Boxes Box 12.1 Compilation and Revision Schedule: An Example...16 Box 12.2 Presentation of Revisions: An Illustration based on Country Practices...

Transcription

1 UPDATE OF QUARTERLY NATIONAL ACCOUNTS MANUAL: CONCEPTS, DATA SOURCES AND COMPILATION 1 CHAPTER 12. REVISIONS 2 Table of Conens 1. Inroducion User Requiremens and Resource Consrains Waves of Source Daa and Relaed Revision Cycles The Compilaion and Release Schedule Communicaions of Revisions Revisions Analysis for Assessing QNA Reliabiliy Summary of Key Recommendaions Annex 1 Revisions Indicaors Bibliography Boxes Box 12.1 Compilaion and Revision Schedule: An Example...16 Box 12.2 Presenaion of Revisions: An Illusraion based on Counry Pracices...17 Examples Example 12.1 Real-Time Daabase...18 Example 12.2 Revisions Indicaors The QNA manual is being updaed by he IMF Saisics Deparmen. For more informaion on he updae, please visi he websie hp:// 2 Draf posed for commens in Ocober 2015.

2 2 CHAPTER 12. REVISIONS Quarerly naional accouns are subjec o revisions. To saisfy imeliness requiremens, he firs esimae of a quarer generally relies on an incomplee and parial daase. When more comprehensive source daa are made available, previously published esimaes mus necessarily be revised. To avoid criicism from users, revisions should be explained and well communicaed. This chaper describes how o design a sound, coordinaed, and ransparen revisions policy of he quarerly naional accouns. Imporan ingrediens for a sound revisions policy are a compilaion and release schedule; an advance release calendar; and a communicaions sraegy. Furhermore, a framework for conducing revisions analysis should be developed o measure and assess he reliabiliy of quarerly esimaes. 1. INTRODUCTION 1. Revisions are an essenial par of good quarerly naional accouns (QNA) compilaion pracice. Revisions are necessary o incorporae improvemens in source daa, mehods, inernaional sandards, and classificaions ha are coninuously made available. They provide users wih daa ha are as imely and accurae as possible. Resource consrains, in combinaion wih user needs, cause ension beween he imeliness of published daa on he one hand and reliabiliy, accuracy, and comprehensiveness on he oher hand. To reduce his ension, ypically, preliminary daa are compiled on a imely basis; laer, revised esimaes are produced when more and beer source daa become available. Good managemen of he process of revisions requires he exisence of a well-esablished and ransparen revision policy. 2. I is imporan o emphasize ha revisions are conduced for he benefi of users, namely, o provide users wih daa ha are as imely and accurae as possible. Revisions provide he possibiliy o incorporae new and more accurae informaion, and hus o improve he accuracy of he esimaes, wihou inroducing breaks in he ime series. Alhough repeaed revisions may be perceived as reflecing negaively on he rusworhiness of official saisics, delaying he incorporaion of new daa in he published esimaes may increase he magniude of laer revisions (in paricular, if hese go in he same direcion). Furhermore, no passing on known revisions reduces he acual rusworhiness of daa even more because he daa do no reflec he bes available informaion, and he public may know his or find his ou (for insance, he public may wonder why a revision in he monhly producion index is no refleced in he QNA). 3. I is essenial o revise preliminary QNA daa o ake accoun of new and beer informaion for he reasons saed above. Furher, aemping o avoid revisions by producing accurae bu very unimely, and hus less useful, daa may no make he bes use of he informaion available. If he official QNA compilers fail o serve users needs, oher

3 3 organizaions may compile heir own esimaes, resuling in confusion from conflicing esimaes o he poin ha many users may consider he official daa irrelevan. Obviously, ha will resul in reduced presige and respec for he official QNA compilers. 4. Revisions o pas daa are no wihou poenial problems and may draw criicism if no properly handled. Revisions o pas daa are inconvenien o users because hey enail revisions o heir daabases and applicaions. More imporan, frequen revisions paricularly o daa for he mos recen periods may cause users o feel uncerain abou he curren economic siuaion and hus uncerain abou wha policy acions should be aken. Some of his uncerainy may be unavoidable and merely reveal he fac ha he informaion base for he esimaes for he mos recen periods is limied and hus ha he daa should be used wih care. Some of he uncerainy, however, may be caused unnecessarily by he way he revisions are carried ou or presened. On he oher hand, unjusified differences beween naional accouns esimaes and heir source daa may cause users o doub he compeence of he naional accouns compilers wih serious and jusified criicism of he naional accouns daa as a resul. 5. To deal wih he issues surrounding revisions and o avoid unnecessary criicism, a well-designed and carefully managed revision policy is needed. Essenial feaures of a welldesigned revision policy are predicabiliy and openness, advance noice of causes and effecs, and explanaion, as well as easy access o sufficienly long ime series of revised daa. This chaper elaboraes on he elemens ha make for a well-esablished revision policy. 2. USER REQUIREMENTS AND RESOURCE CONSTRAINTS 6. The rade-off beween imeliness on he one hand and accuracy and reliabiliy on he oher is caused by a conflic beween differen user requiremens in combinaion wih limiaions in saisical resources. Naional accouns daa are used for muliple purposes ha have parly conflicing requiremens. To allow correcive policy acions o be aken in ime, policymakers and oher users need a coheren, comprehensive, and reasonably accurae picure of he curren economic siuaion ha is as up-o-dae as possible. For oher purposes, such as ime series and srucural analysis of pas evens, users require long ime series of very deailed annual, or quarerly, naional accouns daa. Finally, users are ineresed in boh he period-o-period raes of change in he series and heir levels. The resources available for saisical purposes, however, are limied. Collecion of sufficienly accurae and deailed source saisics is ime-consuming and expensive boh for he saisical office and for he respondens, and compilaion of comprehensive, accurae, and deailed naional accouns is in iself ime-consuming and expensive. Also, frequen collecion of comprehensive and deailed daa may impose an unwarraned burden on respondens, who hemselves may no even have such daa on a imely and shor-erm basis. 7. As a resul, only a limied se of monhly or quarerly source daa ypically is available on a very imely basis. More deailed and more comprehensive monhly or

