Does Accounting Conservatism Reduce Stock Price Crash Risk? Firm-level Evidence

Transcription

1 Does Accouning Conservaism Reduce Sock Price Crash Risk? Firm-level Evidence Jeong-Bon Kim Ciy Universiy of Hong Kong Liandong Zhang Ciy Universiy of Hong Kong ABSTRACT: This sudy provides srong and robus evidence ha conservaism in financial reporing reliably predics sock price crash risk. Using a large sample of U.S. firms over he period of , we find ha accouning conservaism, as measured by he Khan and Was (2009) CSCORE, reduces he likelihood of a firm experiencing sock price crashes. This finding holds even afer conrolling for firm and year fixed effecs, invesor heerogeneiy, informaion opaqueness, and oher firm-specific feaures ha are possible conribuing facors o negaive exreme oucomes in sock reurns. We furher find ha he predicive power of accouning conservaism wih respec o crash risk is more pronounced for firms wih higher informaion asymmeries, namely hose wih relaively higher R&D invesmen, higher indusry concenraion or lower produc marke compeiion, and lower analys coverage. Overall, our resuls are consisen wih he noion ha accouning conservaism limis managerial incenive and abiliy o oversae performance and hide bad news from invesors, which, in urn, reduces sock price crash risk. JEL classificaion: G12; M41 Keywords: Accouning conservaism; Sock price crashes; Negaive reurn skewness; Informaion asymmery Augus 2011 We hank Sudipa Basu, Paul Brockman, Zhihong Chen, Agnes Cheng, Yuyan Guan, Li Jiang, Mo Khan, Kalin Kolev, Jing Li, Tiemei (Sarah) Li, Yinghua Li, Mingzhi Liu, Cahy Pang, Francis Kim, Annie Qiu, Haina Shi, Dan Simunic, Byron Song, Shyam Sunder, Haiping Wang, Zheng Wang, Ross Was, Franco Wong, Cheong Yi, Yong Yu, and paricipans of research workshops a Ciy Universiy of Hong Kong, Concordia Universiy, Lehigh Universiy Accouning Conference, and Naional Taiwan Universiy for heir useful commens. J.-B. Kim acknowledges parial financial suppor for his research from he Social Sciences and Humaniies Research Council of Canada via he Canada Research Chair program and Accouning & Corporae Governance Uni a Ciy Universiy of Hong Kong. L. Zhang acknowledges financial suppor from he Faculy Research Developmen Program of he John Molson School of Business.

2 1. Inroducion Corporae managers have incenives o oversae financial performance hrough sraegically wihholding bad news and acceleraing he release of good news, hoping ha poor curren performance will be camouflaged by srong fuure performance. This asymmeric disclosure incenive sems from a variey of facors, including formal compensaion conracs and career concerns (Ball, 2009; Graham e al., 2005; Khan and Was, 2009; Kohari e al., 2009; LaFond and Was, 2008). If managers are able o wihhold and accumulae bad news for an exended period, negaive informaion is likely o be sockpiled wihin a firm. However, here is an upper limi o he amoun of bad news ha managers can absorb or successfully accumulae. This is because once he amoun of accumulaed bad news reaches a cerain hreshold, i becomes oo cosly or impossible o coninue o wihhold i. When he accumulaion of bad news reaches a ipping poin, i will all be released a once, leading o large, negaive, marke-adjused sock reurns on he individual socks concerned, ha is, sock price crashes (Huon e al., 2009; Jin and Myers, 2006). In his sudy, we invesigae he firm-level relaion beween conservaism 1 in financial reporing and sock price crashes. Conservaism refers o he accouning endency o require a higher degree of verificaion o recognize good news as gains han o recognize bad news as losses (Basu, 1997). Was (2003a, b) and LaFond and Was (2008) argue ha conservaism is a governance mechanism ha curbs managerial incenives and abiliy o accelerae he disclosure of good news and delay he release of bad news. The asymmeric verifiabiliy requiremen of conservaive accouning offses he managers endency o hide bad news and accelerae good news recogniion in audied financial saemens. Moreover, he asymmeric imeliness of earnings repored in audied financial saemens could, in urn, discipline oher managemen disclosures, such as volunary disclosures, by providing a benchmark ha makes managers ex pos accounable. Specifically, conservaive audied 1 In his sudy, we focus on condiional conservaism, as capured by he asymmeric imeliness of earnings. Throughou he paper, we use he erms conservaism, condiional conservaism, and asymmeric imeliness inerchangeably. 1

3 earnings dampen managerial incenives o disclose unverifiable favorable informaion and, insead, bring forh disclosures of unverifiable unfavorable informaion (LaFond and Was, 2008). Viewed his way, we expec ha he more conservaive a firm s accouning pracices, he lower he probabiliy ha firm-specific bad news is hidden and accumulaed. We herefore predic ha all else being equal, accouning conservaism reduces he likelihood of fuure sock price crashes. Furhermore, accouning conservaism can also reduce crash risk by disciplining managers invesmen decisions, in addiion o heir disclosure behaviors. Using a hidden acion model, Benmelech e al. (2010) show ha when a firm experiences a decline in he growh rae of invesmen opporuniies, he CEO of he firm have an incenive o inves in negaive NPV projecs o mainain he preense ha invesmen opporuniies are sill srong. However, he sraegy canno be kep forever, a some poin he firm experiences a cash shorfall, he rue sae is revealed and he sock price sharply declines as he firm needs o recapialize (Benmelech e al., 2010: 2). Moreover, Bleck and Liu (2007) make a relaed poin ha he hidden, negaive informaion abou a firm prevens he boards from discerning and liquidaing negaive-npv projecs a an early sage. This will allow bad projecs o be kep alive and heir bad performance o accumulae, unil an asse price crash. Building on he frameworks of Benmelech e al. (2010) and Bleck and Liu (2007), we argue ha condiional conservaism can reduce invesmen-aciviy-induced crashes because imely loss recogniion makes managers less likely o make invesmens hey expec ex ane o be negaive-npv, and less likely o coninue operaing invesmens wih ex pos negaive cash flows (Ball and Shivakumar, 2005: 84). For our empirical ess, we obain firm-specific measures of crash risk and conservaism, since we are ineresed in he firm-level relaion beween he wo. Following he lieraure (Chen e al., 2001; Huon e al., 2009), we proxy for firm-specific crash risk using wo measures: (i) he likelihood ha exreme negaive firm-specific weekly reurns occur and (ii) he negaive condiional skewness of firm-specific weekly reurns. Since our research quesion is relaed o he abiliy of 2

