Competition among rating agencies and information disclosure

Transcription

1 Competition mong rting gencies nd informtion disclosure Neil A. Doherty y Anstsi Krtshev z Richrd D. Phillips x My 29, 2008 Abstrct The pper nlyzes why rting gency pools di erent credit risks in one credit grde, nd how informtion disclosure depends on the vlue of informtion to the mrket. We build model to nlyze the optiml disclosure policy of monopoly rting gency depending on the vlue of informtion to investors, nd then describe the potentil mrket nd the strtegy of the entrnt. We nd tht entry of symmetric rting gencies results in symmetric rting scles. It justi es why some compnies obtin multiple rtings nd suggests tht similr rtings from di erent gencies my men di erent credit risks. We empiriclly test the qulittive predictions of the model. Stndrd nd Poor s entry to the insurnce mrket tht ws previously covered by monopoly gency, A.M. Best, is used s nturl experiment to study the impct of competition on the informtion content of rtings. Keywords: rting gency, competition, precision nd disclosure of informtion, insurnce. JEL Codes: D8, G22, G28, L, L43 We re gretful to the seminr prticipnts t University of Toulouse, IIOC 2007, EGRIE 2007, LMU Munchen for vluble comments. y Insurnce nd Risk Mngement Deprtment, Whrton School of the University of Pennsylvni, doherty@whrton.upenn.edu z Corresponding uthor. Insurnce nd Risk Mngement Deprtment, Whrton School of the University of Pennsylvni, krt@whrton.upenn.edu x Deprtment of Risk Mngement nd Insurnce, Robinson College of Business, Georgi Stte University. rphillips@gsu.edu.

2 Introduction Following recent filures of rting gencies to provide timely ccurte informtion (Enron, Worldcom, Globl Crossing, AB&B, AT&T, to nme the big ones), the industry hs been criticized for the signi cnt concentrtion of power in smll number of rting gencies. In response to these criticisms, the U.S. House of Representtives pssed the "Credit Rting Agency Reform Act of 2006" tht imed to substntilly simplify the process of obtining NRSRO (Ntionlly Recognized Sttisticl Rting Orgniztion) sttus nd, ultimtely, promote competition. From June 2007 SEC hs dopted rules to implement provisions of the Act. In prticulr, rting gencies who hve issued rtings for 3 yers nd stisfy certin requirements hve the option of registering s NRSROs. This policy hs substntilly diminished the regultory brriers to entry. However, till now there is no greement bout the impct of competition on the qulity of informtion provided by credit rters. The current pper ddresses informtion qulity, speci clly incentives for informtion disclosure. Credit qulity ssessment implies tht rting gency produces n estimte of probbility of defult 2. However, with the exception of KMV T M, mjor rting gencies do not disclose numericl estimte of credit qulity 3. Insted, they ssign letter grdes, nd compnies or issuers with similr credit risk chrcteristics hve the sme grde. Pooling reduces the mount of informtion disclosed to the mrket. The rst objective of our pper is to nlyze the incentives of monopoly rting gency to pool informtion. We nd tht the vlue of informtion to the ultimte users (investors, policyholders, etc.) is crucil in determining the gency s disclosure policy. The second objective is to chrcterize the optiml entry strtegy for new rting gency. We nd tht new rting gency enters from the top by trgeting the compnies of the highest credit qulity. We empiriclly test the qulittive predictions of the model on the entry strtegy using the dt on rtings of creditworthiness of insurnce compnies. The model hs three groups of plyers - sellers (debt issuers, corportions), rting gencies, nd buyers (investors, consumers, policyholders). We ssume tht purchsing rting is voluntry to seller, nd rting gency cnnot chrge fee contingent on the rting ssigned. The optiml rting system of the gency is designed to trde-o the bility of high qulity sellers to signl their qulity by purchsing rting nd the bene ts for the low qulity compnies to be pooled with better compnies. This trde-o determines the rting gency s disclosure policy These requirements re listed in Section 5E of the "Credit Rting Agency Reform Act of 2006". An ppliction must include informtion on credit rting performnce mesurement sttistics, procedures nd methodologies used to ssign rtings, orgniztionl structure, the list of the 20 lrgest issuers nd subscribers tht use services of the pplicnt (on con dentil bsis). An pplicnt must lso submit written certi ction from t lest 0 quli ed institutionl buyers. 2 In generl the output of the credit risk model is multidimensionl. Other importnt outputs re expected loss given defult, rting stbility, etc. 3 KMV LLC o ered Merton structured defult probbilities in mid-990s. It ws cquired by Moody s in

