0.94 prelim(-1) 0.17 0.30 0.08 0.06 0.02 0.04 -0.05 2.31 4.21 1.41 1.81 0.40 0.67 -0.85 q1 -0.11 -0.12 -0.32 -0.01 -0.10 0.54 0.20 -0.68 1.01 -1.03 -0.12 -0.49 2.73 1.49 q2 0.12 -0.14 -0.20 0.09 0.11 0.50 0.10 0.78 -1.45 -1.21 0.55 0.73 2.73 0.79 q3 -0.18 -0.26 -0.29 -0.10 0.04 0.42 -0.13 -1.09 -3.17 -1.61 -0.71 0.30 2.78 -0.88 oil 0.35 -2.01 -2.19 0.28 -0.27 -2.66 -0.34 0.38 -2.68 -1.51 0.36 -0.27 -2.85 -0.62 stock ret 0.01 -0.27 0.24 0.34 -0.42 0.00 0.01 0.80 -1.45 0.86 1.24 -1.82 0.40 1.72 interest -0.01 -0.03 0.03 0.02 -0.07 -0.05 0.00 -1.15 -1.66 0.64 0.66 -2.08 -2.36 0.29 election -0.08 -0.27 0.01 -0.13 -0.37 -1.71 0.02 -0.90 F 76.1 119.5 44.6 154.2 100.0 507.2 28.9 p-val 0.00 0.00 0.00 0.00 0.00 0.00 0.00 R(cid:22)2 0.32 0.27 0.39 0.65 0.45 0.59 0.08 F(cid:3) (seas) 5.70 10.73 4.01 3.25 1.14 11.43 7.05 p-val 0.13 0.01 0.26 0.36 0.77 0.01 0.07 Notes: See the notes to Table 2. The row labelled F(cid:3) (seas) is for a joint test of the joint hypothesis that the 3 seasonal dummy variables are zero in the regression in the top panel. 20

Table 6: Forecast E(cid:14)ciency Regression without Quarterly Dummies, full sample Canada France Germany Italy Japan UK US constant 0.47 0.43 -0.12 -0.19 0.88 0.92 0.13 2.86 1.48 -0.29 -0.33 3.82 4.86 0.82 prelim -0.47 -0.18 -0.46 -0.71 -0.46 -0.54 -0.10 -6.18 -2.24 -4.01 -8.70 -7.91 -8.84 -1.05 prelim(-1) 0.18 0.33 0.09 0.07 0.03 0.04 -0.04 2.41 5.57 1.55 2.20 0.59 0.68 -0.66 oil 0.21 -2.00 -1.98 0.35 -0.27 -2.69 -0.37 0.21 -2.72 -1.52 0.43 -0.26 -3.25 -0.59 stock ret 0.01 -0.29 0.25 0.34 -0.42 0.01 0.02 0.64 -1.57 0.91 1.22 -1.88 1.16 2.06 interest -0.01 -0.03 0.03 0.02 -0.07 -0.05 0.00 -1.14 -1.68 0.62 0.69 -2.14 -2.42 0.26 election 0.01 -0.32 -0.17 -0.01 0.06 -2.15 -0.79 -0.04 F 69.4 53.1 39.3 122.2 99.3 248.8 18.5 p-val 0.00 0.00 0.00 0.00 0.00 0.00 0.01 R(cid:22)2 0.32 0.26 0.40 0.66 0.45 0.56 0.03 Notes: see the notes to Table 2. 21

−5 −2.5 0 2.5 5 −5 −2.5 0 2.5 5 revision to growth rate −5 −2.5 0 2.5 5 −5 −2.5 0 2.5 5 Canada Figure 1. Preliminary GDP Growth Rates and Revisions revision to growth rate −5 −2.5 0 2.5 5 −5 −2.5 0 2.5 5 France revision to growth rate −5 −2.5 0 2.5 5 −5 −2.5 0 2.5 5 Germany revision to growth rate −5 −2.5 0 2.5 5 −5 −2.5 0 2.5 5 Italy revision to growth rate −5 −2.5 0 2.5 5 −5 −2.5 0 2.5 5 Japan revision to growth rate −5 −2.5 0 2.5 5 −5 −2.5 0 2.5 5 preliminary growth rate UK revision to growth rate preliminary growth rate US