ACCOUNTING FOR CRISES #1. Venky Nagar University of Michigan Gwen Yu University of Michigan

Transcription

1 ACCOUNTING FOR CRISES #1 Venky Nagar Unversy of Mchgan Gwen Yu Unversy of Mchgan November 2009 # We are graeful o George-Maros Angeleos, uz Hal, Sephen Morrs, Hélène Rey, Joel Slemrod, and semnar parcpans a he Unversy of Mchgan, Carnege-Mellon Unversy, 2009 AAA FARS Conference, 2009 Uah Wner Accounng Conference, he 4h Tnbergen Conference on Crashes and Sysemc Crses, and 2009 UNC Global Issues n Accounng Conference for her commens.

2 ACCOUNTING FOR CRISES Absrac We provde one of he frs ess of recen macro global-game crss models ha show ha he precson of publc sgnals can coordnae crses (e.g. Angeleos and Wernng 2006; Morrs and Shn 2002, 2003). In hese models, self-fulfllng crses (ndependen of fundamenals) can occur only when publcly dsclosed fundamenal sgnals have hgh precson; fundamenals are hus he sole drver of crses n low-precson sengs. We affrm hs proposon on 39 currency crses by explong a key publcly dsclosed fundamenal drvng fnancal markes, namely accounng daa. We fnd ha fundamenal accounng sgnals are sronger n-sample predcors of crses n low-precson counres. 1

3 ACCOUNTING FOR CRISES I. Inroducon A key queson n nernaonal economcs, explored n a seres of recen crss models such as Angeleos and Wernng (2006), Angeleos e al. (2006), Angeleos e al. (2007), Rey (2000), Akeson (2000), Morrs and Shn (2002), and Morrs and Shn (2003), s wheher economc fundamenals or speculaors self-fulfllng belefs drve crses. Whle dverse n her sengs and modelng approaches, he models all pon o he precson of publc nformaon as he key drver. If publc nformaon s precse enough relave o speculaors prvae belefs, can coordnae mulple self-fulfllng speculaor belefs largely ndependen of economc fundamenals. The belef coordnaon role of publc sgnals n crss models s a sark deparure from neoclasscal asse prcng models, where hgh precson sgnals can only ghen he lnk beween fundamenals and prces, no weaken hem. Ye, hs cenral and robus role of publc sgnals n he analycal crss models has receved scan emprcal aenon, perhaps due o he dffculy n fndng a publc sgnal ha nsuonally compors wh he crss models and whose precson s measurable. Ths paper argues ha accounng daa provde such a sgnal and uses accounng daa o es he cenral resul of recen crss models. All he crses models above share hree basc seps (see Fgure 1): frs, he nal fundamenals are realzed. Nex, he speculaors coordnae (he marke s oo large for any ndvdual speculaor) over an acon such as nerm fnancng or whdrawng capal. Ths coordnaed acon n conjuncon wh he nally realzed fundamenal has a real effec on he underlyng asse s value. In he fnal sep, hs new value ranslaes no asse prce. 1 1 As any observer of he recen Wall Sree urmol wll mmedaely recognze, boh speculaor belefs abou fundamenals and speculaor belefs abou oher speculaors wllngness o supply nerm fnancng play a key role n asse valuaons (Mollenkamp e al. 2008). 2

4 The man pon of hese crss models s ha speculaors belefs abou each oher n he second sep can be self-fulfllng, especally when publc sgnals have hgh precson, hus dvorcng he fnal value of he asse from s nal fundamenal realzaon. For example, suppose each speculaor s unsure of he measure of speculaors akes o unravel a currency. An accurae publc sgnal released n he fnancal markes says ha hs measure s 30%. Ths 30% s hus he nal fundamenal srengh of he currency. If every speculaor beleves every oher speculaor wll aack (.e., whdraw capal), everyone wll aack. The currency wll hen fall, for he measure of aackers s 100%, whch exceeds 30%. On he oher hand, f every speculaor beleves ha no one wll aack, a small ndvdual aack only wases ha speculaor s money. No one wll aack and he currency wll sand. The accurae publc sgnal hus suppors wo self-fulfllng equlbra over s realzaon range (0%, 100%),.e., any realzaon of he nal fundamenals n hs range can lead o eher of he aack or no-aack oucome. The above example assumes ha he publc sgnal s precse. If hs s no he case, he crss models we es show ha speculaors wll wegh her prvae nformaon more heavly. Because hs prvae nformaon vares across speculaors, each speculaor s unsure wha he oher speculaors are hnkng. Each ndvdual speculaor s hen more fearful of losng money n an unsuccessful uncoordnaed aack. In such sengs, poor nal fundamenals become he man pon of coordnaon and hus he man drver of crses. In hs paper, we es hs key comparave sac of he crss models descrbed above usng accounng daa. Specfcally, we argue ha realzed accounng daa should more srongly predc currency crses n counres where he accounng daa are less precse. We use accounng daa for several reasons. Frs, publc corporaons accounng daa are a key source of publc nformaon on asses raded n a counry s fnancal markes. Second, he accounng leraure has exensvely examned he noon of accounng qualy, or he precson wh whch accounng measures reflec 3

5 economc realy. In parcular, hs leraure argues ha here s consderable varaon n he precson of accounng nformaon across counres, due o varaon n legal, enforcemen, and rule-makng nsuons (a Pora e al. 1998; Ball e al. 2000). Such nsuonal varaon s essenal o our emprcs. More mporan, hese precson measures capure he nose n he accounng sgnals. By conras, he varance of alernave publc sgnals such as sock prces s he sum of he fundamenal varance and he nose varance. Addonally, prce sgnals ha are he oupus of fnancal marke radng are hemselves subjec o mulplcy and excess volaly (Angeleos and Wernng 2006), confoundng emprcal esmaon. Accounng daa, by conras, are npus no he fnancal markes and bypass hs problem. Fnally, our proxy for accounng daa qualy measures publc sgnal precson relave o prvae sgnal precson, as requred by he crses models. Our seng comprses 39 currency crses n 21 counres from 1981 o Usng pror leraure, we consruc a compose score of accounng precson for each counry (based on all s frms). We use hs compose score o spl he counres no wo groups: hgh-precson and low-precson. Our brfurcaon of he sample follows Angeleos and Wernng (2006, Fgure 1) who show ha he swch from unqueness o self-fulfllng equlbra s no gradual as he publc sgnal precson ncreases, bu occurs suddenly a a precson hreshold. We hen aggregae all he frms n each counry every year o yeld hree annual counry-based measures of realzed performance: earnngs, accounng accruals (accounng adjusmens o cash flows o yeld earnngs), and volaly of earnngs. We es he n-sample power of hese measures o predc crses. Graphcal analyses (Fgure 2) are conssen wh our predcons. Fgure 2 ndcaes ha he realzed performance measures are relevan o crses: relave o ranqul years, hey are que choppy around he crses onse years, especally for low precson counres. Furher, he confdence nervals of realzed 4