3 to pool compnies into rting ctegories nd the optiml coverge of the mrket. The optiml rting scle derived from the model resembles the intervl disclosure rule employed by the mjor credit gencies. We study how the mrket coverge nd informtion precision depend on the vlue of informtion to buyers. In the model, rting gency decides how credit risk informtion is disclosed to the mrket. Lizzeri (999) estblishes striking result; monopoly rting gency s optiml rting scle is to pool ll compnies in one rting. Surprisingly, ll compnies py to be rted in spite the fct tht de fcto the gency discloses no informtion. The crucil ssumption tht leds to no disclosure result is tht ll prties re risk neutrl. It implies tht no prty vlues the precision of informtion contined in rting. We relx this ssumption nd show tht when precision of informtion hs vlue to end users of rtings (investors, buyers of product, etc.), the no disclosure result no longer holds. Also we study how the mrginl vlue for precision of informtion ects the optiml rting scle. As the vlue of informtion increses rtings become more precise. We use the bsic model to nlyze the entry strtegy of new rting gency to the mrket tht is lredy served by the incumbent. We show tht new gency enters by trgeting high qulity compnies in ech rting ctegory. Interestingly, the entrnt competes by di erentiting its rting scle such tht compnies my be rted di erently by the incumbent nd entrnt. Also the number of rtings ech rm obtins depends on the credit qulity of rted compny. However, there is no congruency between the number of rtings nd the qulity of the compny. High nd low compnies cn be rted by both gencies while the intermedite risk compny obtined only one rting. The theoreticl model yields number of predictions regrding the entry strtegy of the new gency, the types of compnies tht will demnd rting from the new gency, nd the stndrd (i.e., stringency) the entrnt will use to ssign rtings. We test these predictions using dt on the U.S. property-libility insurnce mrket. The insurnce industry provides n idel nturl experiment to test our hypotheses for two resons. Unlike the mrket for bond rtings, there re no regultory brriers to enter the mrket for insurnce rtings Second, until recently, the mrket for insurnce rtings hs lrgely been dominted by single monopoly gency - the A.M. Best Compny. Stndrd & Poor s mde its initil fory into the insurnce rtings mrket in the lte 980 s nd drmticlly incresed the number of rtings it provided to insurers during the 990 s. For exmple, in 992, S&P issued full rting opinions on only 23 property-csulty insurers nd this number incresed to over 340 insurers by the end of the decde. By yer 2004, S&P ws the second lrgest insurnce rting gency nd now rtes lmost 800 compnies representing more thn 90 percent of the industry s ssets. We exmine the strtegies employed s the new entrnt cme into the insurnce rtings business. We employ two methodologies. First, we use hzrd model to estimte one-yer probbility of insolvency using publicly vilble dt for ll U.S. property-libility insurers nd 3

4 then use the results to compre the stndrds tht were necessry for rm to receive similr rtings from both the incumbent nd the entrnt gency. Although the results of the hzrd model re interesting nd useful, the nlysis su ers from two limittions. First, the hzrd model relies completely on publicly vilble informtion to determine the one-yer probbilities of defult. Therefore, we re unble to control for ny privte informtion tht might be lerned through the rtings process itself. Second, compring the probbilities of defult for insurers tht receive rting from both Best s nd from S&P ignores the possibility tht rms will decide strtegiclly whether to request second rting from the new entrnt. This second problem potentilly bises our results. The second empiricl test is designed to control for both limittions of the hzrd model nd we investigte di erences in rting opinions cross the incumbent nd the entrnt using Heckmn-style smple selection methodology. By wy of preview, we nd tht high qulity insurers were ttrcted to receive second rting from S&P nd tht S&P required higher stndrds in order for n insurer to chieve similr rting. Both results re consistent with our theory. The reminder of this section reviews the prior. Section 2 presents the model. Section 3 exmines the strtegy for the monopoly. Section 4 nlyses which mrket segments re pro tble for n entrnt. We present our empiricl nlysis in Section 5 nd the conclusion follows. Relted Literture The pper belongs to the growing literture on incentives of informtion intermediries to mnipulte informtion disclosed to interested prties. Since Akerlof s (970) "lemon mrkets" pper, it is recognized tht informtion intermediries my ply crucil role for mrkets under dverse selection (see Bigliser (993)). Boot, Milbourn nd Schmeits (2005) show tht intermediries cn help to coordinte on desired equilibrium. However, if n intermediry cnnot perfectly ssess the qulity of the good nd/or it hs discretion bout how the results of the ssessment re communicted to buyers, the incentive problems my reduce the mount nd the precision of informtion disclosed to the mrket. The theory we develop in this pper builds on Lizzeri (999) who studies optiml disclosure policies of n intermediry who cn lern perfectly the informtion bout the qulity of the seller nd communicte it to the buyer. Lizzeri shows unique equilibrium in which ll types of sellers py to be rted. However, the intermediry does not disclose ny informtion except tht the seller hs obtined rting. This disclosure policy is equivlent to ssigning unique rting to ll types of debt issuers. The intuition goes s follows. Since the bene t of rting is higher for better types of sellers, the coverge of the mrket is determined by the lowest rted type. To increse the willingness of this type to py, the intermediry pools it will ll better types. Risk neutrlity is essentil for this result. It implies tht the buyer is redy to py the sme price regrdless of whether the qulity is known for sure or is uncertin. In other words, 4

