Predicting early data revisions to US GDP and the effects of releases on equity markets

Transcription

1 Predicing early daa revisions o US GDP and he effecs of releases on equiy markes Aricle Acceped Version Clemens, M. P. and Galvão, A. B. (2017) Predicing early daa revisions o US GDP and he effecs of releases on equiy markes. Journal of Business & Economic Saisics, 35 (3). pp ISSN doi: hps://doi.org/ / Available a hp://cenaur.reading.ac.uk/42102/ I is advisable o refer o he publisher s version if you inend o cie from he work. To link o his aricle DOI: hp://dx.doi.org/ / Publisher: Taylor & Francis All oupus in CenAUR are proeced by Inellecual Propery Righs law, including copyrigh law. Copyrigh and IPR is reained by he creaors or oher copyrigh holders. Terms and condiions for use of his maerial are defined in he End User Agreemen.

2 CenAUR Cenral Archive a he Universiy of Reading Reading s research oupus online

3 Predicing Early Daa Revisions o US GDP and he Effecs of Releases on Equiy Markes Michael P. Clemens ICMA Cenre Henley Business School Universiy of Reading Ana Beariz Galvão Warwick Business School Universiy of Warwick Ana.Galvao@wbs.ac.uk M.P.Clemens@reading.ac.uk July 20, 2015 Absrac The effecs of daa uncerainy on real-ime decision-making can be reduced by predicing daa revisions o US GDP growh. We show ha survey forecass effi cienly predic he revision implici in he second esimae of GDP growh, bu ha forecasing models incorporaing monhly economic indicaors and daily equiy reurns provide superior forecass of he daa revision implied by he release of he hird esimae. We use forecasing models o measure he impac of surprises in GDP announcemens on equiy markes, and o analyse he effecs of anicipaed fuure revisions on announcemen-day reurns. We show ha he publicaion of beer han expeced hird-release GDP figures provides a boos o equiy markes, and if fuure upward revisions are expeced, he effecs are enhanced during recessions Key words: survey forecass, daa revisions, economic indicaors, sock reurns, macro announcemens. JEL code C53. Michael Clemens is also an Associae member of he Insiue for New Economic Thinking a he Oxford Marin School, Universiy of Oxford. Ana Galvão acknowledges suppor for his work from he Economic and Social Research Council [ES/K010611/1]. Corresponding auhor: Dr. Ana Beariz Galvao; ana.galvao@wbs.ac.uk. 1

4 Predicing Early Daa Revisions o US GDP and he Effecs of Releases on Equiy Markes The effecs of daa uncerainy on real-ime decision-making can be reduced by predicing daa revisions o US GDP growh. We show ha survey forecass effi cienly predic he revision implici in he second esimae of GDP growh, bu ha forecasing models incorporaing monhly economic indicaors and daily equiy reurns provide superior forecass of he daa revision implied by he release of he hird esimae. We use forecasing models o measure he impac of surprises in GDP announcemens on equiy markes, and o analyse he effecs of anicipaed fuure revisions on announcemen-day reurns. We show ha he publicaion of beer han expeced hird-release GDP figures provides a boos o equiy markes, and if fuure upward revisions are expeced, he effecs are enhanced during recessions. Key words: survey forecass, daa revisions, economic indicaors, sock reurns, macro announcemens. JEL code C53. 2

5 1 Inroducion Orphanides (2001) brough o he aenion of economiss he difference beween aking policy decisions in real-ime using he early esimaes of real oupu and inflaion ha are hen available compared o using he final-revised daa only available a number of years laer. Revisions o naional accouns daa are large enough o cause he policy rae implied by he real-ime Taylor rule o differ significanly from he rae compued wih revised daa. Daa uncerainy also affecs financial marke paricipans. The calendar of marke-moving indicaors published on he Econoday websie ( includes no only he advance esimae of real GDP, published up o one monh afer he end of he observaion quarer, bu also he second and he hird esimaes, released, respecively, wo and hree monhs afer he end of he observaion quarer. Indeed he resuls of Gilber, Scoi, Srasser and Vega (2015) on he impac of macroeconomic news on bond and currency markes esablish ha markes reac o surprises (differences beween he published values and he marke expecaion) in he second release of real GDP. Gilber (2011) also provides evidence ha equiy markes reac no only o surprises in he iniial release, bu also o expeced fuure revisions, indicaing ha markes care abou he revised values of economic aciviy measures. In his paper we consider he exen o which he early monhly daa revisions of GDP are predicable. Following curren usage, we refer o he Bureau of Economic Analysis (BEA) hree GDP esimaes released a monhly inervals, following he reference quarer, as he advance, second and hird esimaes (see, e.g., Fixler, Greenaway-McGrevy and Grimm (2014)). (The second and hird esimaes were formerly known as he preliminary and final esimaes). We begin wih he survey forecass of he second and hird esimaes, made subsequen o he advance and second esimaes, respecively, having been released. Of ineres is wheher he survey forecass are able o predic he daa revisions conained in he second and hird releases, and how he accuracy of hese forecass compares wih ha of forecasing models which make judicious use of monhly economic indicaors and daily financial daa available a he ime he survey forecass were made. If survey 3

