NBER WORKING PAPER SERIES LEARNING ABOUT CEO ABILITY AND STOCK RETURN VOLATILITY. Yihui Pan Tracy Yue Wang Michael S. Weisbach

Transcription

1 NBER WORKING PAPER SERIES LEARNING ABOUT CEO ABILITY AND STOCK RETURN VOLATILITY Yihui Pan Tracy Yue Wang Michael S. Weisbach Working Paper hp:// NATIONAL BUREAU OF ECONOMIC RESEARCH 15 Massachuses Avenue Cambridge, MA 2138 March 213 We hank Hengjie Ai, Frederico Belo, Miriam Schwarz-Ziv, Berk Sensoy, Yingdi Wang, Jun Yang, Jianfeng Yu, and seminar paricipans a Universiy of Minnesoa for helpful suggesions. The views expressed herein are hose of he auhors and do no necessarily reflec he views of he Naional Bureau of Economic Research. NBER working papers are circulaed for discussion and commen purposes. They have no been peerreviewed or been subjec o he review by he NBER Board of Direcors ha accompanies official NBER publicaions. 213 by Yihui Pan, Tracy Yue Wang, and Michael S. Weisbach. All righs reserved. Shor secions of ex, no o exceed wo paragraphs, may be quoed wihou explici permission provided ha full credi, including noice, is given o he source.

2 Learning abou CEO Abiliy and Sock Reurn Volailiy Yihui Pan, Tracy Yue Wang, and Michael S. Weisbach NBER Working Paper No March 213 JEL No. G32,G34,M12,M51 ABSTRACT When here is uncerainy abou a CEO s qualiy, news abou he firm causes raional invesors o updae heir expecaion of he firm s profiabiliy for wo reasons: Updaes occur because of he direc effec of he news, and also because he news can cause an updaed assessmen of he CEO s qualiy, affecing expecaions of his abiliy o generae fuure cash flows. As a CEO s qualiy becomes known more precisely over ime, he laer effec becomes smaller, lowering he sock price reacion o news, and hence lowering he sock reurn volailiy. Thus, in addiion o uncerainy abou fundamenals, uncerainy abou CEO qualiy is also a source of sock reurn volailiy, which decreases over a CEO s enure as he marke learns he CEO s qualiy more accuraely. We formally model his idea, and evaluae is implicaions using a large sample of CEO urnovers in U.S. public firms. Our esimaes indicae ha here is saisically significan and economically imporan marke learning abou CEO abiliy, even for CEOs whose appoinmens appear o be unrelaed o heir predecessors performance. Also consisen wih he learning model is he fac ha he learning curve appears o be convex in ime, and learning is faser when here is higher ex ane uncerainy abou he CEO s abiliy and more ransparency abou he firm s prospecs. Overall, uncerainy abou managemen qualiy appears o be an imporan source of sock reurn volailiy. Yihui Pan Deparmen of Finance 1655 Campus Cener Drive Universiy of Uah Sal Lake Ciy, UT Yihui.Pan@business.uah.edu Tracy Yue Wang Carlson School of Managemen Universiy of Minnesoa h Avenue Souh Minneapolis, MN wangx684@umn.edu Michael S. Weisbach Deparmen of Finance Fisher College of Business Ohio Sae Universiy 21 Neil Ave. Columbus, OH 4321 and NBER weisbach.2@osu.edu

3 1. Inroducion In recen years, CEO changes have become highly visible evens, and are ofen porrayed as porens of a rosier fuure for he company. Presumably, a new CEO can influence a corporaion s aciviies, and ulimaely is profis, in a meaningful manner. Ye, a new CEO also brings subsanial uncerainy o he firm; i is impossible o know for sure wha he paricular decisions he CEO will make and he sraegic direcion he will ake, le alone he overall effec of he CEO on he firm s value. When a firm ges a new CEO, his uncerain abiliy o change he firm s value will be revealed over ime o he marke. The process hrough which he marke learns abou his abiliy will affec he way in which i responds o news abou he firm, and consequenly will impac he firm s sock reurn volailiy. This paper explores he idea ha he marke s learning abou CEO abiliy will affec sock reurn volailiy from boh heoreical and empirical perspecives. We firs presen a model in which he firm s cash flow sream follows a random process, wih he drif of ha process depending on an unknown abiliy of he CEO o add value. The firm s value depends on he marke s assessmen of he CEO s abiliy, and he marke updaes his assessmen when i receives any relevan informaion abou i. Thus, when here is news abou he firm, he firm s value changes for wo reasons: Firs, here is a direc effec of he news on he firm s expeced cash flows, and second, he news will change he marke s expecaion of he manager s qualiy and herefore influence is expecaion of fuure cash flows. Over ime, as he CEO becomes more of a known quaniy, he marke s updaes of is expecaion abou his qualiy become smaller condiional on a paricular signal, so ha he firm s sock price will move less for a given piece of informaion. Therefore, a firm s sock reurn volailiy should decline wih he CEO s enure. The model conains a number of predicions abou he relaion beween he firm s sock reurn volailiy and he CEO s enure. The model implies ha he sensiiviy of sock reurn volailiy o CEO enure depends on he raio of he variance of he unknown abiliy o he variance of he firm s fundamenals. If uncerainy abou he CEO s abiliy is resolved over ime, hen volailiy should decline wih CEO enure. The rae of his decline should be higher when uncerainy abou CEO abiliy is higher. Thus, as uncerainy abou he CEO s abiliy decreases because of marke learning, he rae a which he 1