6 performance measures around crses are much larger n low precson counres. However, despe her nose, we fnd ha publc sgnals have he power o predc crses n low-precson counres. Specfcally, we es he n-sample power of realzed accounng performance measures o predc crses. We conrol for a rch ls of oher leadng ndcaors as well as counry fxed effecs and conemporaneous cross-seconal shocks (conagon). In he overall sample, he ncluson of accounng measures sgnfcanly mproves our n-sample ably o predc crss one year n advance: he explanaory power ncreases from o 0.283, a 15 percen ncrease. Much of he predcve power arses from accruals (.e. accounng adjusmens o cash flows). Ths s a parcularly neresng resul, for shows ha he applcaon of accounng rules o measure frm operaons s wha generaes crcal asse-prcng nformaon. More mporan, we show ha mos of he predcve mprovemen comes from he sub-sample of counres wh low accounng precson. The F-sasc for he realzed accounng performance measures one year pror o crses s n low precson counres, bu 1.31 for he hgh precson counres. By conras, when we examne accounng daa one year afer he crss onse year, we fnd ha boh low precson and hgh precson accounng counres show sgnfcance: he F-sas are now and respecvely. So accounng daa n hgh precson counres do appear o reflec he consequences of a crss; hey jus don have ex ane predcve power. We hen conduc a seres of comparave sacs nvolvng nsuons, he naure of he crses, he specfc precson hresholds, and specfc ndusry secors. Our man resuls connue o hold n all hese ess. Our fndngs make hree conrbuons o he leraure. Frs, we emprcally valdae a key predcon of recen crss models, namely ha fundamenals are more mporan han self-fulfllng belefs n precpang crses when fundamenals have low precson. Second, and perhaps more mporan, s dffcul o es self-fulfllng belefs drecly: one can only show ha fundamenals don 5

7 maer. Bu hen, s no clear f he resul s due o self-fulfllng belefs or lack of sascal power. By showng sengs where fundamenals maer as well as sengs where hey don, our sudy overcomes hs objecon. Fnally, our very use of accounng daa s an nnovaon n emprcal crses sudes, whch have ypcally used macro, nsuonal, and rade facors o predc crses (Marn and Rey 2006; Rancere e al. 2008). Our resuls ndcae ha accounng daa offer sgnfcan ncremenal power beyond hese facors o predc crses. Secon II places our research queson n he conex of pror leraure. Secon III descrbes our daa and our emprcal consrucs. Secon IV presens he man resuls. Secon V ess he robusness of our resuls. Secon VI concludes. II. Background Perhaps he smples way o frame he crses leraure s o use equaons from undergraduae macro (e.g., Mankw 2003, Ch. 13), ( ) ( ) ( ) Y C( Y ) I( r ) G NX ( ) (1) M P ( ) ( ) (, Y ) where r (2) e ( ) ( ) NX ( ) CF( r * r ). (3) The hree endogenous varables are he domesc GDP Y, he domesc real neres rae r (he nomnal neres rae s smply he nflaon adjused verson), and he exchange rae ε (he prce of foregn currency). Equaon (1) saes ha he GDP s smply consumpon plus nvesmen plus governmen expendure plus ne expors. Invesmen declnes when he cos of borrowng s hgh. Ne expors follow he Marshall-erner condons and ncrease when he domesc currency s cheaper. Equaon (2) says ha he supply of real money equals money demand. People demand more money 6

8 when he GDP s hgh and less money when he opporuny coss or he neres raes are hgh. The supply of money M s se by moneary polcy. Equaon (3) s smply an accounng deny: any mbalance n he rade of goods has o be balanced wh an IOU or capal flows. CF s capal ouflows from he domesc counry, whch s more lkely f he foregn neres rae r* s hgh. Ths smple model llusraes many well-known feaures of nernaonal economcs. The governmen canno use M o conrol wo endogenous varables, and ε smulaneously, unless s wllng o resrc he CF funcon. Ths s he famous nernaonal polcy rlemma. Purchasng power pary s smply a specfc srucure of he NX funcon (hgh elascy around ε = 1), whle he open neres pary requremen mposes smlar resrcons on he shape of he CF funcon. More broadly, he key mplcaon of hs model for our paper s ha any explanaon of exchange rae (ncludng s sudden drop) has o be grounded n ssues such as moneary polcy, rade, and capal flows. Ths s precsely he roue ha he pror leraure has aken. Krugman shows how raonal speculaors n fxed exchange economes foresee he drop n foregn exchange componen of he moneary reserves M, and drve he currency down va capal flows CF. Emprcal ess of earler crses (e.g., Blanco and Garber [1986]) suppored hs heory, bu laer crses appeared o less nfluenced by facors such as reserves (equaon (2)), and drven more by acves n he fnancal markes supporng producve acves n he economy (equaons (1) and (3)). The search for oher facors led Obsfeld (see hs 1996 summary) o model crses as arsng from speculaors self-fulfllng belefs. He modeled a fnancal marke where each speculaor s oo small o affec he currency. Bu f speculaors collecvely coordnae and whdraw suffcen capal from a counry, s currency wll collapse. Consequenly, f a counry s fundamenals are moderaely srong, bu a large measure of speculaors s pessms hese speculaors belefs by hemselves can precpae 7