4 news-dependen, condiional conservaism o forecas fuure crash risk, we proxy for conservaism using he firm-level asymmeric imeliness measure, CSCORE, developed by Khan and Was (2009). Using a sample of 137,571 firm years over he period of , we find ha a greaer exen of conservaism, or he greaer imeliness of losses versus gains, in financial reporing significanly reduces he likelihood of a firm experiencing fuure sock price crashes. Moreover, resuls from he Cox (1972) proporional hazard model esimaion show ha conservaism also reduces he insananeous likelihood of crash occurrence, condiional on he pas hisory of crashes. The above resuls hold afer conrolling for firm and year fixed effecs, he measure of heerogeneiy of invesor beliefs of Chen e al. (2001), he measure of informaion opaqueness of Huon e al. (2009), and several oher firm-specific facors idenified in he lieraure as associaed wih sock price crashes. The resuls of robusness ess using he Basu (1997) and Ball and Shivakumar (2006; 2008) piecewise linear regressions are, overall, consisen wih our main resuls using he Khan and Was (2009) conservaism measure. In examining he cross-secional variaion in he relaion beween conservaism and crash risk, we find ha he predicive power of conservaism wih respec o fuure crash risk is sronger in an environmen where invesors are faced wih larger informaion asymmeries. Specifically, we find ha he predicive abiliy of conservaism is greaer for firms wih inensive research and developmen (R&D), firms wih higher indusry concenraion or lower produc marke compeiion, and firms wih lower levels of analys coverage. In shor, our resuls provide srong and robus evidence ha conservaism significanly reduces sock price crash risk. Our evidence is in line wih he noion ha conservaism is an equilibrium response o a sandard agency problem associaed wih managemen incenives o hide firm-specific bad news for privae gain (LaFond and Was, 2008; Was, 2003a). 3

5 This paper conribues o he lieraure in several imporan ways. Firs, our sudy adds o he conservaism lieraure. Ever since Basu (1997) firs provided sysemaic evidence for he exisence of conservaism, many sudies have examined various counry-wide and firm-specific facors ha explain he demand for conservaism in financial reporing. 2 However, exising research pays lile aenion o he economic consequences of or benefis from conservaive accouning. Three noable excepions are Ahmed e al. (2002), Wienberg-Moerman (2008), and Zhang (2008), who documen he benefis of conservaism in he deb marke. To our knowledge, however, our sudy is he firs o provide sysemaic evidence ha conservaism has desirable consequences in he equiy marke because i reduces fuure crash risk, consisen wih he heory in LaFond and Was (2008) ha equiy invesors demand conservaism. This evidence is inriguing and imporan, paricularly because crash risk has received increased aenion from boh academic researchers and he invesmen communiy, as demonsraed by he prominen volailiy smirk phenomenon observed in opion markes afer he 1987 sock marke crash. Second, our sudy conribues o he undersanding of accouning facors ha explain sock price crashes and negaive reurn skewness. This sudy provides evidence ha accouning conservaism, as a governance mechanism, is an imporan facor ha deermines fuure crash risk, and is abiliy o predic fuure crashes is incremenally significan over and beyond he measure of invesor heerogeneiy of Chen e al. (2001) and he measure of informaion opaqueness of Huon e al. (2009). Third, a unique feaure of his sudy is he employmen of boh he logi and hazard models for forecasing crash risk, whereas prior sock price crash sudies, such as ha of Huon e al. (2009), use only he logi model. Unlike he logi model, he hazard model akes ino accoun prior crash 2 See Was (2003b) for an excellen, srucured review of he earlier lieraure on he exisence of alernaive explanaions for conservaism. 4

6 hisory (occurrence, duraion beween crashes, and magniude) when evaluaing a firm s insananeous crash risk. This is imporan because fuure crash risk is influenced no only by he iming (or he duraion beween crashes) bu also by he magniude of pas crashes, and simply conrolling for he lagged crash indicaor in a logi model seing is no sufficien o address he issue (Jin and Myers, 2006). Using he hazard model approach, we are able o confirm he validiy of Chen e al. (2001) and Huon e al. (2009), as well as ha of our own findings from he logi model esimaion. Fourh, our findings will be of ineres o opions pricing researchers and opions raders. Negaive reurn skewness and negaive jump risk (i.e., risk of fuure crash) are imporan inpus o opions pricing models ha exend he original Black Scholes model, such as sochasic volailiy and jump-diffusion models (Baes, 2000; Pan, 2002). Therefore, he finding ha accouning conservaism is associaed wih lower crash likelihood and/or negaive skewness in reurns is of obvious value o opions marke paricipans who focus on ail evens. Finally, our resuls have implicaions for accouning sandard seing bodies. I appears ha he U.S. Financial Accouning Sandards Board (FASB) and he Inernaional Accouning Sandards Board (IASB) are seeking o eliminae conservaism from heir new concepual frameworks, favoring fair value or neural accouning insead (FASB, 2008). Boh he FASB and IASB argue ha conservaism inroduces biases ino financial reporing and increases informaion asymmery (LaFond and Was, 2008; Was, 2003a). However, he resuls of his sudy sugges ha conservaism is associaed wih lower sock price crash risk. This paper proceeds as follows. Secion 2 offers a brief review of he relevan lieraure and develops our hypoheses. Secion 3 describes he sample and daa, and explains our research 5