4 4 quarerly source saisics ypically become available on a less imely basis, while he mos deailed, comprehensive, and reliable source daa may be annual or less frequen daa ha become available wih varying delays long afer he reference year. To provide sufficienly reliable benchmark daa, many counries conduc periodic benchmark censuses, collecing very deailed and reliable annual daa every five or en years. These are ofen linked o periodic compilaion of supply and use ables. The monhly and quarerly daa commonly are based on smaller samples and less complee sample frames han he corresponding annual daa. Finally, he annual daa may be based on audied business accouns hrough comprehensive quesionnaires ha faciliae a horough checking and ediing of he repored daa, while he quarerly daa may be colleced using simpler quesionnaires ha allow less exensive checking and ediing. 3. WAVES OF SOURCE DATA AND RELATED REVISION CYCLES 8. Quarerly accouns are subjec o hree waves of saisical source daa ha become available. Each of hese waves may lead o revisions of earlier esimaes and he incorporaion of more deails in he published accouns. In accordance, hree revision cycles may be disinguished: A quarerly revision cycle is deermined by he evoluion of he shor-erm saisics as used in he QNA; An annual revision cycle is caused by incorporaion of annual source daa or annual naional accouns (ANA) esimaes based on a separae ANA compilaion sysem ino he QNA hrough benchmarking. Finally, a periodic major revision cycle originaes from incorporaing daa from periodic benchmark censuses, revised inernaional guidelines, and oher changes ha canno be incorporaed on a coninuous basis because of resource consrains. Revisions may, of course, also be caused by compilaion errors, which need o be correced when found. 9. The evoluion of shor-erm saisics used in he QNA may cause revisions for wo reasons: (a) correcions or changes in specific shor-erm source daa and (b) incorporaion of addiional, somewha less imely, shor-erm daa. Changes in shor-erm source daa can be caused by lae responses received afer iniial publicaion of source saisics and by he use of pre-published daa ha are sill open o change. To increase he imeliness of he QNA, he firs esimaes may have o be based on an incomplee se of shor-erm source daa. Monhly and quarerly source daa commonly become available wih varying delays. Thus, when preparing he firs esimaes, only daa for wo monhs of he las quarer may be available for some series, while daa may be missing alogeher for oher series. To fill hese source daa gaps, provisional esimaes are generally based on simple rend exrapolaion or on

5 5 alernaive indicaors ha are more imely bu less reliable. During he course of he curren year, hese provisional esimaes mus be revised o incorporae more and beer daa as he less imely shor-erm source saisics become available. 10. An early (or flash ) esimae of he quarerly GDP should always be revised during he same quarer o incorporae more complee and updaed shor-erm indicaors ha become available afer is publicaion. An early esimae serves a very crucial need for policymakers, who necessiaes a promp measuremen of he curren economic developmens. However, imely bu unreliable official esimaes may lead o wrong decisions. Subsanial changes o an early esimae should be inroduced as soon as possible, because policy decisions may change in accord wih he revised esimae. Counries producing early esimaes should release a leas one more release of he same quarer before he nex quarer commences. The iming of he early and revised esimaes should be decided on he basis of he iming of arrival of he shor-erm indicaors. 11. Incorporaion of more reliable annual daa ino he quarerly esimaes implies several revisions o he QNA esimaes over ime for wo reasons. Firs, he annual daa hemselves may be revised. Second, for echnical reasons, he benchmarking procedure will resul in revisions o quarerly daa for earlier years in addiion o he year(s) wih new annual daa. As explained in Chaper 6, hese addiional revisions o pas esimaes are needed o avoid inroducing breaks (he sep problem ) in he QNA ime series beween successive pairs of years. Benchmarking of QNA on more reliable annual daa has he advanage of conveying he accuracy and reliabiliy of he annual daa o he QNA and allows for a degree of comprehensiveness ha he shor-erm source daa by hemselves do no admi. Annual source daa may become available hroughou he year or clusered around a few imes of he year. 12. Periodic major revisions may be needed o he complee quarerly and annual ime series or o a large par of he ime series. Over ime, periodic benchmark censuses may be conduced, new ypes of annual source daa may become available, and improved compilaion mehods may be developed, all indicaing a need for level adjusmens. In addiion, inernaional guidelines are periodically revised. To inroduce hese improvemens wihou creaing breaks in he quarerly and annual ime series, he complee ime series or a large par of he ime series mus be revised a he same ime. Ideally, his should be done on a coninuous basis, series by series; however, resource consrains ofen do no permi such a frequen back-casing approach. Simplified raio-based back-casing echniques may help in dealing wih his problem. When such major revisions are released, i is desirable o provide o he user coninuous ime series of key QNA aggregaes (such as he quarerly GDP) of a leas five years of lengh.