9 a crss. Obsfeld s sudy generaed a large spae of models (see Fourcans and Franck s [2003] excellen survey). Idenfyng speculaors belefs n he daa, however, proved o be hard. Jeanne (1997) and Jeanne and Masson (2000) used non-lnear emprcal ess wh Markov swchng o denfy hese belefs n he devaluaon of he French franc n Markov swchng s a maxmum lkelhood esmaor ha spos large shfs (he swch) n he me-seres of he franc exchange rae. Because hese swches are unrelaed o he already conrolled-for fundamenals, hey can poenally represen self-fulfllng belefs. Bu he concern wh such ess s ha he nably of he fundamenals o predc crses could arse from low sascal power and no from self-fulfllng belefs. The ease of achevng mulple-equlbra analycally and he dffculy n spong hem emprcally led Angeleos and Wernng (2006) o label economss relaon o mulple equlbra as love hae. Summarzng he sae of affars, Obsfeld (1996) called for more explc modelng of he neracon beween fundamenals and speculaor belefs. Ths nex sep was underaken by Morrs and Shn (1998). They showed ha even f each speculaor has an epslon amoun of prvae nformaon avalable, each speculaor becomes unsure of oher speculaors prvae nformaon and wheher hey wll parcpae n an aack. Ths uncerany s key because speculaors lose money f hey are no n suffcen numbers. As a resul, speculaor coordnaon necessary for self-fulfllng belefs does no oban: poor fundamenals hen reman he key drver of crses. Akeson (2000) and Rey (2000) (as well as Morrs and Shn [2002, 2003]) poned ou ha he key drver of Morrs and Shn s resul was no he precson of he speculaors prvae nformaon per se, bu s srengh relave o publc nformaon. If he publc nformaon were precse enough, would provde a coordnang pon for speculaors o coordnae her aack even f he currency were moderaely srong (see Secon I of hs paper for a numercal example). In parcular, Akeson (2000) 8

10 poned ou ha such publc nformaon could arse from rades among prvaely nformed speculaors. Angeleos and Wernng (2006), Angeleos e al. (2006), and Angeleos e al. (2007) flesh ou hs nuon by endogenzng varous aspecs of he radng model, namely dvdends, prce, payoffs, ec. More mporan, hey also endogenze he polcy response o he crss. Ther man concluson s ha, despe such endogeney, here s a range of realzed fundamenals where endogenous publc nformaon s suffcenly precse relave o speculaors dverse prvae nformaon o rgger speculaor coordnaon and self-fulfllng belefs. Consequenly, fundamenals canno predc crses n hese ranges because self-fulfllng belef rggers can occur over he enre range (noe ha hese belef rggers are no guaraneed a any specfc pon n hs range, only her possbly s). Ths s a sark deparure from radonal asse prcng models where hgh precson sgnals ghen he lnk beween realzed fundamenals and prces. Why he deparure hen? The key assumpon n he crss models s ha he evenual prce depends on he nally realzed fundamenal and a coordnaed acvy (such as nerm fnancng) by all speculaors. Ths coordnaed acvy --- and coordnaon s necessary because he marke s oo large for any ndvdual agen --- can affec he asse s ably o generae cash flows and hus s evenual prce. Agens herefore use publc nformaon no jus as a sgnal o uncover he underlyng fundamenal bu also as an mporan sraegc ool o form hgher order belefs of ohers acons. I s hs sraegc role of publc nformaon ha generaes nformaon exernaly ha leads o self-fulfllng belefs. On he oher hand, f publc nformaon s relavely mprecse,.e., below a ceran precson hreshold, speculaors canno coordnae on her belefs (hey pay oo much aenon o her own dverse prvae nformaon and gnore he common publc nformaon). 2 As a resul, crses happen only 2 Of course, f speculaors prvae nformaon s compleely precse, everyone knows ha everyone else receves he same sgnal realzaon. Ths suaon s equvalen o a precse publc sgnal, and we are back o mulplcy (see Angeleos and Wernng 2006). 9

11 when fundamenals are weak enough for a suffcenly large mass of speculaors o feel confden ha hey wll prof n aackng he currency: poor fundamenals are hus he sole deermnan of crses. These feaures are all especally salen n nernaonal fnancal markes where raders are small relave o he marke and her collecve supply of ongong nerm capal fnancng s key o survval. When nernaonal raders whdraw capal from a counry s fnancal marke, a currency crss ensues. Consequenly, f we assume a) ha varaon of he absolue precson of he publc nformaon across counres reflecs varaon of precson of publc nformaon relave o prvae nformaon, b) ha we can use he daa o nomnae he precson hreshold necessary for self-fulfllng belefs o be feasble, and c) ha counres where self-fulfllng belefs can occur do experence such evens (.e., mulplces are emprcally realzed), we have: Hypohess: Realzed publc dsclosures of fundamenals should predc currency crses more srongly n he subsample of counres where hese dsclosures have low precson. Accounng daa form a naural seng for esng hs predcon, for several reasons. Accounng daa are a key publcly dsclosed fundamenal no jus n deb and equy markes, bu also for bank loans (Dchev and Sknner 2002). Accounng nformaon s relevance o bank loans s especally mporan because banks are an mporan fnancng vehcle n many counres. Second, he use of accounng daa n he prcng of secures and loans has promped exensve accounng research on he noon of accounng precson. Ths research also explores he causes and he consequences of varaon n accounng precson across counres (Secon III has he deals). We explo hs nsuonally drven varaon for our purposes. 10

12 Thrd, we can measure he varance of he nose n he accounng sgnal. By conras, he varance n oher publc fundamenals such as prces ha arse from radng n asse markes ncorporaes boh fundamenal varance and nose varance. Furher, prces are subjec o mulplcy and excess volaly (Angeleos and Wernng 2006), makng dffcul for he emprcal researcher o uncover he underlyng precson of he publc nformaon from realzed values. In many nsances, he asses may no be raded suffcenly ofen o yeld a lqud prce (Greenlaw e al. 2008). Accounng daa crcumven hese problems because hey are npus no he fnancal markes. Fourh, he noon of precson of publc nformaon n he crses models we sudy s defned relave o he speculaor s prvae nformaon. Our measure of accounng precson capures hs concep. Numerous emprcal sudes have shown ha hgher qualy accounng nformaon s assocaed wh reduced bd ask spreads and less prvaely nformed radng (euz and Verreccha 2000; Daske e al. 2008). 3 These fndngs sugges ha hgher precson n he publc accounng nformaon also reduces he precson of he prvae nformaon among agens. Thus, emprcal proxes of accounng precson used n our sudy represen he relave precson of publc nformaon over prvae nformaon as n Angeleos and Wernng (2006). Despe hese advanages, here has been no aemp n he early warnngs leraure (a leas o our knowledge) o use accounng nformaon o predc currency crses. 4 Insead, pror leraure has prmarly focused on macro measures o predc crses. For example, Rancere e al. (2008) and Marn and Rey (2006) show ha counres wh hgh growh skewness and hgh radng coss are more lkely o suffer crses. We nclude counry ndcaors o accoun for such across-counry varaons. Tha s, our ess are mng ess ha operae on a whn-counry model. In addon, we also conduc our analyses 3 For example, Daske e al. (2008) show ha when counres shf o a hgher qualy fnancal reporng regme, boh bd-ask spreads and radng coss of he frms n he counry declne by 12 bass pons. 4 Swanson e al. (2003) sudy he nformaon conen of accounng fgures followng he 1994 Mexcan currency devaluaon. 11