7 procedures. Secion 4 presens descripive saisics and he resuls of he mulivariae regressions. Secion 5 provides addiional analyses and robusness checks. Secion 6 presens our conclusions. 2. Lieraure review and hypohesis developmen Basu (1997) defines conservaism as capuring accounans endency o require a higher degree of verificaion for recognizing good news han bad news in financial saemens. Was (2003) aribues he exisence and prevalence of conservaism for five cenuries o he need for and he use of verifiable accouning numbers in deb and compensaion conracs, shareholder liigaion, regulaory and poliical processes, and axaion. According o Was, conservaism is a governance mechanism ha consrains managerial incenives and abiliy o oversae accouning numbers used in a conrac. More recenly, LaFond and Was (2008) focus on equiy marke demand for conservaism. They argue ha informaion asymmeries beween corporae insiders and ouside equiy invesors engender conservaism in financial reporing. This is because conservaism reduces informaion asymmery by curbing managers incenives, opporuniies, and abiliy o oversae income and ne asse values. While LaFond and Was provide empirical evidence consisen wih heir argumen, heir analysis does no focus on he economic consequences or benefis of conservaism in he equiy marke. This paper aims o complemen he line of research ha examines he informaional role of conservaism by examining he firm-level relaion beween conservaism and sock price crash risk. As menioned in he Inroducion, managers have a general endency o sraegically wihhold bad news or delay he disclosure of bad news and accelerae he release of good news. This endency may sem from a variey of managerial incenives, such as earnings-based compensaion conracs, career concerns, repuaion concerns, and empire building (see Ball (2009) for an exensive discussion). Empirically, Kohari e al. (2009) provide evidence suggesing ha managers end o delay he release of bad news o ouside invesors. The managerial endency o conceal bad news 6

8 from ouside invesors engenders crash risk, or, more generally, negaive reurn skewness. This is because he asymmeric disclosure behavior of managers leads o sockpiling wihin a firm of negaive informaion unknown o ouside invesors. When he accumulaed bad news reaches some ipping poin or when managerial incenive for hiding bad news collapses, he large amoun of negaive informaion will suddenly and immediaely be released o he marke, which leads o an abrup decline in sock price or a crash (Huon e al., 2009; Kim e al., 2010). Moreover, he hiding of bad news allows firms wih aggressive accouning o keep bad projecs for a longer period, compared wih firms wih conservaive accouning. When he accumulaed bad performance evenually surfaces, one observes sock price crashes (Bleck and Liu, 2007). This sudy predics ha accouning conservaism reduces crash risk for he following reasons. Firs, he asymmeric verifiabiliy requiremen for he recogniion of losses versus gains acceleraes he recogniion of bad news as losses, while delaying he recogniion of unverifiable good news as gains in audied financial saemens. This hus offses he managerial endency o hide bad news from ouside invesors and accelerae he release of good news o he marke (LaFond and Was, 2008). As a resul, bad news flows ino he marke in a imelier manner, compared wih unverifiable good news. One can herefore expec ha conservaism prevens bad news from being sockpiled, and hus reduces he likelihood ha a large amoun of bad news will be released o he marke a once. As a resul, he higher he level of conservaism, he lower he probabiliy ha bad news will be hidden and accumulae, and hus, he lower he crash risk. Second, by heir naure, conservaive accouning repors provide verifiable, hard informaion ha can be used as a benchmark for evaluaing he credibiliy of compeing, alernaive sources of unverifiable, sof informaion, such as managemen forecass and oher volunary disclosures of nonfinancial informaion (LaFond and Was, 2008). The availabiliy of his hard informaion may play he role of disciplining managers volunary disclosure behavior hrough ex 7

9 pos accounabiliy for heir own volunary disclosures (Ball, 2001; Ball e al., 2009a). Moreover, any reicence (wih respec o bad news) or puffery (wih respec o good news) in volunary disclosures will be discovered sooner for conservaive firms han for non-conservaive firms. For nonconservaive firms, he misleading volunary disclosures are unlikely o be discovered ill he manager has moved on, and hence, his manager is more likely o mislead ouside invesors hrough volunary disclosures. For conservaive firms, misleading volunary disclosures are likely o be discovered sooner, so heir managers are less likely o mislead ouside invesors hrough volunary disclosures. Thus, conservaism consrains he incenives and abiliy of managers o delay he release of bad news and accelerae he release of good news in volunary disclosures. This reduces crash risk, as well as he likelihood of inflaing sock price bubbles, an imporan source of crash risk. Third, while he above discussion focuses on how conservaism reduces crash risk hrough improving he flow of boh hard and sof informaion o he marke, conservaism could also reduce crash risk via is impac on real decision making. The imelier recogniion of losses han gains can be an early warning mechanism ha enables shareholders and board of direcors o promply idenify unprofiable projecs and force managers o disconinue hem. This prevens he bad performance of bad projecs from accumulaing and reduces he probabiliy of asse price crashes (Ball, 2001; Bleck and Liu, 2007). For example, Francis and Marin (2010) find ha conservaive firms ac more quickly o dives unprofiable acquired companies. The above discussions lead o he following hypohesis in alernaive form: H1: Conservaism in financial reporing reduces he likelihood of fuure crash occurrence, ceeris paribus. Noe ha alhough he crash risk models such as Jin and Myers (2006) is buil on he concep of bad news hoarding, i should no be aken lierally ha i only applies o bad news in realiy. Managers can also hide bad performance by recognizing unverifiable good news in accouning 8