5 the buyer does not vlue the precision of informtion disclosed by the intermediry. We chnge this ssumption nd ssume tht buyers cre bout the qulity of informtion contined in the rting. In this respect our nlysis is relted to the literture on informtion qulity nd mbiguity version (Veronesi (2000), Epstein nd Schneider (2008)). There re other explntions for why n informtion intermediry my mnipulte informtion. Mnipultion cn lso occur due to collusion between the intermediry nd the seller. Strusz (2003) shows tht the thret of collusion mkes honest certi ction nturl monopoly. Peyrche nd Quesd (2005) rgue tht mndtory certi ction mkes intermediries more prone to collusion by incresing prticiption of poor types. When intermediries compete for clients, nd re not certin bout their bility s experts, reputtion concerns my led to misreporting of informtion. Schrfstein nd Stein (2000) nd Ottvini nd Sorensen (2006, 2006b, 2006c) study the impct of reputtion concerns on the reports of nlysts. These ppers consider chep tlk models in which intermediries re concerned with estblishing reputtion of being well informed. In order to signl its bility to provide informtion with high precision, the intermediry bises its privte observtion in fvor of prior belief. Mrino (2006) ddresses similr issue in the context of rting gencies. In spite the fct tht most informtion intermediries function in oligopolistic mrkets, there is little reserch on the impct of competition on the disclosure of informtion. Lizzeri (999) shows tht competition leds to full disclosure nd zero fees for certi ction. In this pper we study the impct of new entry into previously monopolistic mrket for rtings. 2 The model There re three groups of gents: sellers nd buyers of credit sensitive product, nd rting gencies. Sellers hve privte informtion bout their qulity v: Higher v corresponds to higher qulity. Rting gencies nd buyers shre common prior bout the qulity of seller. For simplicity, we ssume tht v is distributed uniformly on [0; ]: There is unit mss of identicl buyers. A buyer purchses t most one unit of good from one seller. The buyer s willingness to py for the good depends on the prior on qulity nd the ccurcy of informtion bout qulity. To model demnd for ccurcy we ssume tht buyers hve men-vrince preferences. Given informtion I vilble to buyers, their vlution of good is equl to u(i) E[vj I] V r[vj I]; where E[vj I] is the expected qulity, nd V r[vj I] is the vrince of qulity. > 0 mesures the mrginl vlue of informtion ccurcy to buyers. Buyers re price tkers, nd u(i) is the price they py for the good. Under the prior distribution, buyers vlution is equl to u 0 = 2 2 : 5