13 usng only he radable secors. Fnally, Yuan (2005), among ohers, shows ha correlaon across equy markes can propel crses (conagon). We use cross-seconally correlaed errors o accoun for conagon. III. Daa and Varable Defnons III.A Currency Crses and Fnancal Daa Snce our goal s o predc n-sample crses, we lm ourselves o counres ha have had crses. Gven he fnancal marke seng of our analycal models, wha currency crses are emprcally bes sued o analyze our hypoheses? Whle crses are heerogeneous, he fnancal markes are ypcally an mporan conrbung facor (Kamnsky and Renhar 1999, Renhar and Rogoff 2008). We herefore use all he crses as our sample, and hen conduc sensvy analyses over specfc crses subses. Kamnsky and Renhar (1999) defne currency crss as an even of a seep decrease n exchange raes and/or reserves. They provde an exensve ls of crss evens (Kamnsky and Renhar 1999, Table II), whch Capro e al. (2005) and Kamnsky (2003) subsequenly updae. We defne he crss onse year as he year a crss sared n he Kamnsky (2003) and Capro e al. (2005) daases. Ths procedure yelds 68 crses epsodes from 21 dfferen counres as shown n Table I. 5 Table I classfes he dfferen ypes of crses based on Kamnsky (2003, Table IV). Table I shows ha 78 percen of he crses evens can be classfed as eher fnancal excess or soveregn deb. These ypes of crses ypcally arse from fnancal llqudy problems followng a perod of hgh expansonary cred growh (Tornell and Weserman 2005). Fnancal markes hus are mporan drvers of hese crses, makng hem an approprae seng for our sudy. 5 Some counres experence mulple ypes of crses n he same year. Our analyses coun hese evens as one even. 12

14 We collec frm level fnancal daa from Thompson Daasream, whch conans accounng nformaon from annual repors of publcly raded companes around he world. To be ncluded n he sample, a counry mus have more han fve frm-year observaons wh non-mssng values for a number of accounng varables such as oal asses, curren asses, curren lably and ne operang ncome. We acknowledge ha our analyss s lmed o he publcly raded secor; acves of prvae companes ha do no rade publcly are no capured drecly (excep hrough her effec on publc frms). However, noe ha publc frms ypcally end o be he larger frms, and can hus have a large mpac on he counry s economy. Thompson Daasream defnes each frm observaon by he un of equy ssues. Thus, f a frm ssues eques on wo dfferen exchanges, wll coun as wo frm observaons. Snce secures lsed on a foregn exchange are also subjec o accounng rules of he foregn counry, we delee all secures cross-lsed on foregn sock exchanges. Ths deleon ensures ha he accounng sgnals of each counry are an oucome of he local accounng sandards. 6 Our procedure yelds 101,492 frm-year observaons from 21 dfferen counres n our fnal sample. The lmed avalably of frm year observaons n earler years resrcs our analyss o crses epsodes afer Ths runcaon removes some early reserves based crses and makes he sample more relevan o our fnancal marke based hypohess. We hen aggregae he frm-years no counry-years (we do no over-wegh counry-years wh more frm-year observaons). Ths aggregaon yelds 331 o 371 observaons dependng on he regresson. These counry-years nclude 39 crses. Table II repors he onse year of each crss n dfferen counres, as well as he number of publc frms n our sample for each of he counry-years. There s consderable varaon n he number of frm- 6 The resuls are robus o ncludng he cross-lsed frms. 13

15 year observaons across counres, reflecng dfferences n he level of ndusralzaon, fnancal marke developmen, and also daa avalably. The shaded areas n Table II show consderable varaon n he spread of crses across counres and me. Crses have a slgh endency o be clusered n he early 1990s and lae 1990s, reflecng he exsence of he well known conagon effec (Allen and Gale 2000; Kamnsky e al. 2003; Yuan 2005). III.B Precson of Accounng Sgnals B.1 Measures Our man predcon s ha accounng fundamenals are a sronger predcor of crses n counres wh low accounng precson. We now descrbe our compose measure of accounng precson for each counry. The accounng leraure --- see summares n Dechow and Sknner (2000) and Healy and Wahlen (1999) --- has exensvely researched he precson or he ably of accounng measures o capure economc fundamenals. The source of accounng (m)precson arses from he followng problem: perod cashflows are no perod economc earnngs. For example, he manager may have spen cash on nvesmens ha wll pay back n he fuure, so he cash ouflow s no a pure economc loss. Alernavely, asses may have declned n value leadng o an economc loss, bu here s no cashflow mpac because he asses are no sold. Accounng herefore adjuss he cashflows o consruc a measure of earnngs or profs. Ths adjusmen s called accruals. The nose n hese adjusmens s hen our proxy of he precson of he publc sgnals. Noe agan ha we are no measurng he varance of he overall performance sgnal, we are measurng he nose n he accounng adjusmens. Ths s precsely he measure ha he crss models requre. 14