4 volailiy decreases wih CEO s enure also declines. Consequenly, he model implies ha he volailiyenure slope should be convex. We evaluae hese predicions empirically using a sample of 1,873 CEO urnovers in 1,582 U.S. publicly raded firms occurring beween 1992 and 26. If CEOs were irrelevan for firm value, hen here would be no relaion beween volailiy and CEO enure; however if CEOs creae or desroy value, hen he marke should updae is assessmen of heir abiliies o do so, leading o more precise esimaes of abiliy and lower subsequen sock reurn volailiy. Our esimaes indicae ha here is a robus relaion beween CEO enure and he firm s sock reurn volailiy: Volailiy increases around he ime of CEO urnover, and hen decreases subsequenly. The magniudes of he effecs are subsanial; idiosyncraic reurn volailiy declines by 14% and oal reurn volailiy declines by 1% over he 36 monhs afer he CEO ook office. This paern is consisen wih he predicions of he learning model, in ha here is likely o be large uncerainy abou he new leadership a he ime of he urnover, and afer he CEO change, volailiy declines as he CEO s abiliy becomes known more precisely. An alernaive inerpreaion o learning for hese resuls is ha CEO urnovers end o occur a imes of high fundamenal uncerainy, so ha he pos-urnover decline in volailiy simply reflecs reversion o a normal level of volailiy. We empirically assess he exen o which he paerns in volailiy over CEO enure reflec learning or endogenous iming of urnovers. To do so, we esimae he sensiiviy of volailiy o enure subsequen o a subsample of urnovers ha are arguably exogenous: urnovers due o deahs, healh issues, and reiremens of he deparing CEOs. For his subsample of urnovers, here is sill a decline in volailiy wih he enure of he replacemen CEO, alhough he decline is smaller for his subsample of urnovers han for he subsample of forced urnovers. This finding suggess ha alhough many CEO urnovers are nonrandom and end o occur a imes of high fundamenal volailiy, here noneheless is learning abou CEO abiliy subsequen o all urnovers ha is refleced in sock reurn volailiy. Anoher reason why firms fundamenal volailiies could change subsequen o CEO urnovers is ha CEO urnovers are ofen followed by subsanial changes in he firm. These changes eiher reflec he 2

5 vision of he new leadership (e.g., expansion, divesiure, new produc developmen) or occur because of revelaion of (negaive) informaion abou he firm s fundamenals ha had been wihheld by he previous managemen (e.g., accouning wrie-off, earnings resaemen, fraud invesigaion). Pos-urnover real acions or informaion releases could affec volailiy eiher direcly by changing he risk (or he perceived risk) of he asses, or indirecly by conveying informaion ha affecs he marke s learning abou he CEO s abiliy. We conrol for boh he direc and indirec effecs of various acions enaced by new CEOs. We find ha he volailiy-enure sensiiviy is saisically significan and economically imporan regardless of wheher here are subsanial acions afer urnover. The model also conains predicions abou he ime series and cross-secional paerns in he speed a which he marke learns abou he CEO s abiliy: In paricular, i suggess ha he learning speed should decrease over ime, and i should increase wih he iniial uncerainy abou he CEO s abiliy and wih he informaiveness of signals available o he marke. To es hese predicions, we firs use boh polynomial and spline specificaions o esimae he curvaure of he volailiy-enure sensiiviy, which reflecs he learning speed. Our esimaes indicae ha he learning curve is convex, wih learning being much faser in he firs year of he new CEO s enure han in he second and hird years. The convexiy in he learning speed is consisen wih he inuiion ha a given signal affecs learning more a he beginning of a CEO s enure when uncerainy abou he managemen is highes. To es he predicions abou he cross-secional deerminans of he learning speed, we esimae he sensiiviy of volailiy o enure for each new CEO in our sample and hen measure he exen o which i is relaed o he firm s informaion environmen and he level of prior uncerainy abou he CEO s abiliy. The resuling esimaes sugges ha learning abou CEO abiliy is faser in more ransparen firms and for CEOs wih higher prior uncerainy (i.e., ousider CEOs, younger CEOs, and less experienced CEOs). These findings are consisen wih he noion ha learning abou CEO abiliy is faser when here is more uncerainy abou he abiliy, and also when signals abou ha abiliy are more informaive. An implicaion of he model is ha a given piece of news will have a larger impac on he firm s sock price when uncerainy abou he CEO s abiliy is larger. We es his implicaion direcly by 3

6 considering he way in which he absolue value of sock price reacions o news varies over he CEO s enure. We consider four ypes of announcemens: expansions/downsizing, new producs, dividend changes, and earnings surprises. For each ype, he absolue value of sock price reacions declines significanly over a CEO s enure, wih he rae of decrease becoming smaller over ime, similar o he finding of he convex volailiy-enure slope. This resul is consisen wih he view ha a componen of sock price reacions o news is informaion abou he CEO s abiliy, and ha his componen declines in imporance as he CEO s abiliy becomes beer known over ime. Finally, he model allows us o quanify he imporance of uncerainy abou CEO abiliy relaive o he firm s fundamenal cash flow uncerainy. Our esimaes show ha a he ime of CEO urnover, uncerainy abou managemen qualiy conribues o 26% o 29% of he oal sock reurn volailiy. The impac of uncerainy abou managemen qualiy also exhibis significan heerogeneiy across differen manager ypes and firm ypes. For example, uncerainy abou younger CEOs is more han wice as much as he uncerainy abou older CEOs, relaive o he firm s fundamenal uncerainy. Alhough hese esimaes are poenially sensiive o he assumpions in he model, hey do provide iniial esimaes of he exen o which uncerainy abou managemen qualiy, as well as he uncerain naure of he policies managemen will adop, can conribue o he overall firm uncerainy and sock reurn volailiy. Uncerainy abou managemen appears o be a non-rivial source of uncerainy ha affecs sock price movemens. Overall, he resuls srongly sugges ha he process of he marke s coninual evaluaion of a firm s managemen qualiy affecs he volailiy of he firm s sock reurn. These adjusmens accoun for a reasonable fracion of he firm s overall sock price movemens. Numerous paerns in he daa sugges ha he process by which he marke learns abou he firm s managemen qualiy can be well characerized by a Bayesian learning model. More imporanly, his analysis implies ha here are subsanial differences in managerial qualiy, and hese differences lead o imporan differences in firm performance. The paper spans he usual dichoomy in finance research beween corporae finance and asse pricing, and is relaed o lieraures in each subfield. This paper builds on a lieraure wihin asse pricing 4