10 income or disclose hem hrough oher channels. For example, Enron launched EnronOnline in 1999, and adoped mark o marke accouning o repor is performance. Enron s managers were able o hide he firm s real losses by recognizing anicipaed fuure profis from any deal of EnronOnline as if real oday. We discuss his poin o emphasize he imporance of our adoping he asymmeric verifiabiliy version of conservaism, which includes boh he concep of imely loss recogniion and he posponing of good news recogniion unil hey are verifiable. Moreover, a key poin underlying H1 is ha conservaism curbs he managers incenive o hide privae negaive informaion. However, he amoun of value-relevan, negaive informaion ha managers wihhold can vary across firms. As an example, consider a firm wih relaively high R&D invesmen. In he exreme case of no informaion asymmery, managers have no incenive for sraegic disclosure, and hus conservaism plays no role in conrolling manager disclosure behavior. However, he long-erm effecs of R&D projecs are difficul for ouside invesors o evaluae, and he informaion abou his effec is idiosyncraic o managers. This creaes informaion asymmery beween managers and ouside invesors. In an environmen of high informaion asymmery, he coss from managers wihholding negaive informaion abou an R&D projec are likely o be lower, and he associaed benefis, including he proecion of proprieary informaion, are likely o be higher. As a resul, managers are more likely o be moivaed o wihhold negaive informaion. Given evidence ha informaion asymmeries in he equiy marke engender conservaive financial reporing (LaFond and Was, 2008), we argue ha in an environmen of high informaion asymmery, conservaism plays a more imporan role in counering managerial incenive o wihhold negaive informaion, and, hus, he impac of conservaism on reducing crash risk is more pronounced for firms wih high informaion asymmery. To uncover sysemaic evidence for he above argumen, we es he following hypohesis: 9

11 H2: Conservaism in financial reporing reduces he likelihood of fuure crash occurrence o a greaer exen for firms wih high informaion asymmery han for firms wih low informaion asymmery, ceeris paribus. Our main hypoheses are based on he noion ha conservaive accouning limis he incenive and abiliy of managers o wihhold and accumulae adverse privae informaion from ouside invesors, which, in urn, leads o lower fuure crash likelihood for conservaive firms. One can argue, however, ha ouside invesors can ge access o adverse privae informaion in a imely manner via heir privae informaion search aciviies, which, in urn, reduces he likelihood of fuure crashes for non-conservaive firms. In oher words, o he exen ha he cos of privae informaion search is no prohibiively high, i could subsiue for conservaism. In such a case, here would be no significan difference in fuure crash likelihoods beween conservaive and non-conservaive firms. However, Aboody and Lev (2000), among ohers, argue ha privae informaion search is cosly and opimal informaion acquisiion by ousiders will generally fall shor of compleely exhausing a manager s privae informaion. We herefore expec ha he impac of conservaism on fuure crash risk remains significan even when marke paricipans engage acively in privae informaion search. 3. Sample and measuremen of key variables 3.1. Sample and daa Iniially, our sample is drawn from he inersecion of daa from he Cener for Research in Securiy Prices (CRSP) and Compusa for he period We hen impose he following selecion crieria: Firs, similar o Khan and Was (2009), we require ha oal asses and book values of equiy for each firm be greaer han zero and ha he share price a he fiscal year-end be greaer han $1. Second, o be included in he sample, a firm mus have a leas 26 weekly reurns for each fiscal year. Third, following Khan and Was (2009), we exclude firms in each sample year ha fall in he op and boom perceniles of earnings, annual reurns, marke value of equiy, marke-o- 10

12 book raio, or leverage. 3 Throughou his paper, our sample year is defined as he 12-monh period ending hree monhs afer fiscal year-end. We delee firm years wih missing daa for he research variables used in our regressions. Afer applying hese selecion crieria, we obain a final sample of 137,571 firm years spanning he period Measuremen of firm-specific conservaism Since our hypoheses are relaed o he abiliy of news-dependen, condiional conservaism o forecas he likelihood of fuure crash occurrence, we measure he degree of accouning conservaism for each firm in each sample year, using he firm year condiional conservaism measure, CSCORE, developed by Khan and Was (2009). To obain he CSCORE measure, we begin wih he Basu (1997) model, which is designed o capure he asymmeric imeliness of earnings in recognizing bad news versus good news. Specifically, he Basu model can be wrien o allow coefficiens o vary across firms and over ime as follows: X D R D * R , (1) where j indexes he firm, indexes he year, X represens earnings before exraordinary iems divided by marke value of equiy, R is he cumulaive marke adjused reurn over he fiscal year period, D is a dummy variable ha equals one if R < 0, and zero oherwise, and is he error erm. In Eq. (1), 4 measures incremenal imeliness for bad news over good news, or he exen of conservaism for each firm in each year. The firm year-specific coefficiens 3 (imeliness of good news) and 4 (conservaism) are hen expressed by linear funcions of firm year-specific characerisics ha are correlaed wih he imeliness of good news and conservaism: 3 All he empirical resuls remain idenical if we do no rim daa. 11

13 GSCORE MKV 3 MB 4 LEV, (2) CSCORE MKV 3 MB 4 LEV, (3) where MKV is he naural log of he marke value, MB is he marke-o-book raio, and LEV is he deb-o-equiy raio. Replacing 3 and 4 in Eq. (1) by Eqs. (2) and (3), respecively, yields he following empirical model: X D 1 D 1 2 * R 1 R ( MKV ( MKV ( MKV MB LEV D MB 4 MB 3 4 LEV LEV ) 5 4 MKV D ) MB D 6 LEV ). (4) We esimae Eq. (4) using five-year rolling panel regressions 4 and calculae our measure of conservaism, CSCORE, using Eq. (3) wih he esimaed coefficiens 1, 2, 3, and 4 from Eq. (4). Here, firms wih a higher CSCORE are considered more conservaive. Khan and Was (2009) conduc a series of ess on he properies of his conservaism measure and conclude ha he CSCORE measure capures variaions in conservaism very well. Since our research quesions are relaed o he asymmeric disclosure behavior and he asymmeric imeliness of earnings (newsdependen, condiional conservaism), he CSCORE measure fis our purpose well. Moreover, his paper hypohesizes ha for conservaive firms, he higher levels of monioring and beer governance reduce he amoun of privae informaion wihheld by managers. This hypohesis, based on hidden privae informaion, allows he use of he Basu model (and hence CSCORE) since he Basu model does no require ha he marke be efficien wih respec o privae informaion. Basu simply assumes ha he marke knows more han wha is in earnings. 4 Therefore, our CSCORE is he PC_SCORE, as in Khan and Was (2009). We use his specificaion because Khan and Was repor ha his measure of conservaism performs bes in heir horse racing ess. However, our resuls are robus o he use of Khan and Was C_SCORE. 12