6 6 4. THE COMPILATION AND RELEASE SCHEDULE 13. A crucial par of a well-esablished and ransparen revision policy is devising an appropriae compilaion and release schedule. When esablishing a compilaion and release schedule, i is imporan o decide (a) how imely he iniial quarerly esimaes should be; (b) how frequen new quarerly source daa should be incorporaed; (c) how early and how frequen annual source daa should be incorporaed; (d) how frequen regular major revisions should be conduced; and (e) how long he revisions period should be exended. 14. Major elemens in deermining he compilaion and release schedule are (a) iming of arrival of major daa sources, and he source daa revision policy; (b) iming of preparaion of imporan economic poliical documens; (c) aiudes oward he rade-off beween imeliness and accuracy, as well as oward size and frequency of revisions; (d) disseminaion modes; and, finally, (e) workloads and he design of he naional accouns compilaion sysem. 15. To minimize he number of revisions needed wihou suppressing informaion, i is advisable o coordinae saisical aciviies. The revision schedule is, or should be, largely driven by he arrival of source daa, and coordinaing heir arrival would subsanially help reduce he number of revisions needed. Tying inroducion of new conceps and mehods, or new inernaional guidelines, o he ime of oher planned revisions would also help reduce he number of revisions. Alhough he iming of censuses and new surveys may no be a he discreion of naional accounans, hey may have a srong say in his, and hey are well advised o use heir influence o achieve maximum consisency wih heir revision policy. 16. Accoun needs o be aken of he coordinaion of QNA wih relaed economic policy documens, such as he general governmen budge and oher imporan documens relaed o he parliamen s or legislaure s budge discussions. Release of new esimaes shorly afer he governmen budge has been presened or in he mids of a budge debae may cause problems (alhough his should no change he release schedule once i has been fixed). 17. As discussed in he previous secion, care should be aken o preven ha iniial esimaes for a quarer are prepared and released oo early. Improved imeliness could require use of a higher proporion of incomplee source daa, resuling in an unaccepable reducion in he accuracy of he esimaes and larger revisions. The informaion conen of esimaes based on very incomplee source daa may be limied and, in some cases, more misleading han informaive. In hose cases, he users would be beer served by less imely iniial esimaes for a quarer. 18. Finally, he design of he naional accouns compilaion sysem has imporan implicaions for how frequenly i is possible and appropriae o incorporae new source daa. Large and complicaed compilaion sysems wih deailed and exensive balancing and reconciliaion procedures (e.g., based on quarerly or annual compilaion of inegraed supply

7 7 and use ables and a complee se of inegraed secoral accouns) make i cosly o incorporae new source daa very frequenly. 19. Timeliness of release of he iniial esimaes for a quarer varies grealy from counry o counry, mainly reflecing differen perspecives on he imeliness-accuracy-revision radeoff. Early releases of QNA daa among saisically advanced counries come wihin 45 days afer he reference quarer. In oher counries, a more common release ime for he iniial esimaes is around wo o hree monhs afer he end of he quarer. 3 To provide very early annual esimaes, some counries release heir iniial esimaes earlier afer he end of he fourh quarer han for oher quarers. Correspondingly, here is ypically a shif of focus in he presenaion from he esimaes of he quarers o he esimaes for he full year. While he main focus may be on he esimaes for he full year, he fourh-quarer daa need o be published in heir own righ because failing o do so will cause users who need inegraed annual and quarerly daa o wrongly derive he fourh quarer as he difference beween he annual oal and he sum of he hree previously published quarers. If he iniial esimaes for he fourh quarer are released earlier han for oher quarers, i is preferable o highligh he lower qualiy of he fourh-quarer esimae, for example, by noing is revisions in previous years and he specific shorcomings in he daa used. 20. How frequenly new quarerly source daa are incorporaed varies. Counries ha release heir iniial esimaes wihin a monh afer he reference quarer ypically release revised and more deailed esimaes shorly hereafer. These early esimaes are ofen revised once or wice in he firs quarer afer he reference quarer. The esimaes may be open o quarerly revisions hereafer. A more common pracice, followed by counries ha are less imely in releasing heir iniial esimaes, is o revise he esimaes quarerly linked o he preparaion and release of he iniial esimaes for he following quarers. To reduce he number of revisions, i may be emping o allow he esimaes o be revised only once during he ongoing year. However, emporarily suppressing informaion may resul in larger revisions laer. Suppression of informaion may also someimes be echnically difficul o implemen and hus may resul in compilaion errors. The preferred pracice is o le all esimaes be open o revision during he ongoing year. This is paricularly relevan for seasonally adjused daa, which may change subsanially jus wih he addiion of a new observaion in he series (see Chaper 8 for echnical deails on revisions due o seasonal adjusmen). 21. Annual source daa can be incorporaed ino he QNA esimaes eiher series by series, when he new annual source daa for a series become available, or simulaneously for all series. The choice mainly depends on he design of he ANA and QNA compilaion sysems. The series-by-series approach has he advanage of allowing new annual 3 The Special Daa Disseminaion Sandard (SDDS) specifies imeliness for he iniial QNA esimaes a hree monhs afer he end of he quarer.