6 forecasers are no able o predic early daa revisions, or if hey underperform relaive o he model, hen he usual pracice of proxying marke expecaions by survey expecaions is suspec for he early GDP releases. For example, he even sudies lieraure invesigaes he response of financial markes o new informaion provided by he release of measures of economic aciviy (see, e.g., McQueen and Roley (1993), Andersen, Bollerslev, Diebold and Vega (2003) and Faus, Rogers, Wang and Wrigh (2007)), and generally relies upon survey forecass o calculae wha consiues new informaion. Our findings sugges here are sources of informaion - no incorporaed in survey expecaions - which can be used o predic he hird release of US GDP. We provide an assessmen of he effecs of surprises in he second and hird releases of GDP on daily equiy reurns, allowing ha marke expecaions may no be accuraely measured by survey expecaions. Equiy markes are found o respond o unanicipaed news abou GDP esimaes, and he magniude of he response o he hird release is increased when model forecass are used o calculae surprises, consisen wih he model forecass beer proxying marke expecaions. We also find ha during recessions invesors respond o he informaion he GDP release carries abou fuure daa revisions, consisen wih Gilber (2011). We exend he analysis o consider he effecs of expeced and surprise revisions beween he new release and he rue value. We find ha, during recessions, upward revisions in he hird release GDP figures boos equiy markes arising boh from expeced and surprise revisions. The predicabiliy of revisions depends on he naure of he revisions o already published daa. Uncerainy abou he curren sae of he economy will decrease wih he publicaion of revised esimaes which incorporae new informaion, and which may also reduce he measuremen noise componen of he earlier esimaes. Following Mankiw and Shapiro (1986), economiss classify daa revisions as news when hey add new informaion, and noise when hey reduce measuremen error. If daa revisions are noise, hey can be prediced based on he curren esimae. Mankiw and Shapiro (1986) and Faus, Rogers and Wrigh (2005) provide empirical evidence ha daa revisions o US real GDP are largely news, while Aruoba (2008) and Corradi, Fernandez and Swanson (2009) found some limied predicabiliy of daa revisions, in paricular of iniial revisions. Clemens and 4

7 Galvão (2012) exploi muliple-vinage models o show ha real-ime esimaes of oupu and inflaion gaps can be improved by using predicions of daa revisions following he encouraging resuls of Garra, Lee, Mise and Shields (2008). Predicable daa revisions sugges ha we are able o reduce curren daa uncerainy in real ime. Much of he lieraure has used he quarerly vinages recorded in he Real Time Daa Se for Macroeconomiss (RTDSM: see Croushore and Sark (2001)), where he firs esimae is he advance esimae (as used here) bu he esimae available in he following quarer corresponds o he hird esimae. Consequenly, he predicabiliy of he monhly revisions o he advance esimae (i.e., of he second esimae relaive o he hird, and of he hird relaive o he second) has no been addressed. Noe ha given he earlier esimae is in he informaion se, hen forecasing he new esimae, or he revision beween he wo esimaes, is obviously equivalen, and we can refer o he predicabiliy of revisions or esimaes. The naure of he process by which he naional accouns daa are revised suggess ha he iniial monhly revisions may be predicable even if he revisions are news in he sense ha hey are unpredicable based on informaion a he ime he firs esimae (or an earlier esimae more generally) was made. As described by Landefeld, Seskin and Fraumeni (2008), 25% of he GDP componens a he ime of he release of he firs esimae are rend-based daa obained from exrapolaions suppored by relaed indicaor series. The proporion of rend-based daa in he second and hird esimaes is 23% and 13% respecively. As a consequence, i migh be possible o exploi he economic indicaor daa published prior o he release of he GDP figure o predic ha figure. We evaluae differen mehods of forecasing he BEA s early releases of GDP daa: survey daa, forecasing models wih economic indicaors, and models wih financial indicaors. The qualiy of survey forecass of new observaions has been exensively evaluaed (see, e.g., Ang, Bekaer and Wei (2007) for a recen appraisal), bu we are no aware of any explici assessmens of survey forecass of he revisions o iniial releases, ha is, of he second and hird esimaes. Evans (2005) compares model-based real-ime measures of oupu growh wih he MMS (Inernaional Money 5

8 Marke Services) survey median forecass of he hree iniial releases of GDP growh. However, he comparison was made o evaluae he models, wih he MMS forecass aken as he arge values. The plan of he res of he paper is as follows. In secion 2 we describe he survey daa, and he accuracy of he median forecass of he second and hird esimaes of oupu growh. Secion 3 evaluaes he forecas accuracy of forecasing models exploiing informaion ses comprising monhly economic indicaors and daily financial daa. Secion 4 analyses he impac of he mis-measuremen of marke expecaions on esimaes of he effecs of daa release announcemens on equiy reurns. I also analyses wheher announcemen-day reurns are affeced by fuure expeced revisions o he GDP figures induced by he announced value. Secion 5 offers some concluding remarks. 2 Using Surveys o Predic Revised Esimaes When predicing he second and he hird GDP releases in real ime, we are able o use an earlier release. The advance esimae (published on average 30 days afer he end of he quarer, and denoed y +1/3 ) can be used o predic he second release (published on average 60 days afer he end of he observaional quarer, denoed y +2/3 ), and he second esimae can be used o predic he hird esimaes (published on average 30 days laer han he second esimae, and denoed y +1 ). Here and hroughou we use he convenion ha he superscrips are he release daes (in monhs, as fracions of quarers), and he subscrips are he daes he observaions refer o (in quarers). A no-change forecas suggess ha he revision is no predicable. The no-change forecas of he second esimae is: ŷ +2/3 (made when y +2/3 is known) is simply ŷ +1 = y +1/3, and he shor-horizon no-change forecas of he hird esimae = y +2/3. The accuracy of no-change forecass serve as a benchmark for he forecasing models in secion 3, and also for he survey forecass in he remainder of his secion. Noe ha if we are able o predic GDP second and hird esimaes more accuraely han he benchmark, we are effecively reducing he daa uncerainy surrounding real-ime policy and economic decision-making. Before he announcemen of marke moving economic daa, business websies such as Bloomberg 6