7 focusing on he way in which learning abou fundamenals affecs sock reurn volailiy. Early work by Timmermann (1993) shows ha such learning can help resolve he excess volailiy puzzle posed by Shiller (1981). Pasor and Veronesi (23) develop a sock valuaion model in he presence of learning abou he average profiabiliy. The model predics ha sock valuaion increases wih uncerainy abou average profiabiliy, and declines over a firm s lifeime as such uncerainy decreases due o learning. Pasor and Veronesi (29) survey a number of oher relaed papers, which show how learning can help explain a wide range of asse pricing phenomena, including predicabiliy of reurns, sock price bubbles, porfolio choices, among ohers. Mos recenly, Pasor and Veronesi (212) use he learning framework o undersand he impac of uncerainy abou governmen policy on sock prices. The model presened below combines insighs from his asse pricing lieraure on he effecs of learning, wih specific learning feaures inheren when a firm s profiabiliy depends on he unknown abiliy of he manager. As such i draws on a lieraure inspired by Holmsröm (1999) ha explains aspecs of managemen incenives and governance using he learning process abou managemen abiliy as one key ingredien (see Gibbons and Murphy (1992), Hermalin and Weisbach (1998, 212), Hermalin (25), and Taylor (21)). 1 Two paricularly relaed papers are Clayon, Harzell and Rosenberg (25) and Taylor (212). Clayon e al. (25) documen an increase in sock reurn volailiy around CEO urnovers. Our work exends his analysis, formalizing he relaion beween CEO urnover and sock reurn volailiy in a framework of Bayesian learning abou CEO abiliy and esing he model predicions abou he learning process. Taylor (212) uses a model similar o ours o sudy he way in which CEO pay is relaed o he CEO s bargaining power. Taylor s model does conain a predicion abou he relaion beween sock reurn volailiy and he marke s esimaes of CEO qualiy, and he uses his relaion o idenify parameers of his srucural model. In conras, our sudy focuses on esing wheher his and oher predicions in learning 1 In addiion, several sudies apply he learning framework o undersand managerial incenives in he money managemen indusry (e.g., Berk and Green (24), Chung, Sensoy, Sern and Weisbach (212), and Lim, Sensoy, and Weisbach (213)). 5

8 models of managemen abiliy are consisen wih he daa, and evaluaing he exen o which uncerainy abou managemen conribues o sock reurn volailiy. The remainder of our paper proceeds as follows. Secion 2 presens he formal model. Secion 3 describes he daa and he empirical approach. Secion 4 presens evidence of a robus relaion beween sock reurn volailiy and CEO enure, documens ha he learning curve is convex, and also considers he possibiliy ha hese findings could occur because of nonrandom iming of urnovers or subsanial posurnover changes enaced by he new leadership. Secion 5 analyzes he cross-secional deerminans of he learning speed, and he relaion beween marke reacion o corporae news and CEO enure. I also provides esimaes of he imporance of uncerainy abou he managemen relaive o ha of fundamenal uncerainy. Secion 6 discusses he implicaions of he paper s findings and concludes. 2. Uncerainy abou CEO Abiliy and Sock Reurn Volailiy: A Simple Model In his secion we develop a simple model based on Pasor and Veronesi (23, 29) o formalize he link beween uncerainy abou CEO abiliy and sock reurn volailiy. In he model, here is an unknown managerial abiliy ha affecs profis. Over ime, marke paricipans draw inferences abou his abiliy when news arrives abou he firm. When here is uncerainy abou CEO abiliy, news abou he firm has wo effecs on he firm s expeced fuure prospecs. Firs, he news can lead he marke o updae is expecaion abou he firm s fuure profis direcly. Second, he news can also lead he marke o updae is assessmen of he manager s abiliy, and hus indirecly change he expeced fuure profis from he change in he assessmen of abiliy. For example, if here is posiive news abou he firm s cash flows, he marke will value he firm s cash flows a a higher level and consequenly will increase he firm s value. In addiion, he posiive news is likely o reflec well on he managemen, increasing he marke s esimae of his abiliy and furher increasing is expecaion of he firm s fuure cash flows. This indirec effec hrough learning abou managerial qualiy will augmen he direc effec of news on expeced profiabiliy, leading o higher sock reurn volailiy. 6

9 Wha we refer o as abiliy in he model can be hough o exis for a number of reasons, each of which provides a mechanism hrough which a CEO could add value o a paricular firm. Firs, abiliy in he model could refer o raw alen ha will improve performance in any siuaion. Second, abiliy could arise from he qualiy of a mach beween a paricular CEO and firm (Pan, 212). In his case, he mach qualiy could be uncerain even for esablished execuives who have been CEOs in oher companies, or for execuives who have been wih heir curren firms in oher posiions. Third, abiliy could refer o a corporae sraegy ha he CEO is hired o enac. If he success of he sraegy is uncerain, hen marke paricipans will updae heir priors abou he sraegy s profiabiliy exacly as described by he model. We assume ha sock prices are formed based on an efficien marke wih a represenaive agen: P = 1 1+r E(P +1 + D +1 I ), (1) where P is he sock price a ime ; D +1 is he dividend (or equiy earnings) a ime +1; r is he expeced rae of reurn; I denoes he common informaion se of invesors a he end of. Suppose ha he firm s dividend process follows he geomeric Brownian moion: dd D = αd + σdw, (2) where α is he (rue) CEO abiliy ha deermines he average dividend growh rae, σ reflecs he volailiy of he firm s dividend or earnings growh, and dw is a Wiener process. We refer o σ as fundamenal volailiy because i represens he volailiy ha he firm would have absen uncerainy abou he CEO s abiliy. We assume ha α follows a runcaed normal disribuion wih prior mean θ and variance, and α<r wih probabiliy one. While invesors canno direcly observe α, hey coninually updae heir belief abou i according o Bayes rule. A any ime, we have α ~ N(θ, δ 2 ), α < r (3) 2 δ dθ m [ dd θ D d], wih m = 2 (4) σ σ δ δ = 2 2 σ + δ (5) 7