6 If the mrginl vlue of informtion is low, 0 < < 6, the reservtion price u 0 is positive. In this cse, providing new informtion is not essentil for functioning of the mrket. When > 6, buyer does not purchse good unless it hs some dditionl informtion bout seller. When = 0 this model is equivlent to Lizzeri (999). We ssume tht seller cnnot credibly communicte its nncil strength to buyers. A rting gency o ers n evlution service for fee nd cn perfectly observe the type v of seller. We ssume tht the fee is the sme for ll sellers purchsing rting, nd the rting gency cnnot screen compnies by demnding higher fee for more fvorble rting. (Also rted compny does not hve n option to withhold its rting 4.) - discuss it lter The disclosure policy of the gency de nes how the estimtes of qulity re communicted to buyers. One prticulr cse is full disclosure, where the rting gency communictes the observed qulity v. In generl, disclosure policy is mesurble function from the set of signls [0; ] into the set of Borel probbility distributions on rel numbers. The optiml disclosure policy in our model (Proposition 4) is similr to the discrete system of rtings employed by the mjor rting gencies. Under this system n gency prtitions the set of reliztion of v in subintervls, nd discloses tht its estimte of qulity belongs to subintervl. We ssume tht obtining rtings is voluntry to sellers. The decision to be rted is bsed on the cost of rting nd its e ect on the buyer s vlution. The informtion impct of rting depends on the disclosure rule employed by the gency nd on the set of rted types. The expected pyo to seller of type v depends upon the buyers vlution nd is equl to u R (v) t, if it is rted; u N (v); if it is not rted, where u R (v) nd u N (v) re the expected pyo s of type v with nd without rting, respectively. Denote the mss of sellers demnding rting. Then the pyo of the rting gency is equl to V = t: The gme consists of three stges.. Sellers lern their types v. A rting gency designs its disclosure policy, nd sets fee. 2. Sellers observe the disclosure policy of the rting gency nd decide whether to purchse rting. The prticipting sellers re evluted, nd the results re disclosed to buyers ccording to the disclosure policy of the gency. 3. The buyers observe the disclosure policy nd the rting if the seller is rted. They decide whether to purchse the credit sensitive product. Sellers receive pyo which depends on rting sttus. 4 Fure-Grimud, Peyrche nd Quesd (2005) show tht rms my hve incentives to hide their rtings only if they re su ciently uncertin bout their qulity. In our setting rms hve perfect informtion bout their qulity, nd thus will not pply for rting unless it increses their reservtion price. 6

7 We study sequentil equilibri of the gme. Strtegies of ll plyers must be optiml t every stge of the gme given the beliefs bout other plyers informtion. Beliefs must be consistent with the Byes rule whenever possible. 3 Monopoly Rting Agency 3. Preliminry results: Full disclosure This section describes the demnd for rtings nd the pro ts of the rting gency under full disclosure. Though this system is not optiml for the rting gency, the nlysis cn be useful to highlight the gency s gins from pooling di erent risk types in the sme rting grde. Proposition Suppose tht monopoly rting gency commits to full disclosure, nd the fee for the rting services is such tht t < Then the unique sequentil equilibrium of the subgme hs threshold structure: There is type v F 2 [0; ] such tht ll types bove v F purchse rting, nd no type below v F is rted. Proof. Let s consider some type v 2 [0; ] nd ssume tht ll types bove v nd no types below v re rted. Under full disclosure rted seller of type v is pid by buyers u R (v) v: If seller type v is not rted, it is pooled with types [0; v]. The reservtion price of non-rted sellers is then equl to u N (v) 2 v 2 v2 ; where 2v is the expected qulity nd 2 v2 is the vrince of qulity of non-rted types [0; v]. If this price is negtive, the non-rted sellers do not trde. A necessry condition tht seller type v purchses rting is tht it increses its reservtion price net of the rting fee, u R (v) t mxfu N (v); 0g: () Note tht s v increses, the di erence between the two prices increses, nd d dv (u R(v) u N (v)) = 2 + v > 0; (2) 6 0 if v = 0; u R (v) u N (v) = > 0 if v = : If the fee t is below 2 + 2, then there is seller type v F 2 (0; ) for which the prticiption constrint () is binding. (2) implies tht ll types bove v F strictly prefer to obtin rting, nd no type below v F obtins rting. The buyers beliefs tht the compny s qulity is bove 7