9 ( and Econoday ( provide he consensus forecas of he pre-announcemen value of oupu growh. The consensus forecass are he medians of he forecass made on he Friday before he announcemen. These deermine he forecas horizon we consider. Aggarwal, Mohany and Song (1995) and Hess and Orbe (2011) have evaluaed survey forecass of he releases compued by he MMS (Inernaional Money Marke Services). Our preliminary resuls sugges he choice of survey provider maers lile - he accuracy of he forecass for overlapping periods is generally similar. (We compare MMS, Bloomberg, Econoday and Acion Economics). We use he survey median provided in he Econoday repor for he advance, second and hird esimaes of US GDP growh (nowihsanding he pre-eminence of Bloomberg wih praciioners). This covers 2001:M1 o 2013:M12, and so includes boh he 2001 and he recessions, and he pos Financial Crisis recovery period. Figure 1 presens he forecas errors from predicing he second and he hird release of US GDP growh. For he survey forecass, he forecas errors are he released esimaes minus he forecass of hese quaniies. The same is rue for he no-change forecass, bu in his case he forecas errors are also he revisions o he esimaes: for he second esimae, for example, he no-change forecas is equal o he advance esimae. In he figure, he daes refer o he release daes, for example, 2005M2 refers o he second esimae of GDP growh for 2004Q4. No-change forecas errors for he second release are in he range is 2.5 o 1.5%, bu are smaller for he hird release, wih a range of 0.6 and 0.6%. Hence he revisions beween he advance and second esimaes are reasonably large, given ha he average GDP growh rae is 3% (compued wih laes-available vinage daa for he period ), while he revisions beween he second and hird esimaes are markedly smaller. The improvemens offered by he professional forecasers for he second release are eviden from he figure, and reflec he accumulaion of informaion over he monh or so since he earlier esimae. Bu equally clear is ha heir forecas errors for he hird release are similar o hose of he no-change forecass. I may of course be he case ha forecasers pu less effor ino forecasing he ypically small revisions beween he second and hird esimaes, compared o he revisions beween he advance and second esimaes. 7

10 We use he roo mean squared forecas error (RMSFE) o measure forecas performance. Table 1 records he RMSFEs of no-change forecass as benchmarks agains which he survey forecass can be assessed. For compleeness, we also repor he accuracy of forecass of he firs release, using las quarer s final esimae as he benchmark: ŷ +1/3 = y 1. Table 1 includes a es of he null of equal forecas accuracy. The alernaive hypohesis is ha he no-change benchmark is less accurae han he survey median (i.e., a one-sided es). This is he -saisic of Diebold and Mariano (1995). Rejecions a 1, 5 and 10% significance levels are indicaed by, and, respecively. As expeced, he survey forecass are much more accurae for he advance esimae. Bu he resuls also indicae sizeable improvemens in accuracy for he second release of GDP growh. The RMSFE is a half of ha for he no-change benchmark. By conras, he hird esimae is no prediced any more accuraely by he survey forecasers han if we were o assume no revision o he second esimae, consisen wih he visual impression provided by Figure 1. We invesigae possible dependence of he resuls on he business cycle phase (see, e.g., Swanson and van Dijk (2006)) by evaluaing forecass separaely for observaions ha fall in expansions and conracions. The spli is based on he observaion dae as deermined by he NBER business cycle chronology. The resuls in Table 1 indicae ha he survey forecass of he hird release are equivalen o he no-change forecass independenly of he business cycle phase. In conras, he survey forecass of he second release record a larger reducion in RMSFE relaive o he no change during conracions. Second release esimaes are also more variable during conracions (compare he no-change RMSFEs for second esimaes across phases). In shor, he firs revision (i.e., he second release) is boh larger and relaively more predicable using survey forecass during conracions. 8

11 3 Using Forecasing Models o Predic Revised Esimaes The resuls in he previous secion show ha survey forecass are significanly more accurae han no-change forecass for he second release of GDP growh bu no for he hird release. In his secion we consider wheher forecasing models ha use monhly economic indicaors and daily financial indicaors are more accurae han no-change forecass for he second and hird esimaes of US GDP. All he daa we use was available o he professional forecasers a he ime hey revealed heir forecass o he survey. We use monhly vinages of US real GDP from 1966:M2 up o 2014:M1 from he Real-Time Daase for Macroeconomiss (RTDSM) of he Philadelphia Fed (see Croushore and Sark (2001)) o esimae he forecasing models. The RTDSM conains he daa available a he middle of each monh. Because early releases are normally published a he end of he monh, we reschedule he real-ime daa se such ha he firs, second and hird monhly vinages wihin a quarer conain he releases of, respecively, he advance, he second and he hird esimaes. We assess which informaion is useful o predic early revisions (namely, he second and hird esimaes) in a real-ime ou-of-sample forecasing exercise. In-sample evaluaions (such as Aggarwal e al. (1995), amongs ohers) may be misleading, especially if here are parameer insabiliies. Agains his, ou-of-sample evaluaions require longer spans of hisorical daa because separae in-sample esimaion and ou-of-sample forecas periods need o be defined, bu neverheless we choose o conduc an ou-of-sample evaluaion. We evaluae forecass from auoregressive models in secion 3.1, from models wih monhly economic indicaors in secion 3.2, and from models wih daily financial daa in secion 3.3. Table 2 summarizes he forecasing models used in his paper for ease of reference, wih deailed explanaions in wha follows. We aim o forecas he second and he hird esimaes, ha is, y +v for v = 2/3, 1. Noe ha = 1, 2,... in quarers, varying for boh vinages (superscrips) and observaions (subscrips). All forecasing models use he revision y +v variable for reasons explained in secion 3.1. y +v 1/3 as he dependen 9