10 Equaions (4) and (5) represen Bayesian updaes of θ and (see, e.g., Pasor and Veronesi, 29). Equaion (4) is an approximaion because α follows a runcaed normal disribuion, and holds exacly only when α is non-runcaed normal. Equaion (4) implies ha he speed of learning abou managerial abiliy is equal o m, which is he raio of uncerainy abou he CEO s abiliy 2 δ o he firm s fundamenal cash flow uncerainy σ 2. Equaion (5) suggess ha uncerainy abou he CEO s abiliy 2 δ decreases over ime due o learning, and 2 δ is convex in. Consequenly, he above equaions imply ha here should be a convex learning curve abou CEO abiliy, in which here is faser updaing abou CEO abiliy in earlier periods han in laer periods during he learning process. In Appendix A, we show ha in his framework, he sock price is given by: r r( τ ) 1 P E e D τ α α, (6) = [ τ d ] = D f ( ) d r α 2 where f (α) is he runcaed normal densiy funcion wih mean θ and variance. This equaion indicaes ha no only he perceived average CEO abiliy ( uncerainy abou i ( 2 δ ) affecs sock valuaion (posiively), bu he ) also does (non-monoonically). However, a more easily esable implicaion of his model concerns he sock reurn volailiy, which is given by: dp log( P / D) vol ) σ [1 + ( ) m ]. (7) P θ Equaion (7) characerizes he way in which marke learning abou CEO abiliy influences he firm s sock reurn volailiy, and implies ha a firm s sock reurn volailiy conains wo componens, he fundamenal volailiy and he volailiy due o he marke updaing is assessmen of he CEO s abiliy (see θ ( δ Appendix A for proof). 2 The erm log( P / D) θ equals he marginal reurn o expeced CEO abiliy: When 2 We can also obain a resul similar o Equaion (7) in a wo-sae coninuous ime Bayesian learning model, in which CEO abiliy is assumed o be high wih probabiliy π and low wih probabiliy (1- π ). Wih his disribuional assumpion, an equaion comparable o Equaion (7) holds exacly. We focus on he case in which abiliy is disribued 8

11 i is posiive, hen a shock o he perceived abiliy ranslaes o greaer movemens in sock prices, and when i equals zero, hen uncerainy abou CEO abiliy will affec neiher firm value nor reurn volailiy. Therefore, he firs empirical implicaion of Equaion (7) is ha when CEO abiliy maers for firm value, hen he firm s sock reurn volailiy should increase wih he amoun of uncerainy abou he CEO s 2 δ abiliy ( ). Second, over ime as he marke learns abou α, should decline, leading sock reurn 2 δ volailiy o decline as well and evenually converge o he level of fundamenal volailiy (vol dp P σ as δ ). 3 Third, since 2 δ is decreasing and convex in, sock reurn volailiy should also be decreasing and convex in. Finally, learning should affec he firm s idiosyncraic risk, bu no is sysemaic risk or expeced rae of reurn (see Pasor and Veronesi (23, 29) for more discussion and proof). Equaions (4), (5), and (7) sugges ha, holding oher facors consan, a more negaive volailiyenure relaion over a CEO s career reflecs a faser learning speed ( m ). The model esablishes a link beween he empirically esimable volailiy-enure relaion and he concep of learning speed formalized by his model. Thus, he model provides a roadmap for inferring he naure of marke learning abou CEO abiliy based on he dynamics of sock reurn volailiy. In summary, his model provides a heoreical link beween marke learning of a firm s CEO abiliy and he dynamics of is sock price movemens. Examining he way sock reurn volailiy changes during he learning process provides us wih esimaes of he exen o which he marke learns abou he CEO s abiliies, he speed of learning, and he facors ha affec his learning process. 3. Empirical Design and Specificaion normally because he poserior variance is characerized by he formula presened in Equaion (5) and is a monoonic funcion of ime, which provides cleaner guidance for he empirical analysis. 3 The model presened here assumes ha CEO abiliy α is consan over ime so ha he uncerainy abou i converges o zero. If CEO abiliy changes over ime, hen he uncerainy abou i converges o a saionary level above zero (e.g., Holmsrom, 1999). In his case, he sock reurn volailiy will always be above he level of fundamenal volailiy. 2 δ 9

12 3.1. CEO Turnover and Sock Reurn Volailiy Evaluaing he model s predicions on he relaion beween uncerainy abou CEO abiliy and sock reurn volailiy is complicaed by he fac ha here is some uncerainy abou he abiliy of every CEO, and for every sock, his uncerainy will conribue somewha o is volailiy. However, he heory presened above suggess ha learning abou CEO abiliy should be mos imporan when uncerainy abou abiliy is highes, presumably when a new CEO akes office. Therefore, if he goal is o measure he way in which he marke learns abou a CEO s abiliy, a naural place o sudy is he period following he succession of a new CEO. For his reason, we consider a sample of CEO urnovers and draw inferences abou he process by which he marke subsequenly learns abou he new CEOs abiliies. Prior o he urnover, uncerainy abou he new managemen is likely o increase because he marke does no necessarily know who will be he new CEO, or even if here will be a new CEO. Afer he urnover, he marke will learn abou he new CEO s abiliy and sraegy for managing he firm, leading i o updae is assessmen o a more precise esimae of he CEO s abiliy α. When α is known more precisely, he impac of new informaion on he marke s esimae of α declines, lowering sock marke volailiy. Thus, assuming ha fundamenal volailiy remains consan, we expec he sock reurn volailiy o rise around he ime of a CEO urnover, and hen o decline over he CEO s enure. The underlying assumpion in his argumen is ha fundamenal volailiy of a sock is unrelaed o he managemen change. I is possible, however, ha CEO urnovers end o occur a imes of high uncerainy abou he firm s fundamenals, and hus fundamenal volailiy ends o be unusually high around urnover. To evaluae his possibiliy and o isolae he effec of learning, we examine he paern around exogenous urnovers ha are likely o be unrelaed o oher sources of uncerainy abou he firm s value. A second reason why fundamenal volailiy could change subsequen o CEO urnovers is ha CEO urnovers are ofen followed by subsanial real changes in he firm. Pos-urnover real changes in he firm s asses could have wo differen effecs on he firm s sock reurn volailiy. Firs, hey can have a direc impac on volailiy since hey change he firm s asse porfolio. Second, hey can serve as signals 1