14 3.3. Measuremen of firm-specific crash risk Following Huon e al. (2009) and Kim e al. (2010), we define crash weeks (exreme evens) in a given fiscal year for a given firm as hose weeks during which he firm experiences firm-specific weekly reurns 3.2 sandard deviaions below he mean firm-specific weekly reurns over he enire fiscal year, wih 3.2 chosen o generae a frequency of 0.1% in he normal disribuion. 5 The firmspecific weekly reurn, denoed by W, is defined as he naural log of 1 plus he residual reurn from he following expanded marke model regression: r, j 1 jr r r r r, (5) j m, 2 2 j m, 1 3 j m, 4 j m, 1 5 j m, 2 j where r is he reurn on sock j in week and m, r is he reurn on he CRSP value-weighed marke index in week. We include he lead and lag erms for he marke index reurn o allow for nonsynchronous rading (Dimson, 1979). Specifically, he firm-specific weekly reurn for firm j in week is W ln( 1 ) j, j.. Our firs measure of crash likelihood for each firm in each year, denoed by CRASH, is an indicaor variable ha equals one for a firm year ha experiences one or more crash weeks (as defined above) during he fiscal year period, and zero oherwise. Following Chen e al. (2001) and Kim e al. (2010), our second measure of crash likelihood is he negaive condiional reurn skewness (NCSKEW) measure. Specifically, we calculae NCSKEW for a given firm in a fiscal year by aking he negaive of he hird momen of firm-specific weekly reurns during he same fiscal year, and dividing i by he sandard deviaion of firm-specific weekly reurns raised o he hird power. Specifically, for each firm j in year, we obain NCSKEW as n( n 1) 3 W j ( n 1)( n 2) W j NCSKEW. (6) Our definiion of crash resuls in subsanial negaive weekly reurns. Unabulaed saisics show ha he mean (median) firmspecific reurn for crash weeks is -20.7% (-18.6%), and he mean (median) raw reurn is -22.2% (-20.0%). 13

15 We inroduce his second measure of crash risk for wo major reasons. Firs, one may suspec ha less conservaive firms are, in general, relaed o longer ails; ha is, hey have no only more crashes bu also more posiive jumps. The use of negaive skewness as an alernaive measure miigaes his concern. 6 Second, some opion and asse pricing applicaions require fuure reurn skewness as an inpu. Building a model ha predics skewness could hus conribue o his line of research (Barberis and Huang, 2008; Boyer e al., 2010). 4. Empirical resuls 4.1. Descripive saisics and correlaion marix Table 1 presens descripive saisics for he major variables discussed in Secion 3, along wih addiional variables ha are used as conrol variables in our mulivariae analysis. Appendix I provides deailed definiions of all variables. The mean value of CRASH is 0.12, suggesing ha, on average, 12% of firm years experience one or more firm-specific weekly reurns ha fall wihin 3.2 sandard deviaions below he annual mean. Though no abulaed, a closer look a he daa reveals ha only less han 0.2% of firm years experience wo crash evens during a sample year, and only one firm year experiences more han wo crash evens (hree) during a sample year. The mean and median values of NCSKEW are and , respecively. Here, NCSKEW is slighly lower han he values repored by Chen e al. (2001), which is expeced since hese auhors use daily reurns o consruc heir variables (Das and Sundaram, 1999). The mean and median values of CSCORE are and 0.084, respecively, similar o hose repored by Khan and Was (2009). Table 2 presens a Pearson correlaion marix for all he variables used in our regression analysis. The wo measures for crash risk, CRASH and NCSKEW, are significanly and posiively 6 To furher address his concern, we also consruc a variable COUNT, which is he difference beween he frequency of exreme negaive reurns and he frequency of exreme posiive reurns (Jin and Myers (2006)). We hen rerun all regressions by replacing NCSKEW wih COUNT. Though no repored, we find ha all he regression resuls repored in he paper are qualiaively similar o hose using his alernaive dependen variable. 14

16 correlaed wih each oher. The year conservaism measure, CSCORE, is significanly and negaively correlaed wih he wo measures of year + 1 crash risk, which is consisen wih our predicion ha more conservaive firms have lower crash risk Tes of H Logisic regressions of crash likelihood on conservaism To es wheher more conservaive firms experience lower crash risk (H1), we firs esimae he following logi model ha links conservaism in year wih he likelihood of sock price crash in year + 1 (firm subscrips are suppressed): CRASH m h 1 0 1CSCORE q ( q ConrolVariables ), (7) q 2 where CRASH +1 is an indicaor variable ha equals one if a firm experiences one or more crash evens in year + 1, and zero oherwise, and CSCORE refers o he Khan and Was (2009) conservaism measure in year. Hypohesis H1 ranslaes as 1 0. To isolae he effec of conservaism on crash risk from he effecs of oher variables, we include several conrol variables known o influence crash likelihood. Our main conrol variables are hose used in Chen e al. (2001), ha is, derended share urnover (DTURN ), negaive skewness of firm-specific weekly reurns (NCSKEW ), sandard deviaions of firm-specific weekly reurns (SIGMA ), firm-specific average weekly reurns (RET ), and firm size (SIZE ). We conrol for he derended share urnover in year because Chen e al. show ha i proxies for differences of opinion among invesors and has a significan posiive impac on negaive reurn skewness or crash risk in year + 1. Firms wih high reurn skewness in year are likely o have high reurn skewness in year + 1 as well (Chen e al., 2001). To conrol for his possibiliy, we include NCSKEW in Eq. (8). We 15