12 We presen a real-ime analysis of forecasing early GDP monhly releases. The ou-of-sample periods mach he release daes covered by he survey (2001:M2-2013:M12). A each new forecas origin we re-esimae each model wih an expanding number of observaions obained from he real-ime daase available a he ime he forecas was made. As in secion 2, we provide resuls for he whole period and also he spli by business cycle phase. We assess wheher model forecass are more accurae han he random walk using he -es of Diebold and Mariano (1995) (DM), assuming quadraic loss. An alernaive es for nesed models is he encompassing saisic of Clark and Wes (2007) (CW), which makes an allowance for he effec of parameer esimaion uncerainy (in esimaing he nesing model). Effecively he CW es assumes ha we have an infiniely large sample, ha is, ha we are able o use he populaion values of he model s parameers o generae forecass. The DM approach ess wheher he model is more accurae han he random walk allowing ha he model needs o be esimaed, and unlike CW, will only rejec he null when he mean squared forecas error of he model s forecass is smaller han ha of he random walk. Thus he DM approach seems preferable for our purposes. 3.1 Informaion from Pas Vinages We sar by considering forecasing models wih an informaion se resriced o pas daa in he US real GDP real-ime daase. If pas daa vinages of oupu growh help predic he second and hird esimaes in comparison wih he no-change forecas benchmark, hen revisions a leas in par embody a reducion in noise or measuremen error. The firs panel of Table 2 summarizes he five forecasing models. The firs model assumes ha daa revisions are serially uncorrelaed, possibly wih a non-zero mean. The second model adds an auoregressive erm relaed o he spillover effec : see e.g., Jacobs and van Norden (2011). The hird model allows revisions o depend on he value of he earlier release. Similar regressions are commonly used o es wheher revisions are news. If β 0 = β 1 = 0 hen daa revisions are unpredicable (news) as defined by Mankiw and Shapiro (1986). Noe ha by comparing he ouof-sample forecasing performance of hese models wih he no-change forecas, we are assessing 10

13 he ou-of-sample predicabiliy of daa revisions. If he DM es rejecs he null, we conclude ha revisions are no pure news. Clemens and Galvão (2013) have shown ha models of muliple daa vinages are able o predic quarerly daa revisions o oupu growh and inflaion by exploiing informaion on pas revisions, and in paricular, he annual revisions which ake place in he hird quarer of each year. The fourh model in he firs panel of Table 2 is a vinage-based model: a simplified singleequaion version of heir model (see also Koenig, Dolmas and Piger (2003), and Croushore (2011a) for a recen survey of forecasing wih daa vinages). We experimen wih q = 5, 14, where q is he number of lags. Swanson and van Dijk (2006) repor ha he biases of he revisions o indusrial producion depend on he sae of he business cycle. To capure possible business cycle asymmeric effecs, we consider a hreshold specificaion ha allows he response of he revision o he earlier release o depend on he sign and size of he earlier release: his is he fifh model in Table 2, a hreshold model. Noe ha in he specificaion of his model I() is an indicaor funcion (so I (x) = 1 when x is rue, and I (x) = 0 oherwise) and c is he value of he hreshold. The hreshold is joinly esimaed wih he slope parameers by condiional leas squares. The esimaion employs a grid search for he hreshold value c based on he resricion ha each regime mus have a leas 15% of he observaions (see, e.g., Hansen (2000)). Table 3 presens he raios of he RMSFEs of each of he 5 models o ha of he no-change benchmark. By and large, here is lile indicaion ha any of hese own-informaion pas vinage models improves on he no-change benchmark, and in paricular, here is no evidence o suppor he use of a hreshold specificaion. Broadly, hese findings are in agreemen wih he lieraure suggesing here is limied predicabiliy of earlier revisions o US GDP growh (as, for example, Mankiw and Shapiro (1986) and Faus e al. (2005)). 11

14 3.2 Informaion from Monhly Economic Indicaors As discussed, early monhly esimaes of real GDP are based on exrapolaions, and subsequen releases incorporae new informaion as i becomes available (see e.g., Landefeld e al. (2008)). Forecasers migh be able o predic upcoming daa releases by using he new informaion published since he previous release (eiher he advance esimae, or he second) bu before he arge release (respecively, he second, or he hird) is announced. We consider monhly economic indicaors which are someimes caegorized as marke moving (see, e.g., Econoday, reflecing heir perceived imporance as indicaors of he sae of he economy. The variables included in his sudy are lised in Table 4, along wih he daa sources, and heir imeliness (or delay), which we discuss below. Their imporance derives from heir correlaions wih GDP and is componens, and heir early availabiliy. Our choice of variable is also deermined by he availabiliy of a real-ime se wih monhly vinages over a long period. Our monhly indicaors include indusrial producion and employmen. Marke paricipans in general perceive he announcemens of hese variables as carrying informaion on subsequen GDP growh announcemens. The nonfarm payroll announcemen in paricular receives much media aenion. Reail sales is jusified as an indicaor of curren consumpion, and he producion manufacuring index (NAPM) and durable good orders measure curren aggregae producion. An alernaive measure of consumer spending is provided by he Universiy of Michigan Consumer Senimen Index, which is generally regarded as a leading indicaor, as opposed o a coinciden indicaor. We also include wo housing aciviy measures, housing sars and new home sales, as well as he CPI as a measure of inflaion (he GDP deflaor is released a he same ime as GDP, and so canno be used as a predicor). Finally, we consider he monhly rade balance compued from he Balance of Paymen accouns. An addiional moivaion o include his variable is ha expors and impors are GDP componens subjec o mean absolue revisions (beween he iniial monhly esimaes and he laes-available esimaes) which are hree or four imes larger han for personal consumpion expendiures, even hough heir proporion of real GDP hey accoun for is 12