8 8 informaion o be incorporaed in as imely a manner as possible. Some counries compile heir quarerly and annual esimaes using basically he same ime-series-oriened compilaion sysem ypically wihou deailed and exensive balancing and reconciliaion procedures making his approach he naural choice. Mos counries use a separae sysem for compiling heir annual esimaes, however, which makes i naural o filer he annual source daa hrough he annual accouning sysem before incorporaing he informaion ino he QNA esimaes. In hose circumsances, o avoid inconsisencies beween quarerly and annual accouns, he second approach may be he naural choice. Some counries use a combinaion of he wo approaches. In any case, because he QNA sysem is an accouning sysem highly inerconneced, compilers should make sure ha changes due o new or updaed annual source daa be refleced consisenly across QNA series. 22. Counries wih an independen ANA compilaion process ypically revise heir annual esimaes from wo o four imes before he books are closed unil a major revision is underaken. These regular revisions o he annual esimaes are normally underaken once a year, alhough a few counries conduc hem more frequenly. The iming wihin he year of hese annual revisions varies widely. The emphasis is ypically on providing accurae and deailed daa for srucural analysis, wih less emphasis on imeliness. They are nearly always more deailed han he QNA and may encompass a more complee se of he inegraed economic accouns, including supply and use ables. All hese feaures make back-casing a demanding ask and hus resric he frequency wih which level adjusmen originaing from new daa sources and new mehods can be incorporaed. 23. Box 12.1 gives an illusraion of a possible compilaion and release schedule followed by counries wih independen ANA compilaion sysems. In his example, he annual accouns are revised only once, bu in many counries he annual accouns are revised several imes before hey are declared final. These subsequen revisions of he ANA should also be pu hrough in he QNA so ha he number of revisions of QNA evenually depends on he number of revisions of he ANA. If a major overhaul of he ANA sysem is performed laer, i should also be pu hrough in he QNA ime series. I should be noed ha in he benchmarking procedures recommended in his manual, revisions of pas years will also necessiae revisions in he quarers of laer years, including he quarers of he curren year. 24. An imporan elemen of he revisions policy is o esablish a revisions period, namely he number of quarers ha should be revised a each revision sage. The revisions period varies depending on he ype of revisions. For revisions due o quarerly and annual sources, he revisions period of QNA esimaes should cover a leas he quarers and years wih revisions o he source daa. As menioned earlier, he applicaion of benchmarking and seasonal adjusmen echniques may require exending he revisions period backward for avoiding seps beween unrevised and revised periods. For benchmarked daa, he QNA revisions period should comprise (a leas) wo years prior he firs revised annual daa; revisions o more disan years end o be small and negligible and could be more difficul o explain o users. For seasonally adjused daa, he revisions period should be longer han he

9 9 revisions period of unadjused daa o accommodae changes in he esimaed seasonal facors (see Chaper 8 for specific deails on revisions policy of seasonally adjused daa). In he case of major revisions, he enire QNA series should be revised as far back as possible. The saring period of he new series should be communicaed clearly, so ha he user is aware ha previously published series for earlier periods are no direcly comparable wih he new series. 25. Someimes, a change in he revisions policy may be required o respond o unforeseen circumsances. For example, a major misake in he reamen of source daa should be fixed as soon as possible, possibly wih an exraordinary QNA release no planned in he calendar. During periods of srong changes in he economy, i may be necessary o increase he frequency of revisions o incorporae new and updaed quarerly source daa as soon as hey become available. Unexpeced changes in he source daa may also jusify a change in he QNA esimaes, or an exension of he revisions period. All of hese changes are warraned o improve he accuracy of he curren esimaes for he benefi of users. However, users may misinerpre changes in he revisions policy and quesion he inegriy of he saisics agency. To avoid such risk, i is exremely imporan o develop a communicaions sraegy of QNA revisions. Elemens of a clear and effecive communicaions policy are deailed in he nex secion. 5. COMMUNICATIONS OF REVISIONS 26. To inform users and avoid unmeried criicism, revisions should be communicaed in a clear and ransparen way. A number of imporan elemens should be aken ino accoun when communicaing revisions o he users: Give advance noice of revisions o he naional accouns daa. Advance communicaion is criical when he revision is expeced o produce significan changes o he levels and raes of change of he quarerly GDP and oher macroeconomic aggregaes. Compilers should proacively oureach o he users before, during, and afer any benchmark revisions of he QNA. Users should be noified in advance abou major revisions wih significan impac on he curren esimaes. For planned revisions, an advance noice should be given in he QNA press release of he quarer preceding he revision; Provide sufficienly long, consisen ime series. A leas five years of coninuous quarerly series for he mos imporan QNA aggregaes should always be available o he public. Five years of daa is ypically a minimum requiremen for macroeconomic or forecasing models. Revised ime series should be accessible wih a sufficien level of deail, so ha users can appreciae he differences in scope and coverage wih he previously published esimaes;