8 v F if it is rted nd is below v F if it is not rted, re consistent with the equilibrium. The proof of the uniqueness is shown in the Appendix. Under full disclosure, the fee chrged by the rting gency is t = u R (v F ) mxfu N (v F ); 0g. It is equl to the mount tht the lowest rted seller is willing to py for rting. The demnd for rtings is F = v F. So the pro t of the rting gency is mx v F ( v F )(u R (v F ) mxfu N (v F ); 0g): Since incresing the fee reduces the demnd for rtings, the optiml fee for the rting gency nd the resulting coverge of the mrket, F = v F, re derived from the trde o between the mrginl bene t of chrging higher fee nd the mrginl cost of the reduced demnd for rtings. Proposition 2 Under full disclosure, the optiml mrket coverge, 8 p 2(+3) >< ; 0; F = 6 ; 0 2; >: 2 ; 2; is decresing in the vlue of informtion for < 0 nd incresing for 0 < < 2, nd is independent of for > 2: The pro t of the rting gency is 8 (+2)(+3)( 6)+(6+ >< 2 +36) 3 2 ; 0; 62 F = 2 6( 6) ; 0 2; >: 2 4 ; 2: Pro t is incresing in the mrginl vlue of informtion. Di erent regimes for the optiml coverge re driven by the bility of non-rted sellers to trde. When the mrginl vlue of informtion is reltively low, obtining rting is not essentil for trde. However, rted sellers cn chrge v insted of selling t the verge vlution of non-rted types, u N. The fee chrged by the rting gency is equl to the gin of rting to the lowest rted type, nd it increses s the gp between v nd u N becomes higher. Due to this e ect the rting gency bene ts from reducing the mrket coverge. As the mrginl vlue of informtion increses, it becomes impossible to trde without rting. For intermedite vlues, 0 < < 2, non-rted compnies hve zero vlution, nd the e ect on the mrket coverge is reversed. Mrket coverge increses s informtion becomes more vluble. For high vlues of informtion, > 2, non-rted sellers cnnot trde, u N < 0. The fee chrged by the rting gency is the vlution of the lowest rted type, v F. The optiml mrket coverge does not depend on the vlue of informtion, nd it is derived from trde o between the gin of chrging higher fee to better types nd the cost of reduced mrket coverge. 8

9 Is full disclosure optiml? Suppose tht, insted of reporting the type v F ; the rting gency nnounces tht this type is from n intervl [v F ; v F + ], > 0: Thus v F is pooled with better types nd the rting gency my be ble to chrge higher fee without reducing demnd for rtings. When the vlution of pooled types is higher thn the vlution of the lowest rted type, then v F > v F : If the mrginl vlue of informtion is zero, = 0, one obtins Lizzeri s (999) result tht ll types should be pooled nd ssigned the sme rting grde. When precision of informtion mtters, > 0, pooling imposes cost in lost ccurcy. This intuition suggests tht the optiml disclosure policy trdes o the bene ts of pooling due to higher fees nd the cost of pooling due to reduced informtiveness of rtings. 3.2 Optiml disclosure of monopoly rting gency In this section, we nlyze pro t mximizing rting system of monopoly rting gency. A disclosure policy is signl ssigned to ech rted type, s : [0; ]! S, where S is the signl spce chosen by the rting gency. A signl s(v) obtined by type v results in the expected vlution (s(v)) nd the vrince 2 (s(v)), where (s(v)) = E[v 0 : s(v 0 ) = s(v)]; 2 (s(v)) = V r[v 0 : s(v 0 ) = s(v)]: So rted seller type v chieves buyers vlution equl to (s(v)) 2 (s(v)): To nlyze the optiml disclosure policy of n gency, we proceed in two steps. In the next proposition we estblish two properties of n optiml disclosure policy. Then we use these properties to chrcterize the policy. Proposition 3 An optiml disclosure policy of monopoly rting gency stis ed the following properties.. Demnd for rtings. If seller type v 0 purchses rting, ll types v > v 0 lso purchse rtings. 2. Monotonicity of rtings. Higher types re ssigned rtings tht result in higher sellers vlutions. Tht is, for ny v 0, v 00 such tht v 0 < v 00, it holds (s (v 0 )) (s (v 0 )) (s (v 00 )) (s (v 00 )): 9