15 small (see, e.g., Fixler e al. (2014, Table 1, p.5)). The majoriy of he variables in Table 4 are subjec o revision. This means ha for such variables we ypically have (i) daa published afer he announcemen of he curren GDP esimae, including new observaions and revisions o he pas daa; and (ii) values and observaions already available before he announcemen of he curren GDP esimae ( pas informaion). We can organize he new informaion ino: new revision, new observaion and updaed observaion. By comparing he relaive forecasing accuracy of models which exploi new, updaed and pas informaion, we can assess he effi ciency of early GDP releases for laer releases, and discover which informaion helps predic subsequen GDP releases. To illusrae he use of he differen ypes of informaion, consider he five indicaors (which include indusrial producion and employmen) which are published wih a delay shorer han 21 days from he end of he observaional monh. Assume ha X refers o he observaion in he las monh of quarer, while X 2/3 is he observaion in he firs monh of quarer. We employ quarerly differences of he monhly variables, ha is, x +v revision is given by x +v = (X +v X 1 +v ). The monhly x +v 1/3. Suppose we wish o predic he second esimae of GDP in quarer, namely y +2/3. The informaion se consiss of he advance GDP esimae, y +1/3, as well as he second-monh vinage for x, comprising he firs esimae of x for he monh following he reference quarer (x +2/3 +1/3 ), he second esimae of x (x +2/3 ) and revised values for earlier monhs, as well as daa in he firs-monh vinage for x and earlier periods, ec. Tha is, { y +1/3 ; x +2/3 +1/3, x+2/3,... ; x +1/3, x +1/3 1/3 }.,... The New Revision regression model uses he published revision of he indicaor x +2/3 x +1/3 o predic y +2/3 y +1/3. (This model and he ohers discussed in his secion are given in he second panel of Table 2). The New Observaion model uses x +2/3 +1/3 Updaed Observaion model uses x +2/3 o predic y+2/3 (as opposed o he revision, x +2/3 consider he use of Previous Release daa, i.e., he use of x +1/3 x +1/3 o predic y +2/3 y +1/3. The ). We also y +1/3. Tha is wheher daa a he ime of he publicaion of he firs release (x +1/3 ) helps predic he revision o GDP. 13

16 As shown in Table 2, his illusraion of predicing he second esimae generalizes o predicing he hird esimae. For example, he New Revision model uses x +1 x +2/3 esimae, y +1 y 2/3 ; he Previous Release model uses x +2/3 o predic y +1 For he hree indicaors published wih a longer delay, x +2/3 o predic he hird y +2/3, and so on. indicaes a firs esimae (raher han a second esimae), and he models have o be adaped accordingly. For example, he New Revision model is only applicable for forecasing he hird release of GDP (v = 1). The Updaed Observaion is he iniial release of he indicaor for predicing he second release of oupu growh, bu is he revised value oherwise. The New Observaion regression model (employing x +v +v 1/3 as a predicor) is only feasible for he five indicaors published wih a shor delay. Noe ha he resuls of he Previous Release model have a bearing on wheher daa esimaes are effi cien (or wheher subsequen revisions are predicable). If Previous Release model forecass are significanly more accurae han no-change forecass, early GDP esimaes are no effi cien since hey do no use all available informaion. If daa revisions are no predicable from pas informaion, hen revisions are ypically classified as news (see, e.g., Croushore (2011b)). Noe ha in conras wih much of he lieraure, he use of a shor forecas horizon of up o one-week-ahead allows for he possibiliy ha a given release may be news, in he sense of being unpredicable based on daa a he ime of he earlier release, bu may sill be predicable from more recen informaion: he new revisions, observaions, or updaed informaion. By exploiing economic releases beween he curren and arge GDP release, we consider wheher he arge release is predicable up o one-week in advance. For he wo survey-based variables in Table 4 ha are no subjec o revisions and are published wih shor delays (NAPM and consumer confidence), we apply he New Observaion model wih x +v 1/3 o exploi new informaion, and he Previous Release model wih x +v 2/3 o consider pas informaion. The resuls for using monhly economic indicaors as predicor variables are given in Table 5. We consider all he variables aken ogeher (firs wo panels of Table 5), as well as he predicive power of he indicaors one-by-one (hird and fourh panels). Given he poor performance of models 14

17 wih auoregressive componens (see Table 3) we omi such erms, while non-zero mean revisions are accommodaed by he inclusion of inerceps in he regressions. We record RMSFE raios of he models using each ype of informaion agains he benchmark. (The second panel of Table 2 summarizes he models we esimae.) The firs panel of Table 5 shows he resuls for he second GDP esimae, and he second panel he resuls for he hird esimae (using all he indicaors). For he second esimae, boh he Updaed Observaion and Previous Release model resul in saisically significan improvemens in accuracy (a he 10% level). The usefulness of boh hese ypes of informaion is greaer in conracionary quarers. For he hird esimae, he only gains are from he use of Previous Release survey daa, and hen only in conracionary periods. Resuls for each individual predicor in he hird and fourh panels of Table 5 sugges lile is los by considering all he variables ogeher. In summary, here is some evidence of predicabiliy during recessions for boh releases, using economic indicaors, and he second esimae is predicable overall from boh new informaion (he Updaed Observaion model) and pas informaion (he Previous Release model). However, while survey forecass of second esimae GDP growh improve on he benchmark by 50% on RMSFE, for he hird esimae he model-based gains (over he no-change forecas) are markedly less (and end o be realized in recessions). 3.3 Informaion from Daily Financial Variables Our hird informaion se consiss of daily financial variables. Tha financial variables may have predicive conen for growh daa revisions is suggesed by Andreou, Ghysels and Kourellos (2013), who show daily financial indicaors help o nowcas revised values of GDP growh. They find shor-erm ineres raes, bond spreads and sock reurns are among he indicaors wih he bes forecasing accuracy for oupu growh one-quarer-ahead. Secondly, Gilber (2011) argues ha on days ha advance esimae announcemens are made, equiy reurns respond o incorporae informaion on expeced fuure daa revisions o measures of economic aciviy such as nonfarm payroll employmen and oupu growh. This implies ha equiy reurns (observed during he firs 15