13 abou he new managemen s qualiy and hus indirecly affec volailiy hrough he learning channel. For his reason, we use daa on pos-urnover real changes in he firm boh o ensure ha any relaion we esimae beween CEO enure and sock reurn volailiy is no spurious because of he endency of real changes o occur following urnover, and also o evaluae he exen ha real changes are signals ha provide informaion abou he CEO. A hird and relaed concern is ha CEO urnover could increase he likelihood of he revelaion of (negaive) informaion abou he firm s fundamenals ha had been wihheld by he previous managemen. The new informaion could accelerae he marke s updae abou he firm s expeced profiabiliy and conribue o he increase in reurn volailiy around CEO urnover. To address his concern, we conrol for informaion disclosure immediaely afer CEO urnover hrough announcemens of accouning wrie-offs, earnings resaemens, securiies fraud invesigaion, divesiures, and erminaion of invesmen. In addiion, we consider a subsample of maure firms for which he uncerainy abou he fundamenals is presumably low, and had exogenous CEO urnovers wihou major pos-urnover acions o confirm he robusness of he learning paern Sample Consrucion We sar wih 24,78 firm-year observaions from 1992 o 26 for which we can idenify he CEOs from he ExecuComp daabase. We use he informaion on job ile, he dae becoming CEO, and he CEO annual flag provided by ExecuComp o idenify CEOs a he firm-year level. For each firm, we compare he designaed CEO in each fiscal year wih he CEO in he previous year o idenify wheher here is a CEO urnover in ha year. For each CEO, he calendar year-monh in which he CEO akes office is designaed as even monh. We exclude urnover evens involving ransiory CEOs such as urnaround specialiss and inerim CEOs (wih enure shorer han 3 years). This process leads o a sample of 1,873 CEO urnovers a 1,582 firms. Panel A of Table 1 describes he disribuion of urnovers over ime. We classify CEOs based on heir succession origin. Using informaion on he ime of a CEO joining company from ExecuComp, supplemened by he daa on saring job from Boardex, we classify CEOs who have been wih he firm for less han wo years when becoming CEO as ousider CEOs, 11

14 and hose who have been wih he firm for a leas wo years as insider CEOs. Based on his classificaion, abou 33% of new CEOs in our sample are considered as ousider CEOs. This fracion is consisen wih hose repored in oher sudies such as Parrino (1997), Murphy and Zabojnik (27), and Cremers and Grinsein (211). Since he purpose of our empirical work is o examine he pos-urnover dynamic of firm sock reurn volailiy, i is imporan o know he reasons for he CEO urnover. Unforunaely, firms are generally secreive abou he rue reasons for CEO changes and usually offer bland, uninformaive reasons when announcing CEO deparures (e.g., he wans o spend more ime wih his family ). 4 I is possible, however, o classify a subse of urnovers as eiher exogenously occurring, or forced. We follow Fee e al. (213) and use he Faciva news search o idenify CEO deparures due o healh issues and deahs. We classify urnovers as reiremen-relaed if he deparing CEO is older han 65. We consider urnovers caused by illness, deah, or reiremen of he deparing CEOs o be exogenous urnovers. We also use he Faciva news search approach o deermine wheher a urnover is overly forced (e.g., forced o leave or lef under pressure). 5 Through his process, we end up wih 211 exogenous urnovers, 56 of which were relaed o healh issues and deahs, and 11 forced urnovers Sock-Reurn Volailiy We rely on hree measures of monhly firm level equiy-based volailiy: Opion-implied Volailiy, Realized Reurn Volailiy, and Idiosyncraic Reurn Volailiy. Opion-implied Volailiy is he monhly average of he implied volailiy calculaed based on he daily prices of he 3-day a-hemoney call opions wrien on he firm s common sock. This measure represens an esimae of volailiy based on he marke s forward-looking assessmen of he firm s risk. Realized Reurn Volailiy is he sandard deviaion of daily sock reurns wihin a monh, based on daa from CRSP. To esimae 4 See Warner, Was and Wruck (1988) or Weisbach (1988) for more deail. Schwarz-Ziv and Weisbach (213) use privae daa on board meeings o documen deails of specific cases in which CEOs are forced ou of heir jobs, bu for which one could never ell so using publicly-available informaion. 5 We hank Edward Fee, Charles Hadlock, and Joshua Pierce for kindly providing us wih heir classificaion of urnovers. 12

15 Idiosyncraic Reurn Volailiy, we follow he mehod in Ang e al. (26) and calculae he monhly volailiy of he residual sock reurn of he following Fama-French hree-facor model. r i i i i i i = α + β MKT + β SMB + β HML + ε. MKT SMB HML Idiosyncraic Reurn Volailiy is defined as Var ε ) from he above equaion. All hree volailiy ( i measures are calculaed for each firm-monh in he hree years afer each CEO urnover in our sample. All hree volailiy measures are aggregaed o he monhly level by muliplying hem wih 21, he square roo of he average rading days in a monh. Panel B of Table 1 repors saisics on he volailiy measures. Boh Realized Reurn Volailiy and Idiosyncraic Reurn Volailiy daa are from 1992 o 29, and Implied Volailiy daa are from 1996 o The average monhly opion implied volailiy is 17%, he average realized monhly volailiy is 12%, and he average monhly idiosyncraic volailiy is 1%. We also repor he summary saisics of he beas on he hree Fama-French facors, which measure he firm s sysemaic risks. The average marke bea in our sample is 1.6, he average SMB bea is.62, and he average HML bea is Oher Variables To conrol for non-managemen relaed facors ha poenially affec volailiy, we also include a se of firm characerisics. Panel C of Table 1 conains summary saisics of hese conrol variables for each firm-year for he hree years afer urnover. The firms in our sample are covered by Execucomp and hus are S&P 15 firms. Abou 55% of hem pay common dividends. The average firm in our sample is abou 22 years old since IPO, has book asses of abou $1.5 billion, 2% leverage (long-erm deb o oal asses), marke equiy o book equiy raio (MB) 2.6, and reurn on equiy (ROE, ne income divided by book equiy) 8%. The volailiy in profiabiliy (VOLP) is esimaed as he annual residual volailiy from an AR(1) model of ROE, and has an average value of 57%. Appendix B provides variable consrucions for he main measures we use in his paper. 6 The daa on opion prices are from OpionMerics and are only available afer 1996, so his measure of volailiy is only available afer