10 10 Publish well known release daes hrough an advance release calendar, as prescribed by he IMF s Special Daa Disseminaion Sandard (SDDS) and General Daa Disseminaion Sysem (GDDS); Provide elemens o appreciae he reliabiliy of he esimaes and he degree of poenial fuure revisions, possibly hrough a revisions analysis as explained in secion 6; Provide deailed daa in an easily accessible forma (e.g., srucured query from a daa warehouse, or pre-formaed spreadshees); Publish ables showing he revisions o he daa wih accompanying ex explaining heir causes. 27. Users should be properly informed of he qualiy of he esimaes and he degree of revisions o expec on predeermined daes in he fuure. Properly informing users of he qualiy of he esimaes involves giving hem candid and easily available documenaion of sources and mehods for he differen versions of he quarerly esimaes, clearly showing he main flows of source daa leading o revisions. When releasing revised esimaes, bes pracice is o simulaneously publish aricles summarizing he main revisions and heir causes since he previous release (see Box 12.2 for an illusraion). Bes pracice also involves periodically conducing and publishing sudies of long-erm rends in he revision paerns. Summaries of hese sudies may accompany he regular quarerly release of daa o remind users ha daa are subjec o revisions. Specific deails on how o produce revisions sudies are given in he nex secion. 28. I is paricularly imporan o inform users properly of he qualiy of he esimaes when releasing QNA esimaes for he firs ime. For a good indicaion of he degree of fuure revisions of he main aggregaes o expec, he complee compilaion process should be simulaed based on hisoric daa before releasing he new esimaes. Tha is, he proposed QNA compilaion sysem should be used o produce QNA esimaes for he pas years as if one were back in ime and were producing he iniial preliminary esimaes for hose years (see he discussion of he racking exercise in Chaper 2). This exercise can provide a firs assessmen of reliabiliy of he QNA sysem, which compilers may decide o communicae o users when he QNA daa are officially launched. 29. Finally, providing easy access o he revised ime series on a sufficienly deailed level should subsanially ease he inconvenience for users of frequen revisions. This involves elecronic release of he complee, deailed ime series, no only he aggregaed daa for he mos recen periods, which will make i easier for users o keep rack of he revisions and updae heir daabases. I should be emphasized ha a release of complee ime series for all revised periods is needed because users ofen use QNA daa in a ime-series forma and need o be alered o any changes in daa for pas periods. No providing hem wih revised hisoric daa will creae breaks in he ime series hey use, which will seriously hamper he

11 11 serviceabiliy of he daa. As menioned earlier, a minimum requiremen for QNA users is a leas five years of coninuous series of he key macroeconomic aggregaes. 6. REVISIONS ANALYSIS FOR ASSESSING QNA RELIABILITY 30. The analysis of revisions is an ideal framework for assessing he reliabiliy of QNA esimaes. The analysis of revision aims a measuring quaniaively he characerisics of he revisions process of QNA esimaes. In he QNA, he firs esimae of a quarer is always based on a parial se of informaion. For his reason, preliminary QNA esimaes are righly subjec o a number of revisions in subsequen publicaions. These revisions can ake differen direcions for each quarer, depending on he naure of he revision a each sage of publicaion. To undersand he quaniaive impac of revisions, he full revisions process is convenienly summarized by means of descripive saisics of he quarerly revisions hisory. Such descripive saisics called revisions indicaors form he basis o conduc a revisions analysis of QNA variables. 31. Monioring revisions indicaors is relevan for boh compilers and users of QNA daa. For users, a se of revisions indicaors offer a sandardized framework o assess he reliabiliy of preliminary esimaes wih respec o laer, revised esimaes. Evidence of bias (or lack hereof) from he revisions hisory can help users make beer decisions. For example, modelers may decide o incorporae sysemaic elemens from pas revisions in an effor o improve heir projecions. Policy makers may decide o hold off on imporan decisions unil more consolidaed esimaes are made available. A he same ime, revisions indicaors may help compilers idenify failures in he QNA esimaion process. A sysemaic bias in he early esimae of a specific QNA variable may indicae he necessiy o improve he source daa or revise he curren esimaion mehod. 32. A real-ime daabase is essenial o calculae revisions indicaors. A real-ime daabase is a abular represenaion of he QNA series as originally published by he auhoriies. When he frequency of publicaion corresponds o he frequency of measuremen (e.g., a quarerly series published once per quarer), he able assumes he characerisic form of a riangle. In such cases, a real-ime daabase is also known as revisions riangle. Example 12.1 provides an illusraion of a real-ime daabase (or revisions riangle) using arificial daa. The rows conain he quarerly raes of change as published a differen release daes; he columns refer o he sequence of esimaes relaed o he same quarer. A revisions riangle shows ha a published QNA series comprises esimaes ha have reached a differen revision sage: he laes quarer is published for he firs ime, while more disan quarers have already undergone several revisions and are likely o be more reliable han recen quarers. 33. Real-ime daabases can be prepared for any ypes of daa (e.g., seasonally or no seasonally adjused, curren prices or in volumes) and forms of presenaion (e.g., levels, changes, raios). Each form may highligh ineresing aspecs of he revisions process.