18 monh of he curren quarer, + 1/3) migh help predic he second and hird esimaes released in + 2/3 and + 1. We use Mixed Daa Sampling (MIDAS) regressions o exploi he informaion in daily financial variables for predicing he quarerly daa releases (see for example he review aricle by Andreou, Ghysels and Kourellos (2011) on MIDAS). The MIDAS regression is described in he hird panel of Table 2. The lag operaor is applied o daily daa, and we assume ha here are m = 60 daily observaions per quarer. The number of daily lags is se o K. The weighing funcion w j (θ, K) is a bea funcion wih wo parameers in he vecor θ. The aggregaion weighs w j (θ, m) sum up o 1 o guaranee he idenificaion of he slope parameer α 1. Galvão (2013) shows ha bea funcions work beer han an exponenial funcion when m is large. The parameers of he weighing funcion are joinly esimaed wih he slope and inercep parameers by nonlinear leas squares. When using informaion up o, he lead parameer l is se o v, and K = 60, ha is, we use all he daily daa from he observaion quarer. When using informaion up o + 1/3, l = 1/3 for v = 2/3, and l = 2/3 for v = 1, while in boh cases K = 20, so only daa from he monh of he firs GDP announcemen is considered. Insead of esimaing he funcion o aggregae high frequency daa, we can also assume fla aggregaion (equal weighing) and se w j (θ, K) = 1/K for all he daily lags, giving he Linear model of Table 2. Galvão (2013) suggess ha regime changes in he slope parameers may also affec he accuracy of oupu growh forecass. The slope coeffi ciens in models which use financial variables o predic oupu growh may shif because of marke regimes (bull/bear) and moneary policy regimes (loose/igh). Therefore, we also employ he Smooh Transiion MIDAS (STMIDAS) regression as a forecasing model o exrac informaion from daily financial variables. The MIDAS model is modified such ha he slope parameers are weighed by a logisic funcion. The values of he logisic funcion (beween 0 and 1) a each poin in ime depend on he difference beween he aggregaed high frequency daa and a hreshold c. The smoohness of he funcion depends on he parameer γ. The STMIDAS regression is described in he las row of Table 2. Noe ha 16

19 he parameers of he aggregaion funcion w j (λ, m) of he ransiion funcion may differ from he parameers of he aggregaion funcion of he indicaor as a predicor (w j (θ, m)). We need a long hisorical sample on each financial variable o esimae hese models for ou-ofsample forecasing. This resrics us o he 5 financial variables described in Table 6, wih daa from he early 60 s for all he variables oher han he Baa spread. The empirical resuls of Andreou e al. (2013) sugges he use of sock reurns (boh SP500 and DJIA) and he shor rae as predicors of economic aciviy variables. Gilber (2011) uses he SP500 o capure he marke reacion o he release of he advance esimae of GDP growh. As well as hese variables, we include a measure of he ineres rae spread (compued as he difference beween he 10-year Treasury bond and a 3-monh Treasury bill), as suggesed by Galvão (2013). We also include he shor rae as well, allowing he model o capure he level and slope of he yield curve (as suggesed by he findings of Wrigh (2006)). Finally, we include he Baa spread, defined as Moody s BAA yield minus he 10-year Treasury Rae. This is moivaed by he evidence in Gilchris and Zakrajsek (2012) on he predicive power of corporae bond spreads for economic aciviy. We experimened wih muliple variable models, as indicaed by he noaion in Table 2. We included all financial variables ogeher, as well as a varian ha included jus he level and slope of he yield curve, and finally a model including jus he wo equiy variables: see he firs wo panels of Table 7. Resuls are given only for he MIDAS and Linear models, as STMIDAS models did no fare well given he proliferaion of parameers when here is more han one variable. Resuls for STMIDAS for single financial predicors (and for he MIDAS and Linear models) are given in he hird and fourh panels of Table 7. Table 7 presens he RMSFE raios wih respec o he benchmark. I also compares he accuracy of models using daily daa hrough quarer (K = 60), wih models wih daily daa for he firs monh of he quarer (i.e., he monh of he firs announcemen, + 1/3, K = 20), and wih models using daily daa up o he day before he second or he hird release announcemens. This las comparison is resriced o observaions from 1975 onwards, for which we have he precise daes of he GDP (or GNP) releases. I is run for K = 60 and he resuls are presened in he las 17

20 columns of Table 7 (indicaed by x +db ). The measures of relaive accuracy show ha equiy reurns from he monh of he firs announcemen have saisically significan predicive power for boh releases. (This is confirmed by he resuls for he individual indicaors). The equiy reurns over his period capure any effec ha he announcemens of he advance GDP esimae and marke-moving economic variables may have had on he sock marke. The able also shows ha daily daa for he period beyond he firs monh (included in he informaion se available wih x +db for y +1 ) is of no value for predicing he daa revision revealed by he announcemen of he hird esimae. Daily sock reurns end o have a significan effec during expansions bu no during conracions. The MIDAS model is generally as good or beer han he Linear model. I is possible ha he predicive conen of equiy reurns sems solely from heir embodying news on he economic indicaors released during he monh. To see wheher his is he case, in Table 8 we consider he incremenal effec in erms of forecas accuracy of adding he equiy variables o he bes forecasing models using monhly economic indicaors. From Table 5, we found he Updaed Observaions model provided he bes forecass of he second esimae, and he Previous Release model he bes forecass of he hird GDP esimae. We repor he RMSFEs for hese models in Table 8, and hen he effec of including daily reurns in hese models via a MIDAS regression wih a bea weighing funcion, and wih all he parameers being esimaed joinly by nonlinear leas squares. Saisically significan reducions of RMSFE from he inclusion of daily financial variables are deeced for predicing he hird release of GDP growh during boh business cycle phases, and for predicing he second release during recessions. The resuls in Table 8 indicae ha equiy reurns conain addiional informaion o ha in he economic variables. 4 Daa Revisions and Equiy Markes The lieraure has idenified wo main problems wih esimaing he effecs of macroeconomic surprises on equiy reurns. Firs, Rigobon and Sack (2008) argue ha if survey forecass are a 18