10 We proceed in three steps. First we ssume tht rting gency o ers disclosure policy such tht types v 2 [v M ; ], v M 0 demnd rtings. Under this disclosure policy it pools types v 2 [v M ; v M + b M ], where b M 0 nd v M + b M ; in one rting grde. We derive the optiml functions v M () nd b M (). Second, we show tht this disclosure policy is optiml for the rting gency. Tht is, there is no other disclosure policy tht obtins higher pro t. Third, we discuss how the rting gency discloses informtion of higher types, v 2 [v M + b M ; ]: For these types intervl disclosure is n equilibrium but it is not unique. Also, s the vlue of precision of informtion increses, the optiml system converges to full disclosure. Denote v M s the lowest seller type tht purchses rting from monopoly rting gency nd ssume tht (i) the gency pools types v 2 [v M ; v M +b M ] in one rting, b M > 0, v M +b M ; (ii) types [v M ; ] purchse rting, (iii) the return from rting is incresing in type v. The pyo of the lowest rted types v 2 [v M ; v M + b M ] is The pyo of non-rted sellers is u(l) = v M + 2 b M u(n) = mx( 2 v M 2 b2 M; (3) 2 v2 M; 0): Under ssumptions (i) (iii) the rting gency fces demnd M = v M, nd erns pro ts ( v M )t M ; where t M is the uniform fee chrged for the rting. These sellers purchse rting only if it hs positive return, u(r; v) t M mxfu(n); 0g for v 2 [v M ; ]; (4) where u(r; v) is the vlution of type v rted R : Assumption (iii) implies tht (4) reduces to u(l) t M mxfu(n); 0g; nd the fee chrged by the rting gency is determined by the willingness of pooled types v 2 [v M ; v M + b M ] to py for the rting, t M = u(l) mxfu(n); 0g: If seller cnnot trde without rting, u(n) < 0, the fee is equl to the vlution when buyer s informtion is tht sellers type is v 2 [v M ; v M + b M ]: When u N > 0, the fee equls to the di erence between the vlutions of pooled compnies [v M ; v M + b M ] nd non-rted compnies [0; v M ]: In equilibrium, non-rted sellers v 2 [0; v M ] must be better o without rting. If seller v 2 [0; v M ] devites nd purchses rting, the rting gency nnounces tht seller s qulity is 0

11 from the intervl [0; v M ]: Then the devition is not pro tble nd purchsing rting cnnot increse the reservtion price chrged by these sellers. Under our ssumptions, n optiml disclosure policy of the rting gency solves mx ( v M)(u(L) mxfu(n); 0g): (v M ;b M ) In the next proposition we summrize the solution to this problem. Proposition 4 Under ssumptions (i)-(iii) the optiml monopoly rting system is summrized in the following tble. v M b M t M M mxfu N ; 0g (+6) 2 3(0 )( 2) (+3) 2 (+3) 3 (2 3)(2 2) ( 6) (4+3) When the vlue of informtion is reltively low, 2, ll seller types re rted nd re pooled in the sme rting grde. As the vlue of informtion increses, the rting becomes more precise, i.e. b M is decresing. The coverge of the mrket is decresing in when the non-rted sellers cn trde, u(n) > 0, incresing in when u(n) = 0, nd decreses in when u(n) < 0: The pro t of the rting gency is non-monotone in the vlue of informtion. It is decresing when u(n) > 0, incresing when u(n) = 0 nd decresing when u(n) < 0: Pro t is the highest when the vlue of informtion is the lowest, = 0: The mrginl vlue of informtion ects the design of the optiml rting system. When is low, 0 2; the optiml disclosure policy of the rting gency is to pool ll sellers in the sme rting grde. Recll tht the vlution by buyers is composed of two components, the expected qulity of the seller nd the precision of informtion bout qulity. Since the vlue informtion precision is low, rting gency cn increse the expected qulity of the lowest rted seller by pooling it with the highest types. For moderte informtion vlues, 2 6, monopoly rting gency hs prtil coverge of the mrket, v M > 0; but ll rted sellers re still pooled in the sme rting grde. Reducing the coverge of the mrket is bene cil for the rting gency becuse it widens the di erence between the vlution of rted nd non-rted compnies, nd llows the gency to chrge higher fee for the rting. At the sme time, the vlue of precision is too low to bene t from incresing precision, so ll rted sellers re pooled in order to increse the expected qulity of the lowest grded type. As the vlue of informtion increses, 6; providing precision becomes more vluble thn incresing expected qulity by pooling. As result, the price tht cn be chrged when subset