16 To users of fnancal saemens, hese accrual adjusmens are relevan, bu her relably can be mperfec. Specfcally, he relably or he precson can be mpared because managemen can make esmaon msakes, or can msuse her dscreon over accruals o conceal economc realy (boh hese facors are evden n he curren U.S. morgage crss, for example). Bu wha facors resran managemen accounng choces? Recen accounng research ndcaes ha he deeper nsuonal facors ha deermne frms accounng choces are accounng rules, legal enforcemen, and he legal regme (e.g., Ball e al. 2000). These facors vary across counres, yeldng us an nsuonally drven cross-counry varaon n accounng precson n our daa. Whle recognzng accounng precson s concepual and nsuonal mporance, he accounng leraure has no converged o a unversally acceped measure of accounng precson. Dfferen accounng sudes pck a dfferen propery of accruals o deduce he precson of accounng measures. We employ sx commonly used measures ha capure varous dmensons along whch accounng nformaon relably reflecs relevan frm fundamenals. Table III defnes n full deal he sx measures we use, as well as her sources n he leraure. We aggregae each measure o he counry level by usng he medan of he frm year observaons. We sgn he measures so ha lower values reflec hgher precson. Our frs measure of accounng precson, accruals qualy (=AQ 1 ), capures he esmaon errors n he accounng process by measurng how well accrual esmaes map no cash flow realzaons. Followng Dechow and Dchev (2002), we operaonalze hs measure as he sandard devaon of he resdual from a counry-level regresson of curren accruals on mul-perod operang cash flow. ow sandard devaon mples hgher accounng precson. Our second measure AQ 2 proxes for he level of managemen dscreon, ofen known as he smoohng behavor (Fudenberg and Trole 1995; Trueman and Tman 1988). Smoohng refers o 15

17 managers msusng her reporng dscreon o conceal economc shocks by over-reporng poor performance and under-reporng srong performance. The accounng leraure has radonally used a srong negave correlaon beween changes n accruals and operang cash flows o proxy for managemen nervenon over and beyond he naural level of accruals accounng (e.g., Francs e al. 2005). The negave of hs correlaon s hen our AQ 2 measure. The remanng four measures of accounng precson (=AQ 3, AQ 4, AQ 5 and AQ 6 ) are varous measures of he magnude of accruals. Sloan (1996) suggess ha large accruals nvolve hgher degree of subjecvy ha can ofen resul n boh nenonal and unnenonal reporng errors. euz e al. (2003), on he oher hand argue ha he larger he absolue magnude of accruals, he more room he manager has o exercse dscreon n reporng earnngs. We measure boh hese conceps boh wh curren accruals (=AQ 3, AQ,4 ) ha arse from operang acves, and oal accruals (=AQ 5, AQ 6 ) ha nclude accruals from boh operang and fnancng acves. We scale he accruals as per he orgnal papers. Then, as defned n Table IV, we consruc a compose measure of accounng precson from he sx AQ measures o elmnae poenal measuremen error. We rank each measure across all counres and ake he mean of he sx ranks as a compose counry ndex of accounng precson. Ths s our counry-based measure of he precson of he publc sgnal. 7 Table IV sors he counres n ascendng order based on he compose ndex wh lower scores reflecng hgher accounng precson. All sx ndvdual measures exhb large varaon across counres bu smlar rankngs n erms of relave magnudes. The magnudes of he measures conform by and large wh pror leraure (Bhaacharya e al. 2003, Table I and III; euz e al. 2003, Table II) wh small dfference due o dfferen sample perods. 7 In Secon V, we repor he robusness of our man resuls o alernave measures of counry-level accounng precson. 16

18 The crss models n Secon II have a specfc precson hreshold a whch self-fulfllng belefs are feasble. We herefore need o paron he sample based on hs precson hreshold. Bu s no clear how o ranslae he analycal precson hreshold o he daa. Table IV herefore dchoomzes he sample a he medan no counres wh hgh and low accounng nformaon precson. In-sample rankng can dsor he relave rankng of he 21 crss counres f accounng precson makes counres more (or less) prone o crses. Hence, we re-examne he accounng precson rankng of he 21 sample counres afer ncludng 27 addonal counres ha dd no suffer any crss. Unabulaed resuls show ha ncludng he 27 counres has mnor effec on he relave rankng of he 21 sample counres. In parcular, he expanded sample caegorzes 10 counres (.e., 47% of he 21 sample counres) as counres wh hgh accounng precson when he sample medan s used as he precson hreshold. 8 In addon, wh some excepons such as Ausrala, he counry classfcaon of hgh and low accounng precson groups confrms pror sudes ha sugges ha nsuonal characerscs (a Pora e al. 1997) and enforcemen of conracs (Ball e al. 2000) are relaed o he accounng nformaon envronmen. For example, Table IV shows hgh ranks for European counres such as Denmark, Fnland, Span and Sweden, whle developng counres lke Argenna, Brazl and Greece rank among he counres wh low accounng precson. The fac ha some counres from common law orgn are classfed n he low precson group (.e., Malaysa and Thaland) s conssen wh Ball e al. (2000), who argue ha common law nfluence does no guaranee accounng nformaon qualy when he enforcemen of legal conrac s weak. In he followng secon, we drecly examne he relaonshp beween our measure of accounng precson and varous nsuonal characerscs. 8 We es he robusness of he resuls o alernave dchoomes n Secon 5. 17

19 B.2 Sources of varaon n accounng precson across counres and over me Accounng pracce emerges n response o sewardshp and valuaon demands for accounng nformaon from nsuons and capal markes. In Table V, we drecly examne he relaonshp beween our accounng precson measures and varous proxes of legal and nsuonal envronmen culled from pror leraure. Table V, Panel A shows he counry ranks of each nsuonal varables sored by he level of accounng precson. We use he well-known an-drecor ndex (a Pora e al. 1998; Djankov e al. 2008) and he credor rghs aggregae score (a Pora e al. 1997; Djankov e al. 2007) o proxy he level of nvesor proecon (EGA c ). To address he common crcsm ha s he law enforcemen raher han he rules self ha defne he legal envronmen, we also examne varous measures of law enforcemen from he pror leraure. The enforcemen varable (ENFORCE c ) s a combnaon of he rule of law ndex from he Inernaonal Counry Rsk Gude and he deb conrac enforcemen measure from Djankov e al. (2006). Fnally, we collec he measure of dsclosure qualy (DISCOSE c ) from a Pora e al. (2006). The correlaons n Table V, Panel B show ha accounng precson s ndeed posvely correlaed wh he qualy of legal nsuons and he levels of law enforcemen. Specfcally, he accounng precson measures (AQ c ) show a srong posve assocaon wh level of enforcemen (ENFORCE c, = 0.505). However, he legal rule self (EGA c, = 0.067) s weakly correlaed. One possble explanaon for hs weak correlaon s he addonal varaon n accounng precson due o frm-level ncenves such as nvesmen opporunes, exernal fnancng and ownershp srucure (Durnev and Km, 2005). Fnally, he DISCOSE c measure s posvely assocaed ( = 0.226) wh our accounng precson measure, provdng addonal consruc valdaon. We aggregae each of our AQ c measures across frms and across me o creae counry specfc measures. However, accounng polces hemselves can evolve n response o crses (Angeleos e al. 18