21 noisy proxy for marke expecaions, he esimaes of he impac of macroeconomic surprises on asse reurns will be aenuaed. Second, good news may be bad news for sock reurns. An unexpeced increase in growh may presage a ighening of moneary policy o allay fears of a build up of inflaionary pressure. In general, his is solved by considering he impac of surprises during expansions and conracions separaely (see, e.g., McQueen and Roley (1993)). Good news may have negaive effecs on sock markes during expansions and posiive effecs during conracions. In his secion we address boh of hese issues. We look a he effec of surprises emanaing from he second and hird GDP releases on daily equiy reurns, allowing differenial impacs across business cycle phases by running separae regressions for recessions and expansions. For marke expecaions we use boh he Econoday survey median (following Andersen e al. (2003)) and a model-based measure moivaed by our earlier resuls. We also assess he evidence for he finding of Gilber (2011) ha invesors respond o he informaion he GDP release carries abou he rue value of GDP. As well as he replicaion of he Gilber (2011) even sudy on our daase using boh SP500 and DJIA reurns, we also explicily decompose fuure revisions ino an expeced and a surprise componen o furher invesigae he response of he sock marke o informaion abou he final value conveyed by he announcemen-day release. This expeced/surprise decomposiion, ogeher wih measures of marke expecaions of forhcoming GDP releases ha draw on daily sock reurns, help shed addiional ligh on how GDP revision announcemens affec he equiy marke. 4.1 The Impac of Announcemen Surprises Our empirical resuls sugges ha he model forecass are superior o he survey forecass for he hird release, alhough he survey forecass perform beer for he second release. However, a combinaion of model and survey forecass migh perform even beer, and poenially provide a superior measure of marke expecaions. We calculae regression-based forecas encompassing ess (see e.g., Clemens and Harvey (2009) for a recen review) o invesigae he poenial for combinaion, bu find ha survey forecass encompass model forecass for he second release, in 19

22 boh business cycle phases, and ha he model forecass encompass he survey forecass for he hird release. This suppors he use of he model forecass as a proxy for marke expecaions for he hird release o lessen he impac of error in he expecaions measure. We repor resuls using boh survey and model forecass for he wo releases. The model forecass are generaed from a MIDAS regression which use he bes model wih monhly economic variables for he release in quesion, combined wih daily sock reurns (SP500) for he monh of he iniial release of GDP (see secion 3.3 and Table 8). We esimae he effecs of surprises in GDP release announcemens on daily sock reurns (measured by he SP500 and he DJIA) on he day of he announcemen. Preliminary resuls sugges ha he size of he effec is similar for boh measures. Given he relaively small sample for his even sudy (52 quarerly observaions), we esimae he wo equaions by pooled ordinary leas squares o obain a more accurae esimae of he effecs of surprises. Announcemen surprises are sandardized and measured as: S +v,k = y+v sd(y +v ŷ +v,k ŷ +v,k ), where ŷ +v,k is he forecas using mehod k (eiher a model or professional forecasers consensus) of he second release (if v = 2/3) or of he hird release (if v = 1). Le re +v,i denoe he reurn o sock index i on he day of he announcemen of a revised GDP figure. Then we evaluae he impac of daa revision surprises by esimaing: re +v,i = β 0 + β 1 S +v,k + ε,i, (1) where i {SP500, DJIA} and runs over he 52 evens. Table 9 repors he esimaes of he slope coeffi ciens (β 1 ) in equaion (1), and he R 2 saisics. The resuls employing he survey forecass as marke expecaions (in he op panel) confirm previous resuls in he lieraure (e.g., Gilber (2011)) ha hird-release surprises have no impac 20

23 on sock reurns. Second-release surprises are shown o have an impac on sock reurns, bu only during recessions. The posiive sign of he coeffi cien implies ha good news has a posiive impac on he sock marke during recessions. These resuls could be inerpreed as suggesing ha equiy markes pay more aenion o daa revision releases during recessions, when heir relaive size and predicabiliy is larger (see secion 2). The use of he model-based measure of marke expecaions increases he magniude of he esimaed response of reurns o second-release surprises (see he second panel of Table 9). The size of he response o hird-release surprises riples, for all observaions (i.e., when we do no differeniae by business cycle phase). Alhough he coeffi cien is no significan a convenional levels, his may simply reflec he small sample size. The increased esimaed response is consisen wih he greaer accuracy of he model forecass for he hird release, and wih hese forecass providing beer proxies of marke expecaions. 4.2 The Impac of Fuure Revisions Gilber (2011) argues ha invesors respond o he informaion conveyed by he iniial release abou he correc value and no only is preliminary esimae. Gilber (2011) defines he oal surprise as he difference beween he final value (y + ) and he forecas of he announcemen (ŷ +v,k ), which can be wrien (in our noaion) as: T S +,k = y + ŷ +v,k = ( y + y +v ) }{{} R + ) y +v ŷ +v,k }{{} S +v,k ( + (2) ha is, as he (non-sandardized) revision R + plus he (non-sandardized) announcemen surprise S +v,k (which we will coninue o refer o as he surprise). By including a sandardized version of R + in he regressions of secion 4.1 (e.g., R+ = R + /sd ( R + ) ) we are able o gauge he response of announcemen-day reurns o fuure revisions, as well as o announcemen surprises (S +v,k ). The firs of hese erms allows reurns o respond o he rue value. 21