16 4. Measuring he Relaion beween CEO Tenure and Sock Reurn Volailiy The heory discussed in secion 2 implies ha he marke is coninually updaing is assessmen of he CEO s abiliy, as well as he expeced change in fuure profis resuling from any change in is esimaes of his abiliy. Since uncerainy abou managerial qualiy is likely o increase prior o a CEO urnover, and decline as he CEO s qualiy becomes revealed over ime, he model predics ha sock reurn volailiy should increase around CEO urnover and hen decrease over he CEO s enure. Addiionally, his paern should be mainly driven by changes in idiosyncraic reurn volailiy, no in he firm s sysemaic risk. Figure 1 porrays a graphical depicion of he relaion beween monhly average sock reurn volailiy and CEO enure from 12 monhs before CEO urnover o 6 monhs following i. 7 Panel A presens he figure using he opion implied volailiy, while Panel B uses realized volailiy, and Panel C uses idiosyncraic volailiy o measure firm-level volailiy. For each measure, Figure 1 indicaes ha volailiy increases subsanially around he ime of he urnover, and decreases subsequenly. The decrease is paricularly pronounced in he firs hree years of he CEO s enure. Figure 2 illusraes firms sysemaic risk over he same period relaive o CEO urnover. Panel A shows he paern of he marke bea, while Panels B and C use he SMB bea and HML bea respecively. This figure indicaes ha, unlike idiosyncraic risk, he beas of sysemaic facors do no have a clear relaion wih CEO enure. The implicaion is ha changes in sock reurn volailiy around CEO urnover are unlikely o be driven by changes in he firm s sysemaic risk and expeced rae of reurn Esimaing he Volailiy-Tenure Sensiiviy The paerns in hese figures are consisen wih he noion ha uncerainy abou he managemen qualiy and he marke learning of i are refleced in sock reurn volailiy, and paricularly in he idiosyncraic volailiy. However, hey do no conrol for oher poenially relevan facors ha could be relaed o boh CEO urnovers and volailiy. Therefore, we esimae mulivariae models predicing a 7 To consruc he sample for his figure, we require CEOs o have a leas 6 monhs of enure and ha he preurnover sample period (-12, ) of he successor CEO does no overlap wih he pos-urnover sample period (, 6) of he deparing CEO. 14

17 sock s volailiy as a funcion of CEO enure, as well as oher relevan facors. We use a number of alernaive specificaions o characerize his relaion, which can be summarized by he following equaion: Vol ij = f ( Tenure) + ij i α + λ + Conrols + ε ij, ij ij where Vol is one of he hree volailiy measures; α is he firm-ceo fixed effec for he pair of firm i and CEO j; λ is he calendar year-monh fixed effec. The variable Tenure is he number of monhs since he CEO ook office, scaled by 12, so ha he variable akes discree values beween and 3. We focus on he hree years following he urnover, since Figure 1 suggess ha he decrease in volailiy occurs primarily in his period. Since he heory predics ha volailiy should be a convex funcion of CEO enure, we use a specificaion ha allows for a nonlinear relaion beween enure and volailiy, denoed by he funcion f(.). We allow f(.) o be eiher a polynomial funcion or a spline funcion o esimae he degree of convexiy in he learning curve. For he wo oal volailiy measures (Opion-implied Volailiy and Realized Reurn Volailiy), we conrol for he firm s sysemaic risk measured by he monhly beas of he hree Fama-French facors, as well as he se of firm characerisics discussed above. For Idiosyncraic Reurn Volailiy, we do no conrol for he facor beas because he idiosyncraic volailiy is calculaed as he residual volailiy afer neing ou hese facors. Panel A of Table 2 repors esimaes of he relaion beween volailiy and enure based on a polynomial specificaion. Models (1) o (3) esimae his relaion using linear and quadraic erms of Tenure. The heory presened above suggess ha he volailiy-enure relaion should be convex, i.e., he volailiy should decrease a a decreasing rae over CEO enure. In his specificaion, convexiy means ha he coefficien on he linear erm should be negaive and on he quadraic erm should be posiive. The esimaes in Panel A of Table 2 follow exacly his paern, and he resuls are saisically significan and robus across differen volailiy measures. 8 8 All he CEOs in our sample have a leas hree years of enure. Thus, he decrease in sock volailiy is no driven by CEOs in high-volailiy firms being fired quickly. We have also esimaed hese equaions including CEOs wih enure shorer han 36 monhs as well, and he resuls are similar o hose repored in Table 2. For example, he coefficien 15

18 In models (4) o (6), we add a cubic erm of Tenure o evaluae he imporance of higher-order erms. In each of he hree models, he coefficien on he cubic erm iself is no saisically significanly differen from zero, and is sign varies across specificaions. However, he linear erm and he quadraic erm sill have he expeced signs and remain saisically significan. These resuls sugges ha he firs wo erms of Tenure are sufficien o characerize he convex shape of he volailiy-enure relaion. In Panel B of Table 2, we presen resuls using a spline specificaion (Friedman, 1991) wih cuoff poins a Tenure = 1 (firs year), 2 (second year). This specificaion allows us o esimae he learning slope separaely in each of he firs hree years of he CEO s enure. The convexiy of he learning speed m in implies ha sock reurn volailiy should decline faser in earlier periods of he CEO enure han in laer periods. In each of he spline models presened in Panel B of Table 2, we find ha he slope esimae is significanly more negaive in year 1 han in year 2. The absolue value of he esimaed slope coefficien in year 2 is less han half of is value in year 1. The slope esimae in year 3 is less negaive han ha in year 2, alhough he difference across hese wo years is no always saisically significan. In Model (4), we include he firs five years of enure in he spline regression for CEOs wih a leas 7 years in office as a robusness check. Using his specificaion, he slope esimaes for he firs hree years are sill negaive and significan. The slope esimaes for he periods afer year 3 are also negaive, alhough no saisically significan. This paern confirms he inuiion ha learning is mos pronounced when uncerainy abou CEO abiliy is highes. In summary, he resuls in Table 2 imply ha he firm s sock reurn volailiy decreases in he firs hree years of a CEO s enure, wih fases decline in he firs year. These resuls are consisen wih he implicaions of he model presened above. Marke learning abou he CEO s abiliy leads o decreasing uncerainy abou he CEO, which in urn leads o decreasing sock volailiy (paricularly idiosyncraic volailiy) over he CEO s enure. The learning curve appears o be convex, wih faser learning in earlier esimae on Tenure is (p-value<.1) and ha on Tenure 2 is.137 (p-value<.1) using he specificaion of Model (3) of Table 2. 16