12 12 However, developing and mainaining real-ime daabase is cosly and ime-consuming. Prioriy should be given o hose daa and presenaions ha are more relevan o he users. Because QNA are mainly used for assessing shor-erm movemens in he economy, a key variable for QNA real-ime daabases is he quarerly seasonally adjused GDP growh. Realime daabases of unadjused daa may be relevan o single ou revisions due o changes in he basic sources (ne of seasonal adjusmen effecs). Revisions riangles of main expendiure and producion componens of he GDP can also be essenial o undersand he driving facors of he headline GDP revisions. Finally, revisions o he levels can be useful o assess he impac of benchmark revisions of naional accouns o he QNA levels Mahemaically, a revision R for a generic quarer is calculaed as he difference beween a laer esimae L and a preliminary esimae P, ha is 5 R=L P. (1) Suppose here are n quarerly revisions available. A revisions hisory R quarerly revisions for he n quarers, ha is R R, R,, R. 1 2 n is he sequence of 35. Revisions (1) can be calculaed beween any wo sages of esimaion. The firs esimae of a quarer always receives he greaes level of aenion from he users, and i is he one ha is mos likely assessed agains laer esimaes. In a revisions riangle, he sequence of firs esimaes is conained in he main diagonal of he riangle. A comparison beween he firs esimae and he second esimae (usually released hree monhs afer he firs esimae) provides he impac of new and updaed source daa received during his period. Longer revision horizons are useful he impac of annual source daa on he quarerly esimaes. When annual daa are received afer a considerable amoun of ime (more han one year), i is helpful o compare he esimae released afer one year wih he esimae release afer wo years. Finally, a comparison wih he laes published series (i.e., las row of he revisions riangle) may be helpful o measure how he curren quarerly informaion compares wih previously published quarerly esimaes. The laes series, however, includes esimaes of differen mauriy. 4 The OECD has developed spreadshee emplaes for creaing real-ime daabases from monhly and quarerly daa. The OECD also provides emplaes for calculaing revisions indicaors from real-ime daabases. Such emplaes are available for download on he OECD websie. For furher deails, see McKenzie and Gamba (2009). 5 Revisions are also calculaed in relaive erms. See Annex 1 for an explanaion.

13 13 Example 12.2 presens hree differen revisions hisories from he daa shown in Example 12.1: second esimaes vs. firs esimaes; esimaes released afer one year vs. firs esimaes; and laes series vs. firs esimaes. 36. Revisions indicaors are calculaed as descripive saisics of a revisions hisoryr. Each saisic aims a measuring a specific characerisic of he revision process. In general, direcion, size, and dispersion of revisions are he mos salien feaures of a QNA revisions process. The following is a lis of he mos common indicaors used for he analysis of revisions of QNA esimaes: Mean Revision (MR). I is he simple average of R. I is a measure of direcion of revisions. An MR value close o zero indicaes ha here are no direcions in he revisions. In his case, he preliminary esimae shows no bias compared wih laer esimaes. When he MR is posiive, he preliminary esimae is on average below he laer esimae (downward bias); when he MR is negaive, he preliminary esimae is on average above he laer esimae (upward bias). Mean Absolue Revision (MAR). I is he simple average of R aken in absolue values. The MAR disregards he sign of revisions. I can be zero when posiive and negaive revisions perfecly offse each oher. The MAR is a measure of he size of revisions. I is usually aken as a measure of reliabiliy of QNA esimaes. The larger he MAR value is, he less reliable he preliminary esimae is compared wih laer ones. Sandard Deviaion of Revision (STDR). I is calculaed as he sandard deviaion of R. I is a measure of dispersion of revisions around he mean value. The smaller he STDR value, he closer are he revisions o he MR value. A high STDR value indicaes large flucuaions in he reliabiliy of early esimaes beween quarers or across years. Range of Revision (RR). I is he difference beween he maximum revision and he minimum revision in R. Always non-negaive, he RR value provides an immediae quanificaion of he specrum of revisions occurred in he pas. Example 12.2 shows he calculaions of hese indicaors using he arificial daa of Example Furher, a mahemaical formulaion of he revisions indicaors is provided in Annex A glimpse a revisions indicaors can provide an immediae insigh ino he characerisics of he revisions process. In Example 12.2, for insance, he small values of he MR indicaor (less han one percenage poin) indicae ha he preliminary esimae is

14 14 relaively unbiased compared o laer esimaes. However, he MAR values show ha i is reasonable o expec revisions o he firs esimae wih a magniude of percenage poins afer hree monhs and percenage poins afer one year. The STDR indicaor also signals an increase in he dispersion of revisions from he second esimae o he esimae published afer one year (0.26 vs poins). 38. Real-ime daabases and revisions indicaors should be developed and regularly mainained as par of he QNA compilaion process. Their availabiliy makes i possible o monior consanly he reliabiliy of he published daa and promply idenify shorcomings in he QNA esimaes. When sufficien resources are available, real-ime daabases and revisions indicaors for he GDP and is main componens should also be made available o he public. 6 A summary of he GDP revisions indicaors should be given in he QNA press release, so ha he general user can immediaely appreciae he level of reliabiliy of he preliminary esimaes and he characerisics of he quarerly GDP revision process. Occasionally, more in-deph revisions sudies should be conduced o deermine he underlying causes of he GDP revisions and discuss how o reduce sysemaic characerisics of he revisions process relaed o daa sources and saisical mehods. 6 Revisions riangles are provided by a number of counries wih advanced QNA sysems. The Unied Kingdom was he firs counry ha disseminaed revisions riangles of he GDP on a srucured and regular basis.