17 conrol for weekly reurn volailiy (SIGMA ) because socks wih high reurn volailiy in year are more likely o experience crashes in year + 1. Chen e al. (2001) provide evidence ha pas reurns have predicive power wih respec o fuure crash risk. In paricular, he auhors find ha fuure crash risk is higher for socks wih higher pas reurns (as far back as 36 monhs). We herefore conrol for pas one-year average weekly reurns (RET ). To conrol for he size effec, we include firm size (SIZE ) measured by he naural log of sales raher han he naural log of marke capializaion, because he laer is one of hree major inpus for compuing our CSCORE measure, as shown in Eq. (3). In alernae specificaions, we also include he marke-o-book raio (MB ) and financial leverage (LEV ) as addiional conrol variables. I should be poined ou, however, ha he resuls from hese specificaions may suffer from mulicollineariy problems, since hese wo variables are also used o consruc CSCORE, as shown in Eq. (3). Furhermore, poenial effecs of MB and LEV on CRASH +1 are likely o be capured by oher conrol variables, such as year sock reurns and reurn volailiies. Therefore, we do no include such variables in our main regression specificaions. Finally, we also esimae alernaive regression specificaions where he Huon e al. (2009) measure of informaion opaqueness (OPAQUE ) and fuure operaing performance (ROA +1 ) are addiionally included as conrols. By including OPAQUE, we inend o (i) validae he effecs of informaion opaqueness on crash risk, as evidenced in Huon e al. (2009) and Jin and Myers (2006), using our sample, and (ii) ensure ha our conservaism measure has he incremenal predicive power for crash risk over and beyond OPAQUE. We measure OPAQUE using he procedures followed by Huon e al., he deails of which are provided in Appendix. Finally, similar o Huon e al., we also include ROA +1, o conrol for possible conemporaneous relaions beween profiabiliy and crash risk. 7 7 We do no include ROA +1 in our main specificaion because our exercise is o forecas crash risk in year +1 given all informaion available in year. 16

18 Table 3 repors he logisic regression resuls for Eq. (7). To address poenial cross-secional and serial dependence in he daa, we repor p-values (wo-ailed) ha are based on robus sandard errors adjused for firm and year double clusering. Following Peersen (2009), all regressions in Table 3 also include year dummies, o conrol for year fixed effecs. Model 1 presens he resuls of our baseline regressions of CRASH +1 on CSCORE, where our major conrol variables, namely, DTURN, NCSKEW, SIGMA, RET, and SIZE are included. Noe ha hese conrol variables are similar o he se of crash deerminans examined by Chen e al. (2001). In model 1, he coefficien of our key variable of ineres, ha is, CSCORE, is highly significan, wih an expeced negaive sign and p = 0.00, suggesing ha conservaism in year reduces crash risk in year + 1, even afer conrolling for oher deerminans of crash risk. In model 2, where we inroduce wo addiional variables, MB and LEV, ino he regressions, he coefficien of CSCORE +1 remains significanly negaive, wih p = To assess he economic significance of our es resuls, using he coefficiens in model 1, we compue he marginal effec of CSCORE, ha is, he change in CRASH (he probabiliy of a crash) arising from a change of one sandard deviaion in CSCORE, holding all oher independen variables a heir mean values. The marginal effec of CSCORE in model 1 is abou , suggesing ha a one sandard deviaion increase in CSCORE resuls in a 3.1% decrease in he probabiliy of a crash. This is economically significan, given ha he average, uncondiional probabiliy of a crash in our sample is 12%, as repored in Table 1. Finally, using a reduced sample of firm years, we esimae our logisic regression in Eq. (8) afer including wo addiional variables ha are considered in Huon e al. (2009), ha is, OPAQUE and ROA +1. The resuls are repored in model 3. We find ha he coefficien of CSCORE remains negaive and significan, wih p = 0.013, suggesing ha he predicive abiliy of conservaism for 17

19 crash likelihood is incremenal over and beyond prior period accouning opaqueness, curren period profiabiliy, and oher firm-specific deerminans of fuure crash risk. Turning o he conrol variables, we observe several ineresing findings. Firs, he coefficien of DTURN is significanly posiive, wih p = 0.00, across all models. In Chen e al. (2001), his derended share urnover variable is he key es variable ha proxies for invesor belief heerogeneiy or differences of opinion among invesors. Chen e al. (2001) examine he effec of DTURN on negaive reurn skewness, bu no is effec on exreme oucomes, namely, crash probabiliy (CRASH). Our resuls herefore provide corroboraing evidence for he heory of Chen e al. ha invesor heerogeneiy increases crash risk. Second, he coefficien of he opaqueness measure (OPAQUE ) of Huon e al. (2009) is significanly posiive, wih p = 0.004, which is consisen wih he auhors findings. Third, wih respec o he relaive significance of CSCORE, DTURN, and OPAQUE, unabulaed resuls show ha he marginal effec of CSCORE is significanly greaer han hose of boh DTURN and OPAQUE. Fourh, we find ha he signs of he coefficiens of pas skewness (NCSKEW ) are significanly posiive, consisen wih Chen e al. (2001). We find he coefficiens of pas reurn volailiy (SIGMA ) in models 1 and 2 o be posiive bu insignifican, while significanly posiive in model 3. We also find ha he coefficiens of he sock reurn (RET ) and marke-o-book raio (MB ) are, overall, significanly posiive, which is consisen wih he sochasic bubble heory, ha socks wih high pas reurns and growh socks are more crash prone (Harvey and Siddique, 2000). The coefficiens of firm size are significanly posiive, wih p = 0.00, in model 3, while hey are insignifican in models 1 and 2. Huon e al. (2009) also repor a significanly posiive coefficien for SIZE, suggesing ha large firms are more likely o experience crashes. However, we are unaware of any heory ha predics a posiive relaion beween SIZE and CRASH +1. The coefficiens of LEV are insignifican in model 2, while significanly negaive, wih p = 0.043, in model 3. Finally, we find 18