24 In he firs panel of Table 10, we replicae Gilber s regression using our panel daase, and repor resuls for announcemen surprises calculaed using survey-consensus forecass, and including in he regression he acual revisions, R +. We approximae y + using daa from he 2013M12 vinage, and consequenly shoren our sample of daa releases. We remove he las wo years so ha he 2013M12 vinage can be used o provide a reasonable measure of fuure revisions, R +. Our resuls for second-release announcemens mach Gilber (2011, Table 8 and 9, p.128). Fuure revisions have a significan negaive effec on sock reurns during recessions, bu no (significan) effec during expansions. However, our resuls for hird-release announcemens differ. Gilber finds a significan negaive effec of hird-release revisions on reurns during recessions, and a posiive effec in expansions, whereas we only find a significan effec in conracions, and he effec is posiive. If we insead measure announcemen surprises S +v,k using model-based forecass (see he second panel of Table 10), we now find evidence of a significan response of equiy markes o surprises in hird-release announcemens. This is consisen wih he superior accuracy of he model forecass (relaive o he survey forecass), as discussed in secion 4.1, which provide more accurae esimaes of he surprises experienced by he marke. Fuure revisions o hird releases coninue o have a significan effec, as when survey forecass are used o define surprises, bu fuure revisions o second-releases no longer have an impac. These resuls imply ha, conrolling for announcemen day surprises, upward revisions in hirdrelease GDP figures boos equiy markes during recessions. To furher invesigae his issue, we decompose he revisions erm R + in (2) ino he expeced revision, ER, and he surprise revision, SR. Tha is, ( y + y +v ) = ( E +v y + y +v ) + (y + E +v y + ). (3) }{{}}{{}}{{} R + E +v R + SR + An issue wih he use of R + o measure fuure revisions is ha he rue value y + will no be realized unil many years laer, and will include benchmark revisions and changes in he mehodology of daa collecion and compilaion, which will be unforeseen a period. One migh suppose 22

25 ha he announcemen-day reurn would only respond o he predicable revision, E +v R +, i.e., how far he curren release is from he prediced rue GDP value. We consider regressions which include he announcemen day surprises S +v,k (as in secion 4), as well as boh E +v R + and SR +, as a way of deermining wheher he expeced revision or he acual fuure revision drives announcemen-day reurns. In populaion, a leas, if he coeffi cien on he surprise revision is no significanly from zero, he resuls would favour he expeced revision. In pracice of course we have a relaively small sample of daa for easing ou he imporance of hese differen facors. We require ha our esimae of E +v y + - he forecas of he rue value - accuraely reflecs he unknown marke expecaions of he rue values. These expecaions are generaed by vinage-based vecor auoregressive models of real GDP growh (as in Clemens and Galvão (2013)) assuming ha he rue value y + is well approximaed by he value in he quarerly vinage released 14 quarers afer he observaional quarer. The model is esimaed on quarerly vinages of daa up o an including he + v vinage, and explois he predicive conen of pas vinages for fuure vinages. The resuls in Table 3 sugges ha a simplified version of his approach was he only auoregressive specificaion able o improve upon he no-change forecas, a leas during recessions. The hird panel of Table 10 records he resuls of regressing reurns on (sandardized versions of) S +v,k, E +vr + and SR +. We find ha neiher expeced or surprise fuure revisions have significan effecs on sock reurns for second releases. The evidence ha fuure revisions affec equiy markes on he day of he hird GDP release is confirmed. The finding ha he expeced and surprise fuure revisions are of he same sign and a similar magniude indicaes ha hird-release announcemen-day reurns respond o he acual fuure revision (as opposed o he expeced fuure revision). This could be explained by markes knowing more abou upcoming hird releases han is indicaed by our model forecass. Tha is, par of he surprise fuure revision, given our forecass, may in fac be included in marke paricipans forecass of fuure revisions. In general, he publicaion of beer han expeced early GDP figures provides a boos o equiy markes during conracionary periods. And if he marke expecs fuure upward revisions 23

26 (especially o he hird release figures), he effecs are enhanced. By and large, he use of modelbased expecaions provides more evidence ha equiy markes reac o daa revisions han when survey forecass are used o measure marke expecaion, principally for hird releases. In our analysis, he models are used o measure marke expecaions of he upcoming announcemen and of fuure revisions o he announced value. 5 Conclusions Daa revisions clearly conribue o he uncerainy abou he curren sae of he economy, and abou he curren condiions of macroeconomic fundamenals, which in urn may affec economic aciviy. An early conribuion was Oh and Waldman (1990), who considered he macroeconomic effecs of false announcemens (see also Oh and Waldman (2005)), and argued ha an upbea esimae of he curren sae of he economy which was subsequenly revised down would lead o sronger oupu growh han would oherwise have ranspired (wih he reverse being rue of an ex pos pessimisic assessmen). Rodriguez-Mora and Schulsad (2007) find ha firs announcemens of GDP growh are a more imporan deerminan of subsequen acual GDP growh han he rue value of GDP growh in he earlier period (see also Clemens and Galvão (2010)). The imporance of expecaional errors for business cycle flucuaions has a long hisory, as indicaed in he cied papers. A recen srand of he lieraure has considered he role of noise shocks in generaing aggregae flucuaions (Lorenzoni (2009), Blanchard, L Huillier and Lorenzoni (2013)). Blanchard e al. (2013) esimae ha noise shocks accoun for more han half of he forecas-error variance of oupu growh a shor horizons: changes in he fundamenals explain a smaller proporion of his variance. Measuremen errors in iniial esimaes of GDP and relaed macro variables (such as produciviy growh) may consiue one source of noise shocks, and as such he exen o which subsequen revisions are predicable may have imporan implicaions for business cycle analysis. Our empirical invesigaion focuses on deermining he predicabiliy of early daa revisions o US oupu growh a shor horizons, namely, he predicabiliy of he revisions revealed by he 24