15 15 SUMMARY OF KEY RECOMMENDATIONS Revisions are an essenial par of good QNA compilaion pracice. Revisions provide he possibiliy o incorporae new and more accurae informaion in he QNA, and hus o improve he accuracy of he esimaes, wihou inroducing breaks in he ime series. Series ha are revised regularly o reflec new and beer informaion are more accurae han hose subjec o lile or no revision. To avoid unnecessary criicism, a well-designed and carefully managed revision policy is needed. Essenial feaures of a well-designed revision policy are predicabiliy and openness, advance noice of causes and effecs, and explanaion, as well as easy access o sufficienly long ime series of revised daa. Quarerly accouns are subjec o hree waves of saisical source daa: (a) quarerly source daa; (b) annual source daa; and (c) periodic census daa. Periodic benchmark revisions are also used o inroduce revised inernaional sandards, major mehodological updaes, and changes in classificaions. A crucial par of a well-esablished and ransparen revision policy is devising an appropriae compilaion and release schedule, which should specify imeliness, frequency of updae, and revisions period of he preliminary esimaes. Revisions should be communicaed in a clear and ransparen way. Users should be noified well in advance of any major revisions wih significan impac on he curren esimaes. When a benchmark revision of naional accouns is released, a minimum of five years of coninuous series for he quarerly GDP and is main componens should be made available o he public. Revisions analysis of QNA daa is essenial o monior he reliabiliy of he esimaes and advise users on he range of uncerainy. Real-ime daabases (or revisions riangles) and revisions indicaors should be developed and regularly mainained as par of he QNA compilaion process. Bes pracices also involve periodically conducing and publishing revision sudies of QNA daa and disseminae real-ime daabases and revisions indicaors of key QNA aggregaes o he public.

16 16 Box 12.1 Compilaion and Revision Schedule: An Example Revisions o Esimaes for a Quarer of Year y Quarerly Revisions Early GDP esimae: up o 1 monh afer he end of he quarer Iniial esimae: 2 o 3 monhs afer he end of he quarer Revised esimae: 5 o 6 monhs afer he end of he quarer All esimaes may be open o revisions during he curren year Firs Annual Round of Revisions: 3-12 monhs afer he end of year y year y Annual Daa for: Preliminary annual esimaes based on a separae annual accouning sysem year y 1 Final annual esimaes based on a separae annual accouning sysem Quarerly Accouns Revised quarerly esimaes of year y + 1 +Revised quarerly esimaes for year y, and y 1 +Revised quarerly paern hrough year y 2 o y 4 o avoid seps beween year y 1 and y 2 Subsequen Annual Rounds of Revisions: monhs afer he end of year y Incorporaion of final annual esimaes for year y and preliminary esimaes for year y + 1 based on a separae annual accouning sysem monhs afer he end of year y monhs afer he end of year y Incorporaion of final annual esimaes for year y + 1, and preliminary annual esimaes for year y + 2 Incorporaion of final annual esimaes for year y + 2, and preliminary annual esimaes for year y + 3 The las wo rounds of revisions are caused by echnical properies of he recommended benchmarking mehods (more rounds wih minor revisions may in some cases be needed). The final annual esimaes may be revised laer as needed, if new daa become available or improved mehods are developed.

17 17 Box 12.2 Presenaion of Revisions: An Illusraion based on Counry Pracices Changes in This Issue Daa for he mining and manufacuring indusries have been revised as a resul of he incorporaion of new annual census resuls for he previous year. As a resul, value added for mos indusries has been revised upward in he previous and curren years. Reail oupu and household consumpion have been revised for he mos recen wo quarers following he processing of lae quesionnaires. The mos recen quarer has been revised down slighly as a resul. Changes in he Nex Issue Release dae: xxxxx. The mehodology for esimaing financial services will be revised in line wih new inernaional sandards. The concepual issues and quaniaive effecs are discussed in a research paper available on reques. Summary Tables of Revisions Table 1: Revisions o Levels of GDP in Currency Unis: Eigh Mos Recen Quarers Table 2: Revisions o Percenage Changes in Volume GDP: Eigh Mos Recen Quarers

18 18 Example 12.1 Real-Time Daabase Reference quarer 2010q1 2010q2 2010q3 2010q4 2011q1 2011q2 2011q3 2011q4 2012q1 2012q2 2012q3 2012q4 Release monh June Sepember December March June Sepember December March June Sepember December March June Example 12.1 shows a real-ime daabase (or revisions riangle). The example refers o (arificial) quarerly raes of change of a QNA aggregae expressed in seasonally adjused form. I assumes a imeliness of 2-3 monhs from he reference quarer; one esimae per quarer; and an open revisions policy for previously published esimaes. The revisions riangle can be read in differen ways. By row, he able shows he series of quarerly raes as published a each release monh. For example, he June 2010 publicaion (firs row) repors he firs esimae of he quarerly rae of change for 2010q1; he Sepember 2010 publicaion (second row) shows he esimaes of 2010q1 and 2010q2 released a ha ime; and so forh. I can be noed ha each row (e.g., each ime series) comprises esimaes wih differen mauriy. By column, he revisions riangle shows he revisions hisory of one specific quarer. For example, he fourh column shows ha he iniial esimae of 2010q4 (1.4 percen) has been revised downward several imes unil i sabilizes wo years afer he firs release a a much reduced rae (0.6 percen). Finally, he riangle can also be analyzed diagonally. The main diagonal includes all he firs esimaes of he quarers; he secondary column he second esimaes; and so forh. The diagonal view is very imporan in he calculaion of revisions indicaors beween wo specific esimaion poins in ime (e.g., firs esimae vs. second esimae), as shown in Example 12.2.