20 ha ROA +1 is negaively associaed wih CRASH +1, which is consisen wih he findings of Huon e al. (2009). To obain furher insighs ino he inverse relaion beween CSCORE and CRASH +1, we consruc decile porfolios based on he ranked values of CSCORE a he beginning of each sample year. We hen compue he implied likelihood of fuure crashes, CRASH +1. In so doing, we use he esimaed coefficiens for Eq. (7) repored in model 1 of Table 4, wih all variables oher han CSCORE se equal o heir sample means (repored in Table 2). We hen plo he implied likelihood values agains he mean CSCORE for each decile. As illusraed in Figure 1, crash likelihood in year + 1 decreases monoonically as we move from he lowes CSCORE decile o he highes. We also noe ha he relaion beween he wo variables, overall, appears o be linear, hough he impac of conservaism on reducing crash risk is more pronounced in he wo exreme deciles. For example, he implied crash likelihood decreases from o as we move from he boom decile o he op decile of CSCORE, which accouns for 23.3% (= 2.8/12) of he variaion in crash risk. In shor, he CSCORE CRASH relaion depiced in Figure 1 provides furher evidence ha more conservaive firms experience a lower likelihood of fuure sock price crashes. Collecively, he resuls repored in Table 3 and Figure 1 reveal ha, consisen wih H1, he higher he conservaism in year, he lower he likelihood of crashes in year + 1, and his relaion is highly significan across all specificaions. This resul holds even afer conrolling for he measures of invesor heerogeneiy of Chen e al. (2001) and he measure of informaion opaqueness of Huon e al. (2009). Our resuls are, overall, consisen wih he view ha conservaism plays a significan role in limiing managemen incenives and is abiliy o wihhold bad news or delay he iming of is disclosure, hereby lowering he probabiliy of bad news being sockpiled wihin a firm and hus reducing he likelihood of a sock price crash. 19

21 OLS regression of negaive reurn skewness on conservaism To uncover furher evidence on he relaion beween conservaism and crash risk, we also use he negaive condiional skewness (NCSKEW) of he weekly firm-specific reurn disribuion (Chen e al., 2001) as an alernaive proxy for fuure crash risk. Table 4 repors he resuls of OLS regressions for Eq. (9) using NCSKEW +1 as he dependen variable. As in Table 3, all repored p-values are adjused using sandard errors correced for firm and year double clusering. As shown in Table 4, he coefficiens of CSCORE are significanly negaive, wih p = 0.00, across all models, which srongly suppors he predicion in H1. This resul is economically significan as well. Consider he resuls in model 1 as an example. The CSCORE coefficien of indicaes ha a one sandard deviaion increase in CSCORE leads o an approximaely 40% (= 1.293*0.07/0.229) decrease in NCSKEW +1. The signs and significance of oher coefficiens are, overall, consisen wih hose repored in Table 3. Consisen wih he finding of Chen e al. (2001), he coefficiens of DTURN and RET are all significanly posiive across all models, wih p = Noe also ha he coefficien on OPAQUE is also significanly posiive, wih p = 0.024, in model The Cox proporional hazard model approach Jin and Myers (2006) argue ha ime can ener invesors assessmen of crash probabiliies because he probabiliy increases as ime passes. To incorporae his ime effec, we employ he Cox (Cox, 1972) proporional hazard mehod: ln h jk ( ) ( m h j( k 1) ) 1CSCORE jk q ( q ConrolVariable jk ) jk, (8) q 2 20

22 where h jk () is he hazard, or insananeous likelihood of crash occurrence, for firm j a ime, condiional on k crashes having occurred in firm j by ime ; 8 j( k 1) is he ime of he (k - 1)h even; and (.) is an unspecified funcion ha capures he baseline hazard. Hypohesis H1 ranslaes as 1 < 0, which can be inerpreed in such a way ha he hazard of crash occurrences decreases wih conservaism, or he insananeous likelihood of crash occurrences decreases wih conservaism, given pas crash hisory. An imporan feaure of he above hazard model ha disinguishes i from he logi model in Eq. (7) is ha he former akes ino accoun boh he iming and magniude of pas crash occurrences when forecasing a firm s insananeous crash risk. The bankrupcy and insurance lieraure shows ha he hazard model is more accurae in idenifying facors ha can predic rare evens (Lee and Urruia, 1996; Shumway, 2001). To esimae he hazard model in Eq. (8), we idenify a sample of firms wih a leas one crash even during he sample period. For each firm crash even, we calculae he crash inerval, which is he lengh of ime (in weeks) from he curren firm crash even o he nex. If no furher firm crash even is observed, he inerval is he lengh of ime from he curren even unil he firm s delising dae or he ending dae of he sample period, whichever occurs firs. The conrol variables are he same as in Eq. (7) and year dummies are included. The model is esimaed by parial likelihoods using he well-known mehod of sraificaion (Cox, 1975). The parial likelihood esimaion makes i possible o esimae 1 o m wihou specifying a paricular funcional form of (.). Firm-level sraificaion allows differen firms o have differen baseline hazard funcions, while consraining he coefficiens o be he same across firms (Allison, 2005). Table 7 repors he esimaed coefficiens 8 Specifically, he hazard funcion h ) lim o firm j by ime. ( j 0 Pr[ N j ( ) N j ( ) 1], where N j () is he number of evens ha have occurred 21

23 and p-values for he sraified hazard model regressions. All repored p-values are adjused using sandard errors correced for firm and year double clusering. As shown in Table 5, he coefficiens of CSCORE are significanly negaive in all models, wih p = 0.00, which srongly suppors H1. To assess he economic significance of our es variable, consider he resuls repored in model 1 as an example: The coefficien of CSCORE is , suggesing ha a one sandard deviaion increase in CSCORE leads o an approximaely 25% (= 1 - e ( ) ) reducion of he subsequen crash hazard rae, even afer conrolling for all oher deerminans of crash occurrence. These resuls can also be inerpreed in such a way ha he insananeous crash likelihood of conservaive firms a ime is lower han ha of aggressive firms, condiional on k crashes having occurred by ime. Table 5 also shows ha he coefficiens of DTURN and OPAQUE are significanly posiive, which lends furher suppor o he findings of Chen e al. (2001) and Huon e al. (2009) Tes of H2: does informaion asymmery maer? Hypohesis H2 predics ha he impac of conservaism on reducing he likelihood of fuure crashes is more pronounced for firms wih high informaion asymmery han for firms wih low informaion asymmery. To es his hypohesis, we consider hree proxies for informaion asymmery beween managers and equiy marke paricipans: (i) The relaive amoun of R&D invesmen. Prior lieraure argues ha R&D invesmen is a major source of privae informaion from he invesor s perspecive (Aboody and Lev, 2000). Many R&D projecs, such as new drugs or sofware programs under developmen, are unique o he firms concerned, whereas mos capial invesmen projecs share common characerisics across firms. Therefore, i is difficul for ouside invesors o infer he produciviy and value of a firm s R&D from observing he R&D performance of oher firms. In addiion, unlike many oher physical and 22