20 2006; Angeleos e al. 2007). As counres mplemen such rule changes, emporal shfs n he crosssecon of accounng precson can occur. We drecly examne our precson measures me seres sably wh AR(1) correlaons. Table V, Panel C shows he AR(1) mes seres correlaon of each AQ measure of 21 counres spannng he years 1981 o Across all AQ measures, he assocaon beween each AQ measure n successve non-overlappng sub-perods of hree or fve years s sgnfcanly posve, suggesng ha he accounng precson s perssen ---- s nsuonally dffcul o change accounng rules and enforcemen quckly (unlke say neres raes). Indvdual accounng rules may change, bu overall accounng precson s unlkely o change rapdly n a counry. Tha sad, he hree year AR(1) correlaons are much sronger han he fve-year AR(1) correlaons suggesng ha he fve year shfs n he daa are more subsanve. Obvously, we canno ell wheher hese shfs are nose or rue varaon, so we repea our man analyses wh he fve-year aggregaon perod. III.C Macroeconomc eadng Indcaors n Pror eraure The general concluson n he crss predcon leraure s ha an effecve warnngs sysem should consder a large varey of ndcaors (Kamnsky and Renhar 1998). We adop he leadng ndcaors proposed n Kamnsky and Renhar (1998, Table IVA) and Edson (2003, Table V). Followng Edson (2003), we group he ls no fve major caegores: curren accoun ndcaors, capal accoun ndcaors, real secor ndcaors, domesc fnancal ndcaors, and global ndcaors. Table VI, Panel A provdes defnons of all he seveneen leadng ndcaors, her daa sources (prmarly he Inernaonal Fnancal Sascs), and he predced drecon of changes pror o a 19

21 currency crss. All ndcaors are defned as a percenage change from he prevous year, excep for he ndcaors already measurng devaon from a rend. 9 The descrpve sascs of all he leadng ndcaors are n Table VI, Panel B. Some leadng ndcaors have exreme values. The exreme values for he currency overvaluaon varable are from Indonesa and Mexco durng perods of hgh nflaon. The exreme values for excess real M1 balances are due o EU counres ha experenced a dsconnuy n M2 measures n To ensure ha hese exreme observaons do no domnae our emprcal ess, we repea all our emprcal ess afer excludng hese wo varables and fnd smlar resuls. Table VII provdes descrpve sascs for each leadng ndcaor varable across dfferen counres, along wh addonal counry nformaon, sored by he counres accounng precson. A comparson of he cos of crses, measured by foregone oupus as well as he acual loss of reserves ncurred n defendng he speculave aack (Bordo e al. 2001), ndcaes ha counres wh low accounng precson appear o have suffered more severe crses. Counres wh low accounng precson also end o have hgher nflaon and GDP growh over he sample perod. Gven ha volale and unsable counres are more lkely o have nsuonally weaker accounng regmes (e.g., Ball e al. 2000), hs able provdes addonal suppor for our accounng precson paron mehod. III.D Realzed Accounng Fundamenals Table VIII provdes he defnon of he hree accounng sgnals we use o operaonalze he realzaon of fundamenals. These measures are a) accruals, b) earnngs or profs, and c) volaly of earnngs. We do no nclude cashflows because hey are smply earnngs less accruals (cashflows and 9 The wo ndcaors are excess real exchange rae and excess real M1 balances. 20

22 accruals are correlaed a 0.9 n our sample). We oban he medan of each measure for each counry year and nomnae as he counrywde measure for ha year. Whle he ls of poenal accounng measures and accounng raos useful for evaluang frm performance s very large (Ou and Penman 1989), he measures we choose are wdely recognzed as key accounng measures of frm performance (Dechow 1994; Dechow and Schrand 2004). More dealed accounng measures and raos may no be equally vald across a dverse se of frms and counres, and also have lmed daa avalably. There s some overlap n he daa underlyng our accounng precson mercs and he daa underlyng realzed accounng fundamenals, nowhsandng dfferen aggregaon procedures. Wh some excepons, he realzed fundamenals are ypcally frs-momen effecs, where as he accounng precson mercs are varance-covarance second-momen effecs. More mporan, our analyses nclude counry fxed effecs, so any across-counry varaon n he measures n Table IV wll have no mpac on he resuls; we wll assess only he whn-counry effec n our regressons. D.1 Realzed Accounng Sgnal: Accruals The frs accounng sgnal we employ, Accruals, represens he adjusmen o cash flows o yeld accounng earnngs. These adjusmens play a key role n reporng frm performance, especally n mes of rapd downurns and upurns, for cashflows are no ye mpaced. For example, banks ypcally do no wa for loans o defaul before wrng hem off. Such advance wreoffs generae large negave accruals. 10 kewse, n upurns, frms may recognze revenue before he cashflows from cusomers have maeralzed. Of course, he exen o whch accruals sysemacally predc fuure frm 10 In hs Congressonal esmony on Feb. 28 h, 2008, Fed charman Ben Bernanke parly mplcaed he wreoffs resulng from he mark-o-marke accounng rule as a drver of he curren U.S. morgage crss (hp:// Also see Greenlaw e al. (2008). 21

23 performance s hghly conroversal n he accounng leraure. Alhough has been well shown ha he accrual componen s less perssen han he cash flow componen of earnngs (Sloan 1996), recen sudes such as Hrshlefer e al. (2009) fnd ha a he aggregae level, accruals are posvely assocaed wh fuure performance. We herefore expec accruals o be large and negave pror o a downurn or a crss. We follow pror leraure (Jones 1991; Dechow e al. 1995; Sloan 1996) and focus on curren accruals ncludng he reversal of ceran non-curren operang asse accruals by subracng deprecaon and amorzaon. We compue accruals from balance shee and ncome saemen nformaon, and hen compue cash flows as operang ncome mnus accruals. We do no use he cash flow saemen o compue accruals because of lmed daa avalably of cash flow nformaon across counres and me. Table VIII, Panel B ndcaes ha he mean of accruals s -0.01, smlar o Sloan (1996, Table I) who repors accruals of Noe ha accruals, hough aggregaed n an enrely dfferen manner, also form he bass of our measure of he precson of accounng nformaon (Secon III.B). Alhough he emprcal measure s dencal, s mporan o noe ha we mplemen he wo n very dfferen ways. The level of accrual as a proxy for accounng precson capures he varaon across counres. On he oher hand, he accruals level as a sgnal for fundamenal s employed o capure whn-counry varaon over me. Therefore, our defnon of accruals as an accounng fundamenal apples only whn each counry. In any even, we revs hs ssue n Secon V where we downplay he mporance of accrual levels as measures of accounng precson. 22