19 19 Example 12.2 Revisions Indicaors Revisions Firs Esimae Second Esimae Second esimae vs. Firs esimae Esimae afer one year Esimae afer one year vs. Firs esimae Laes Esimae Laes esimae vs Firs esimae (1) (2) (3)=(2)-(1) (4) (5)=(4)-(1) (6) (7)=(6)-(1) Reference Quarer 2010q q q q q q q q q q q q Mean Revision (MR) Mean Absolue Revision (MAR) Sandard Dev. of Revision (STDR) Range of Revision (RR) Example 12.2 arranges he quarerly esimaes from he real-ime daabase in Example 12.1 convenienly for he calculaion of revisions indicaors. Column (1) shows he sequence of firs esimaes for he period 2010q1-2012q4. Column (2) shows he second esimaes for he same quarers. The revisions are calculaed in column (3) as he differen beween he second esimae (i.e., laer esimae) and he firs esimae (i.e., preliminary esimae). Columns (5) and (7) show he revisions o he firs esimae made afer one year and in he laes release (June 2013), respecively. Revisions indicaors are repored in he boom par of he able. Taking he case of firs vs. second esimae as illusraion, he Mean Revision (MR) is obained as he simple average of revisions: MR (second esimae vs. firs): ( 0.3) ( 0.3) The Mean Absolue Revision (MAR) is calculaed as he simple average of he absolue value of revisions: 1 MAR (second esimae vs. firs): The Sandard Deviaion of Revision (STDR) is calculaed by aking he square roo of he sum of squared revisions divided by he number of revisions minus one: STDR (second esimae vs. firs): 1 (0.2) 2 (0.4) 2 (0.1) 2 ( 0.3) 2 ( 0.3) Finally, he Range of Revision (RR) is he difference beween he maximum revision and he minimum revision: RR (second esimae vs. firs): 0.4 ( 0.5) 0.90

20 20 ANNEX 1 REVISIONS INDICATORS 1. A revisions hisory can be summarized using sandard descripive saisics. Secion 6 of his chaper discussed and explained he mos common indicaors used in he revisions analysis of QNA variables (i.e., mean revision; mean absolue revision; sandard deviaion of revision; and range of revision). By offering a mahemaical presenaion of hese indicaors, his annex aims o faciliae heir implemenaion in he QNA compilaion sysem as rouine diagnosics of QNA reliabiliy Given a generic quarer, a revision R is defined as he difference beween a laer esimae L and a preliminary esimae P, namely R=L P. (1) The esimaes L and P should be expressed in he same uni of measuremen. Equaion (1) is generally used for measuring revisions on raes of change (quarer-o-quarer, or year-onyear). Alernaively, a relaive measure of revision R, R= L P L, (2) is used o calculae he revisions in erms of he laer esimae (usually in percenage form). Equaion (2) is suiable for measuring revisions in he levels of esimaes. In he QNA, however, he main ineres of compilers and users is on he measuremen of quarerly (or annual) changes. For his reason, his annex will focus on he absolue measure of revisions (1) as he basis for he calculaion of revisions indicaors. However, all he formulae presened apply equally o (2). 3. Le us assume a sample of n revisions, ha is{ R1, R2,, R n }. The Mean Revision (MR) is calculaed as he simple average of he revisions: 1 n R 1 MR= n. (3) The MR is an indicaor of direcion of revision. A posiive MR indicaes an average upward revision o he preliminary esimae (or alernaively, an average downward bias in he preliminary esimae); a negaive MR indicaes an average downward revision o he 7 For a comprehensive lis of revision indicaors, see Di Fonzo (2005) and McKenzie (2006).

21 21 preliminary esimae (i.e., average upward bias in he preliminary esimae). Ideally, he MR should be as close as possible o zero. 4. The Mean Absolue Revision (MAR) is he simple average of he absolue value of revisions: 1 n R 1 MAR = n. (4) The MAR indicaor measures he average size of revisions. I is a sricly posiive number. I is equal o zero when negaive and posiive revisions perfecly offse each oher. 5. The Relaive MAR (RMAR) is calculaed as he raio beween he MAR and he average value of he preliminary esimaes: RMAR = n 1 n 1 R P. (5) The RMAR is very useful for comparing he size of revisions beween variables wih differen magniude (e.g., counries, secors). 6. The Sandard Deviaion of Revisions (STDR) is he square roo of he sum of squared revisions divided by ( n 1): n 1 STDR = R 2 MR. (6) n 1 1 The STDR indicaor is a measure of dispersion of revisions. A small STDR indicaes ha revisions are close o he mean value. 7. Finally, he Range of Revisions (RR) is he difference beween he maximum revisions and he minimum revision: RR = max{ R} min{ R}. (7) The RR indicaor is also a measure of dispersion. I shows he wides range of revisions occurred in he sample. Because i could be affeced by he presence of exreme revisions in he sample, i should always be presened along wih he STDR measure.

22 22 BIBLIOGRAPHY Di Fonzo, T., 2005, The OECD Projec on Revisions Analysis: Firs Elemens for Discussion, OECD Shor-erm Economic Saisics Exper Group, June McKenzie, R., 2006, Underaking Revisions and Real-Time Daa Analysis using he OECD Main Economic Indicaors Original Release Daa and Revisions Daabase, OECD Saisics Working Papers, 2006/02, OECD Publishing. McKenzie, R. and, Gamba, M., 2009, Daa and Meadaa Requiremens for Building a Real-ime Daabase o Perform Revisions Analysis, conribuion o he OECD/Eurosa ask force on Performing Revisions Analysis for Sub-Annual Economic Saisics, available a hp://