19 periods immediaely afer urnover. A signal wih a specified precision helps he updaing of CEO abiliy more in earlier years when uncerainy abou his abiliy is higher Exogenous and Oher Turnovers The high level of sock reurn volailiy around he ime of CEO urnovers and he subsequen decline are consisen wih high uncerainy abou he new CEO s qualiy. However, an alernaive explanaion o learning is ha CEO urnovers end o occur a imes when here is a high level of fundamenal volailiy. Boh he underlying uncerainy abou he firm s prospecs and he uncerainy abou he new CEO s abiliy could poenially lead o heighened sock reurn volailiy around CEO urnover. A long lieraure beginning wih Warner, Was, and Wruck (1988) and Weisbach (1988) documens ha CEO urnovers, and paricularly forced urnovers, are more likely o occur subsequen o poor firm performance, which are also likely o be imes of unusually high sock reurn volailiy. However, his lieraure also documens ha urnovers due o exogenous evens such as illness, deah, and normal reiremens of he deparing CEOs do no occur subsequen o unusual performance (e.g., Weisbach, 1988, Fee e al., 213). Therefore, he exogenous urnovers in our sample should provide a subsample for which i is unlikely ha volailiy will be unusually high for reasons oher han learning. Panel A of Table 3 repors he summary saisics of firm performance and characerisics prior o urnover for he exogenous urnover sample, forced urnover sample, and oher urnovers. Consisen wih he findings in he lieraure, exogenous urnovers in our sample do no end o follow poor performance, while forced urnovers do. Firms wih urnovers classified as exogenous also end o be more maure han oher firms: hey are more likely o be dividend payers and have lower volailiy in profiabiliy. From hese saisics, i seems unlikely ha he urnovers we classify as exogenous end o occur during periods of high firm fundamenal volailiy. In Panel B of Table 3 we separaely esimae he volailiy-enure slope for he subsamples of exogenous urnovers (Model (1)), forced urnovers (Model (2)), and oher urnovers (Model (3)), using Idiosyncraic Reurn Volailiy and he polynomial specificaion of Tenure. The resuls indicae ha subsequen o all ypes of urnovers, here is a negaive and convex volailiy-enure relaion. Even when 17

20 he urnover is due o exogenous reasons, he idiosyncraic volailiy declines wih CEO enure. In hese urnovers, which are unlikely o be caused by prior poor performance and high fundamenal volailiy, he subsequen decline in volailiy likely reflecs he marke s learning abou he abiliy of he new CEO. Panel B of Table 3 also shows ha he volailiy-enure sensiiviy, and hence he learning speed, is significanly lower following exogenous urnovers (-.693) han following forced urnovers (-1.615). This paern is also consisen wih he model, since equaions (4) and (5) imply ha he learning slope ( ) m should be seeper when uncerainy abou he managemen qualiy ( ) is higher. Since mos of he exogenous urnovers end o be associaed wih well-planned successions, here is likely o be relaively less uncerainy abou he new managemen han in he cases of forced urnovers. The fac ha he volailiy-enure relaion is negaive across all ypes of CEO urnovers suggess ha he pos-urnover decline in sock reurn volailiy is no driven by he nonrandom iming of urnovers, and is consisen wih uncerainy abou CEO qualiy decreasing over ime because of marke learning Pos-Turnover Real Changes and Informaion Revelaion CEO urnovers are ofen followed by subsanial policy changes in he firm. These changes ofen reflec he vision of he new leadership o change he firm s sraegies and policies. Pos-urnover real changes could affec he firm s sock reurn volailiy in wo ways: Firs, if hey change he firm s asse porfolio or business policies, hen he firm s fundamenal uncerainy will change as well. Such a change could lead o a shif in he level of volailiy. Second, if such corporae decisions provide signals abou he CEO s abiliy, hen hey can change he speed a which he marke learns abou he CEO. Consequenly, pos-urnover real acions could affec he sensiiviy of volailiy o enure hrough he learning channel. I is also possible ha CEO urnover lead o an increase of addiional informaion abou he firm s fundamenals. Career concerns could moivae incumben managemen o wihhold negaive informaion abou he firm s profiabiliy and o hold ono poorly performing invesmens for oo long. When a new CEO akes over, he has incenives o le he marke know abou he negaive informaion quickly so as o no be held responsible for he poor decisions of his predecessor. For his reason, CEO urnovers can δ 18

21 faciliae informaion revelaion and invesmen re-opimizaion (Kanodia, Bushman, and Dickhau, 1989; Boo, 1992). Consisen wih his argumen, empirical sudies have shown ha subsanial accouning wrieoffs and divesiures are more likely o occur following CEO urnover and he marke seems o be surprised by he new informaion conveyed in hese acions (e.g., see Murphy and Zimmerman (1993) on accouning bah, and Weisbach (1995) and Pan and Wang (212) on downsizing). The addiional informaion revelaion and he corresponding correcive acions could poenially conribue o he increase in reurn volailiy around CEO urnover. To evaluae he effecs of pos-urnover real changes or informaion releases on sock volailiy, we gaher daa on (1) hree ypes of acions ha have real effecs on he firm s asse porfolio: downsizing, expansion, and inroducion of new producs; and (2) hree ypes of acions ha reveal informaion abou he firm s fundamenals: accouning wrie-offs, earnings resaemens, and securiies fraud invesigaions. For each ype of acion excep he revelaion of fraud, we obain informaion abou acion announcemen from he Key Developmens daabase from Capial IQ, which sars in 21. We classify announcemens as downsizing if hey conain announcemens of seeking o sell/dives and disconinued operaion/downsizing. Expansion announcemens are hose conaining seeking acquisiions/invesmen, or business expansion, or M&A ransacion announcemen. New Produc announcemens conain produc-relaed announcemens ha are relaed o new produc releases. Resaemen/Wrieoff/Fraud conains announcemens of resaemen of operaing profis, and impairmens/wrie-offs. We augmen his caegory wih announcemens abou securiies fraud invesigaions, aken from he Federal Securiies Regulaion (FSR) daabase (Karpoff e al., 212), which conains all securiies fraud cases during our sample period. We creae dummy variables ha equal one if here is any one of he above announcemens in a paricular monh. Abou 5% of firm-monh observaions in our sample conain downsizing announcemens, 12% conain expansion announcemens, 13% conain new produc announcemens, and 1% conain resaemen/wrieoff/fraud announcemens. We esimae he effecs of hese changes on he dynamics of volailiy afer urnovers. Since hese acions could eiher affec he level of volailiy, or he volailiy-enure slope, in Table 4 we use Tenure, a 19