24 D.2 Realzed Accounng Sgnal: Operang Profably Operang profably or operang earnngs requre lle movaon. We defne operang profably as he counry medan of frm-level ne operang ncome scaled by begnnng oal asses. Table VIII, Panel B ndcaes ha operang profs average o a reasonable 8.5 percen of asses. D.3 Realzed Accounng Sgnal: Earnngs Volaly Followng sudes such as Rancere e al. (2008), whch mplcae hgher momens as he predcors of crses, we nclude volaly of he repored earnngs as our las accounng sgnal. Volaly s he sandard devaon of operang ncome (scaled by begnnng oal asses) over a hree-year backward rollng wndow. Crses are roubled perods wh hgh uncerany; we herefore predc a posve assocaon beween crss onse and earnngs volaly. D.4 Correlaons Table IX, Panel A presens he correlaon marx of all crses predcors, ncludng accounng sgnals and leadng ndcaors from pror leraure. Smple examnaon of he correlaon ncreases our confdence n he valdy of our measures. For example, ndusry oupu s posvely correlaed wh equy prces (= 0.34, Spearman) and commercal bank depos s posvely assocaed wh domesc real neres raes (= 0.34, Spearman). Also, he assocaons of our accounng sgnals are plausbly sgned: accruals and profably show a posve relaon (0.37, Spearman). More mporan, here also appears o be lle evdence of mulcollneary; our hree accounng sgnals hus capure dfferen dmensons of realzed fundamenals. Table IX, Panel B presens he me seres correlaon of all he hree accounng sgnals. The assocaon of he conemporaneous and lagged accounng measure s sable. For example, he 23

25 correlaon beween Profably -2 and Profably -1 (= 0.549, Spearman) s close o he correlaon beween Profably -1 and Profably (= 0.559, Spearman). More mporan, he AR(1) effec n he realzed accounng fundamenal measures s sgnfcan. Our emprcal ess herefore ncorporae varous lead-lags of he realzed accounng fundamenals o ge a beer undersandng of he predcve mng effecs. We urn o hese ess nex. IV. Resuls IV.A The Sory n Pcures We frs presen a graphcal represenaon of he movemens n accounng sgnals for he perods leadng up o and mmedaely followng he currency crss. Followng Echengreen e al. (1995), we compare he behavor of each accounng sgnal durng ranqul perods as well. Fgure 2 repors he movemen n accounng sgnals hree years before and afer he 39 currency crses. The horzonal axes represen he number of years before and afer he crss (or ranqul) year. The bands represen he upper and lower 25% quarles of he realzaon of each accounng sgnal. Accounng sgnals show much movemen durng crses, especally for low accounng precson counres. Profs declne for hese counres. Accruals do so as well and ener negave errory, suggesng consderable wre-offs. Volaly of prof ncreases as predced. By conras, n he ranqul years, he daa are ndeed ranqul across boh ses of counres, suggesng ha he movemen durng crses years s no enrely spurous. The unvarae naure of he fgures necessaes cauon n any nference. For example, he movemen of accounng sgnals n low precson counres also comes wh larger confdence nervals. We herefore urn o a more formal analyss of he daa. 24

26 IV.B Mulvarae Analyss of Crss Predcon We examne he relaon beween accounng varables and he occurrence of a currency crss n a regresson framework. Our un of observaon s a counry-year. The majory of he early warnngs leraure akes he sgnals approach (Kamnsky and Renhar 1998), where ndcaors ssue a sgnal whenever hey move beyond a ceran hreshold. However, our ably o esmae he opmal hreshold s mpared by he lmed frequency of annual accounng daa. Thus, we use mulvarae prob models as n Frankel and Rose (1996) o es he n-sample sascal power of accounng sgnals o predc currency crses. Berg and Pallo (1999) also use he prob model o assess he ou-of sample performance of bnary ndcaors and fnd ha prob model ouperforms he sgnal approach n erms of scores and goodness-of-f. We esmae he followng prob model for he full sample of counry-years. We nclude counry fxed effecs, a common me rend, and conemporaneous cross-seconal correlaons: 3 18 k n n 1 k 1 D _ Crss Accounng Sgnal eadng Indcaors. (4) Table X repors he resuls for varous lead-lags n he full sample. The coeffcens represen he margnal effec averaged over all observaons. Resuls show ha accounng sgnals wo years pror have no ably o predc crss onses. However, he suaon s dfferen for a one-year lead. Accounng sgnals are now collecvely sgnfcan, and he F-sasc for he hree accounng measures s (p-value <0.001). The goodness-of-f, measured usng McFadden s pseudo R-square, ncreases from o wh he ncluson of accounng sgnals. 11 Fnally, conemporaneous accounng sgnals n he las column are also sgnfcan. However, snce he crss has already occurred hs sgnfcance could parally reflec he oll of he crss on frm performance. 11 Snce McFadden s pseudo R-square can ncrease even wh an ncluson of rrelevan varables, we also assess he model s explanaory power usng adjused-mcfadden s R-square. Unrepored resuls show ha ncludng accounng sgnals one year pror o he crses ncreases adjused-mcfadden s R-square by 8%. 25