24 financial asses, here is no organized marke for R&D and hence no asse prices from which o derive valuaion implicaions of firm-specific R&D. Aboody and Lev (2000) provide evidence suggesing ha R&D is a major conribuor o informaion asymmery beween corporae insiders and ousiders, and hus an imporan source of insider gains. In ligh of H2, we expec ha he impac of conservaism on reducing crash risk is more pronounced for more R&D-inensive firms. (ii) The degree of indusry concenraion or produc marke compeiion. Economiss argue ha produc marke compeiion miigaes managerial agency problems (Giroud and Mueller, 2010). Dhaliwal e al. (2008) provide evidence ha inense produc marke compeiion induces managers o be more conservaive in financial reporing. Ali e al. (2009) find ha firms in more concenraed indusries (wih, herefore, low compeiion) have a more opaque informaion environmen. This finding suggess ha informaion asymmeries are higher for firms wih high indusry concenraion. Thus, we expec ha he impac of conservaism on reducing crash risk is accenuaed for firms wih high indusry concenraion or low produc marke compeiion. (iii) Analys coverage. Financial analyss play an imporan role of informaion inermediaion beween managers and less-informed ouside invesors. Furhermore, analyss play a role in monioring managerial disclosure behavior (Ball, 2001). Evidence shows ha analyss informaion inermediaion and/or monioring is value-adding because i reduces informaion asymmery beween corporae insiders and ousiders (Lang e al., 2003). Yu (2008) finds ha firms wih high analys coverage engage less in opporunisic earnings managemen, a finding consisen wih he monioring role of analyss. The above findings, aken ogeher, sugges ha informaion asymmery in he equiy marke is lower for firms wih higher analys coverage. In ligh of H2, we expec ha he impac of conservaism on reducing crash risk is aenuaed for firms wih high analys coverage. 23

25 Table 6 repors he resuls from he augmened model of Eq. (7), where CRASH +1 is he dependen variable and hree proxies for informaion asymmery and heir ineracions wih our measure of conservaism, CSCORE, are addiionally inroduced. Table 7 repors he same resuls, using NCSKEW +1 as he dependen variable. In boh Tables 6 and 7, R&D is an indicaor variable ha equals one for firms wih R&D invesmen in year, and zero oherwise; HICON is an indicaor variable ha equals one if firms have an above-median Herfindahl index in year, and zero oherwise; and NEGCOV is he naural log of 1 plus he number of analyss following a firm in year, muliplied by minus one. For all hree measures, higher values indicae higher informaion asymmery. In all regressions, we include he same se of conrol variables, ha is, DTURN, NCSKEW, SIGMA, RET, and SIZE. Ai and Noron (2003) and Noron e al. (2004) demonsrae ha boh he effecs and sandard errors of ineracion erms in logi or probi models are biased (wih heir coefficiens someimes exhibiing opposie signs) and sugges a mehod o correc for hese biases. Accordingly, we follow heir suggesion when esimaing he magniude and sandard errors of he ineracion effec in logi models: In Table 6, for non-ineracion erms, we esimae he coefficiens and sandard errors using he double-clusering mehod, as in Table 3. For ineracion erms, we use he procedure of Noron e al. (2004) o esimae he effecs and sandard errors. 9 The resuls in boh Tables 6 and 7 show ha he coefficiens of CSCORE*R&D, CSCORE*HICON, and CSCORE*NEGCOV are all significanly negaive, suggesing ha he impac of conservaism on reducing he likelihood of crash occurrence is more pronounced for firms wih higher informaion asymmery. Similar o H1, we also use he hazard model approach o provide furher suppor for H2, wih Table 8 reporing he resuls. Overall, he resuls in Table 8 are 9 The implicaions are same if we do no use he Noron e al. (2004) procedure. 24

26 consisen wih hose repored in Tables 6 and 7. Overall, he resuls presened in Tables 6 hrough 8 provide srong suppor for H2. 5. Addiional ess and robusness checks 5.1. Longer forecas windows In our logi and OLS regressions, we only examine he relaion beween he curren year s conservaism and crash probabiliy in he one-year-ahead forecasing window. I is ineresing o furher examine how far ou our conservaism predics fuure crash risk. Toward his end, we now expand he measuremen window of crash risk ino wo- and hree-year-ahead windows. Specifically, we esimae CRASH and NCSKEW using firm-specific weekly reurns during he wo- and hree-year periods saring hree monhs afer he curren fiscal year-end. In so doing, we require a leas 100 and 150 weekly reurns available for each firm for he wo- and hree-year window ess, respecively. Using he wo- or hree-year crash risk measure as our dependen variable, we re-esimae our baseline predicion model, namely, model 1, of Tables 3 and 4 and repor he esimaed resuls in Table 9. Panel A of Table 9 displays he logisic regression resuls. As shown in panel A, he coefficiens of CSCORE are significanly negaive for boh model 1 (wo-year-ahead forecasing window, dependen variable: CRASH +2 ) and model 2 (hree-year-ahead forecasing window, dependen variable: CRASH +3 ). Panel B of Table 9 presens he resuls of OLS regressions wih NCSKEW as a measure of crash risk. Again, he coefficiens of CSCORE are significanly negaive for boh he wo- and hree-year forecasing windows (dependen variable: NCSKEW +2 and NCSKEW +3, respecively). The above resuls indicae ha he predicive abiliy of conservaism for fuure crash risk is significan and robus, even when he measuremen window of crash likelihood is 25