22 monhly announcemen indicaor (downsizing in model (1), expansion in model (2), new produc release in model (3), and resaemen/wrieoff/fraud in model (4)), he ineracion of he wo, as well as oher conrol variables o predic he firm s monhly idiosyncraic volailiy. Several resuls are eviden from Table 4. Firs, in all specificaions, he direc effec of Tenure is negaive and significan, wih magniudes comparable o hose in previous specificaions. This resul suggess ha idiosyncraic volailiy decreases wih CEO enure even when here are no significan changes announced. Our main resul on he relaion beween volailiy and enure does no appear o be driven by he higher likelihood of changes in he firm afer CEO urnover. Second, announcemens of pos-urnover changes end o increase firms idiosyncraic volailiy, wih resaemen/wrieoff/fraud having he larges effec. This kind of announcemens usually generaes grea uncerainy abou he firm s fundamenals and fuure prospecs. Finally, he ineracion effec of real changes and enure is negaive, bu is saisically significan a 5% level only for new produc releases. The negaive ineracion effec wih new produc releases suggess ha he inroducion of new producs helps speed up learning abou he managemen qualiy. In models (5)-(8) we repea he analysis using only he exogenous urnovers. We find ha he direc effec of Tenure is sill negaive and saisically significan, wih esimaed magniudes comparable o hose in previous specificaions. These resuls sugges ha idiosyncraic volailiy decreases wih CEO enure even when he urnovers do no coincide wih high firm fundamenal volailiy and when here are no pos-urnover real changes or subsanial releases of informaion. 5. Addiional Implicaions of he Learning Model We have documened ha, holding oher facors consan, a firm s sock reurn volailiy decreases wih CEO enure subsequen o he urnover. This relaion is robus o a variey of alernaive specificaions, and is driven neiher by urnovers occurring a imes of unusually high volailiy nor heir being followed by subsanial changes in he firm. Insead, we argue ha he decline in sock reurn volailiy subsequen o CEO urnovers likely occurs because of he marke s learning of he CEO s qualiy. As he CEO s abiliy 2

23 becomes beer known, signals ha are informaive abou he CEO s abiliy have a smaller impac on firm value, leading o lower sock reurn volailiy. One implicaion of he learning model ha we have already esed is ha he learning curve should be convex, meaning ha volailiy should decrease a a decreasing rae. The learning model also conains a number of oher empirical implicaions. Firs, i predics ha he speed a which he marke learns abou he CEO should be a funcion of he iniial uncerainy of he CEO s abiliy, as well as he qualiy of informaion available o marke paricipans abou his abiliy. Second, he key idea underlying he model is ha he effec of news on firm valuaion should depend on how cerain managemen qualiy is known; consequenly, sock price reacions o news ha poenially reflecs CEO abiliy should be larger in absolue value when CEOs are newer. Finally, he model implies ha uncerainy abou he CEO s abiliy can be an imporan componen of firm-level volailiy: Equaion (4) shows ha he learning speed (m ) equals he raio of he variance of CEO abiliy o he variance of he fundamenals. Therefore, esimaes of he learning speed are also esimaes of he relaive imporance of CEO-relaed uncerainy o fundamenal uncerainy in deermining firm-level volailiy. Equaion (7) provides a way o esimae m based on he volailiy-enure relaion. These esimaes of m also measure he conribuion of uncerainy abou managemen o overall sock reurn volailiy. In his secion we explore hese implicaions empirically Cross-Secional Deerminans of he Learning Speed In he learning model presened above, marke paricipans coninually updae heir assessmen of he CEO s abiliy using Bayes rule. The magniude of hese updaes, which we refer o as he learning speed ( m ), depends on he precision of he marke s prior esimae of he CEO s abiliy, relaive o he qualiy of informaion abou he CEO. Therefore, he model predics ha he learning speed should increase 2 in he amoun of uncerainy abou he CEO s abiliy ( ) and he signal precision ( 1/ σ ). To es hese implicaions, we use a wo-sage procedure. We firs esimae he volailiy-enure slope separaely for each CEO in he sample, using daa from he CEO s firs 36 monhs in office. As discussed in Secion 2, a more negaive volailiy-enure slope corresponds o a faser learning speed. We δ 21

24 hen es wheher facors associaed wih uncerainy abou he CEO s abiliy or he signal precision affec he magniude of he esimaed slopes across CEOs. To esimae CEO-firm-specific volailiy-enure slopes, we rely on he following specificaion: Vol ij ij ij = η + β Tenure + ε, (8) where ij Vol refers o idiosyncraic volailiy under CEO i s enure in firm j, and Tenure is he monh in office coun scaled by 12 (, 1/12, 3). The coefficien β ij capures he average rae of decline in volailiy during he enure of CEO i in firm j. For our purpose here, we refer o ( β ) as he Learning Slope, which should be posiively relaed o he average learning speed. To miigae he noise in he esimaed slope, we normalize i using is empirical cumulaive disribuion funcion, so ha slopes are ranked beween and 1, reflecing he relaive rankings of learning speeds across firms. A learning slope of 1 corresponds o he fases speed. In he second sage regression, we relae hese esimaed learning slopes o firm and CEO characerisics, which according o he model should affec he learning speed. The specificaion for he second-sage cross-secional regression is: LearningSl ope = X ' γ + u ij ij ij The resuls from his wo-sage esimaion process are summarized in Table 5. Panel A groups esimaes of learning slopes from he firs sage by he Fama-French 1 indusry classificaion (see deailed definiions in hp://mba.uck.darmouh.edu/pages/faculy/ken.french/daa_library.hml). The wo indusries for which our esimaes indicae ha learning speeds are he highes are he echnology indusry (compuers, sofware, and elecronic equipmen) and he healhcare indusry (healhcare, medical equipmen, drugs), while he wo indusries wih he lowes learning speeds are he energy and uiliies indusries. The difference beween he esimaed learning slopes beween he op and he boom indusries are saisically significan. We use a number of variables o measure he degree of prior uncerainy abou CEOs abiliies o add value o heir firms. Firs, he exisence of an heir apparen usually indicaes a well-anicipaed 22