27 Ineresngly, Table X shows ha many of he leadng ndcaors also do no have sascally sgnfcan coeffcens. Among he leadng ndcaors, some domesc fnancal varables such as commercal bank depos and Excess real M1 balances appear o be sascally sgnfcan n he predced drecon. On he oher hand, coeffcens of real neres rae dfferenal, domesc real neres rae and lendng/depos rae show reverse sgns. 12 Ths s conssen wh he early warnngs sysems fndng ha even he bes model has lmed predcve power (Kamnsky and Renhar 1998, Table I). IV.C The Accounng Precson Dchoomy We now expand equaon (4) o compare coeffcens across he wo groups of accounng precson. We specfy he followng sacked prob model: 13 3 H CH n 1 D _ crss [ I Accounng Sgnal ] 3 18 k k IC Accounng Sgnal n eadng Indcaorc, n 1 k 1 [ ]. (5) I I ) s an ndcaor equal o 1 when he crss s n a counry wh hgh (low) accounng C H ( C precson and 0 oherwse (here s no nercep erm). The coeffcens H ( ) measure he assocaons beween accounng sgnals from counres wh hgh (low) accounng precson and he onse of a crss. Table XI presens he resul of he prob esmaons. As n Table X, accounng sgnals wo years n advance have no power o predc crses. However, he one-year pror F-ess clearly show ha 12 Ths fndng s conssen wh oher emprcal research n he early warnngs sysem leraure. In parcular, Edson 2003 (Table XIV) fnds ha he real neres dfferenal and real neres raes have he lowes probably of ssung a sgnal durng he 24-monh perod pror o a crss. 13 See Maddala (2001) for a dscusson of sacked regressons. Under he assumpon ha he error erms from each regresson have he same dsrbuon, hs echnque capures any (poenal) correlaons across he error erms. Sackng also allows sascal ess of coeffcens across he sacked equaons. 26

28 accounng sgnals have more sascal power o jonly predc crss among counres wh low accounng precson. In parcular, he n-sample predcon power of realzed accounng sgnals s sgnfcan for low accounng precson counres (F-sa = 14.17, p-value = 0.003), and nsgnfcan (Fsa = 1.31, p-value = 0.726) for hgh precson counres. Ths s precsely he predcon of models such as Angeleos and Wernng (2006). An alernave explanaon for he nsgnfcance of he accounng sgnals n he hgh precson counres s lack of power. The afermah model n he las column of Table XI dspels hs possbly. For he frs year followng a crss onse, he accounng sgnals appear o be jonly sgnfcan for boh he hgh (F-sa= 26.62, p-value< 0.001) and he low (F-sa= 17.67, p-value< 0.001) accounng precson subsamples. Accounng sgnals n hgh precson counres hus reflec he consequence of crses: hey smply canno predc hem. The ndvdual coeffcens of he realzed accounng sgnals n Table XI are somewha dffcul o nerpre. We canno drecly read off he profably coeffcen, for we have o keep he accrual componen consan. We herefore examne he measures ndvdually n Table XII. We ge he same resul as n Table XI, namely ha wo ou of he hree pror-year accounng measures are srongly predcve of crses n low precson counres. 14 Nohng s sgnfcan n he hgh precson counres. Accruals n low precson counres are negavely sgnfcan, suggesng declnes before he crses. Accruals declne f frms ncrease her wre-offs or decrease her nvenory buldup. Invenory buldup, a parcularly mporan measure of economc healh, s a posve accrual because consumes cash bu does no affec earnngs. 14 Unrepored resuls ndcae ha no measure ndvdually has any power o predc crses n boh subsamples wo years n advance. For one year followng a crss, some sgnals are sgnfcan n boh hgh and low precson counres. The resuls hus mrror Table XI. 27

29 The pror-year profably n low precson counres s sgnfcan bu has he oppose sgn ( does so n Table XI as well). One explanaon s ha alhough accruals are declnng, pror-year cashflows are sll boomng, causng oal profs o ncrease before he crss (a la Rancere e al. 2008). Bu hen, once he crss hs, hs boom dsappears. Commensuraely, he profably n he crses years (Table XI) does sgnfcanly declne. In sum, herefore, our resuls for he low precson counres mrror Rancere e al. (2008), who show growh (capured by our prof measures) as mprovng before he crses. Our key addonal nsgh s frms ancpae a growh slowdown and reduce accruals. Accounng adjusmens hus play precsely he role hey are supposed o. Ths s a parcularly valuable resul because shows ha accounng rules maer: s he applcaon of hese rules o measure frm operaons ha generaes crcal asseprcng nformaon. IV.D Insuonal Facors and Tme-Varyng Measures of Accounng Precson One concern wh he above resuls s ha hey reflec he underlyng nsuons n Table V, Panel A and no accounng precson. Ths s especally rue of nsuons ha are srongly correlaed wh accounng precson. In hs subsecon, we examne hs concern drecly. Table V, Panel B shows ha, of all he nsuonal feaures, legal enforcemen s mos srongly assocaed wh accounng precson. We dvde he sample no hgh and low enforcemen counres and examne he predcve power of he realzed accounng sgnals. Table XIII, Panel A shows ha n he year before he crss, realzed accounng sgnals are sronger predcors only n low qualy law enforcemen counres. We nerpre hs resul as follows. Modern research on economc growh has explored several channels hrough whch nsuons affec growh. In mos of hese channels, whch range from fnancal developmen o radng coss (e.g., Marn and Rey 28

30 2006; Acemoglu and Guerrer 2008; Rancere e al. 2008), excess volaly and self-fulfllng crses due o luck and oher sunspo phenomena are more lkely n less-developed counres wh feaures such as poor enforcemen. Prescrpons on capal accoun lberalzaons also rounely sar wh he assumpon ha less-developed counres are more suscepble o sunspo volaly (e.g., Prasad and Rajan 2008). The above lne of reasonng herefore would hen sugges ha fundamenals are less lkely o predc crses n such counres. Our resuls n Table XIII, Panel A are exacly he oppose. So he channel whch does seem o be operang n Table XIII, Panel A s lkely he one n Table XI, namely ha hgh enforcemen counres have hgh precson accounng sgnals (Table V, Panel B), whch hen fall no he purvew of our analycal crss models. Therefore, albe ndrec, Table XIII, Panel A also provdes suppor for our man predcon. Anoher aspec of nsuons s ha hey can change wh me, especally n response o crses. Angeleos e al. (2006) and Angeleos e al. (2007) analycally show ha our man predcon s robus o such endogenous nsuonal changes. As dscussed n Secon III.B.b, we re-sor he counres no hgh and low accounng precson every year, based on he accounng precson score over he prevous fve-year perod. Table XIII, Panel B presens he man resuls n Table XI usng hs me-varyng dchoomy. As n Table XI, he accounng fundamenals can predc crses only n he low precson counres. Our man resul hus connues o hold. V. Addonal Analyses V.A Dfferen Types of Crses Our approach so far has no dfferenaed among dfferen ypes of crss. In hs secon we relax hs assumpon and drop all he seven fscal defc, curren accoun, and sudden sops crses n Table I. 29