12 [v M ; v M + b M ]; v M + b M < ; is pooled in one rting grde is higher thn the price tht cn be chrged when ll rted sellers [v M ; ] re pooled in one rting, even though the lter hs higher expected qulity. The distinction between the lst three regions for 6 is the bility of non-rted sellers to trde. Higher demnds for precision imply tht rting becomes essentil for trde, nd the gency expnds the coverge of the mrket for : However, when the pyo of nonrted sellers is negtive, u(n) < 0, precision becomes secondry to improving the pool of rted compnies. The coverge is incresing for 5 4 : How the rting gency decides the precision of u(l)? From (3), pooling db M sellers in one rting increses the expected qulity of u(l) by 2 db M nd reduces the precision of the rting by ( 6 b M)db M : For low vlues of the impct on precision is dominted by the increse in expected qulity from pooling, nd tht leds to extensive pooling. Note tht the cost of precision is zero when = 0: For higher vlues of ; the interior solution obtins when the mrginl increse in qulity is equl to the mrginl cost of reduced precision, resulting in b M = 3 : As the vlue of precision increses, the mesure of types pooled in one rting goes to zero nd rting L e ectively mens tht the type is v M. The pro t of the rting gency is non-monotone in the vlue of informtion. For reltively low vlues, the rting gency cn bene t from its unique bility to screen sellers nd selectively disclose the results. However, s the vlue of informtion increses, the optiml rting system requires ner informtion disclosure, nd reduces the bility of the gency to increse the fee by pooling types in one rting. Figure shows the boundries for rting L s function of. Types locted below the lower curve re not rted. Types locted between the lower nd the upper curves re pooled in the sme rting grde L. Like under full disclosure, the coverge of the mrket is non-monotone in nd depends on the bility of non-rted compnies to trde. It is decresing in for low informtion vlues becuse the rting gency hs incentive to widen the gp between vlutions of rted nd non-rted sellers. Figure 2 shows the pro t of the rting gency s function of (upper curve): The t line is the limit of pro ts s! + : 2

13 v Figure : Optiml rting scle of monopoly rting gency v Figure 2: Pro t function of monopoly rting gency In the next Corollry we compre the rting scle with two rtings to the full disclosure rting scle. Corollry An optiml rting system derived in Proposition 4 provides higher pro t to the rting gency thn full disclosure rting system. The optiml mount of pooling b M de nes unique boundries for the rting of the lowest group of seller types. However, s the mrginl vlue of informtion increses, nd the rting 3

14 gency does not pool ll types in one rting grde, there re multiple disclosure policies for seller types v > v M + b M. Rting gency s only constrint for these types is tht higher types obtin rting tht results in higher vlution by the buyers. As result, ny sequence of non-overlpping intervls tht results in sequence of incresing vlutions will be optiml. 4 Entry Strtegy of New Rting Agency In this section, we consider the entry of new gency on the following time line. After the rtings hs been ssigned by the incumbent, but before the trnsction between buyers nd sellers, new gency o ers n dditionl rting for fee. If new rting gency ttrcts ny sellers, they re rted by the entrnt. Then buyers form their vlutions bsed on ll vilble sellers rtings (i.e. from the incumbent nd the entrnt), nd trde tkes plce. In this setup, the incumbent hs no possibility to djust its disclosure policy. Our motivtion for this ssumption is tht sellers, buyers nd the incumbent rting gency exhibit inerti in designing nd understnding rting stndrds, nd the industry structure cnnot chnge overnight. The section nlyzes how new rting gency should structure its disclosure policy to crete demnd for its services. A seller will py for n dditionl rting only if it increses its vlue on the eyes of the buyer. This occurs either when the second rting llows the seller to signl higher qulity, or when it improves the ccurcy of informtion of the buyers. If seller is rted by the entrnt, it must be tht u(r m ; R e ; v) t m t e u(r m ; v) t m ; u(r m ; R e ; v) t m t e 0; where u(r m ; R e ; v) nd u(r m ; v) respectively re the pyo s of seller v rted by both gencies nd rted only by the incumbent, respectively, nd t m nd t e re the fees for rtings by two rting gencies. In the next proposition we show tht the demnd for the second rting comes from the sellers with the highest nncil strength. Proposition 5 An entrnt cn lwys design rting system tht ttrcts the best compnies within ech rting intervl of the incumbent. The rting stndrd of the entrnt is more stringent thn tht of the incumbent. A limit exmple of such rting system is the one where the entrnt ttrcts only the highest type v = : This system unmbiguously cretes positive surplus for the seller, nd ll dditionl surplus cn be extrcted by the new rting gency. 4

15 The entrnt s rting is demnded by better thn verge sellers in ech rting ctegory. It implies tht seller with two rtings, from the entrnt nd the incumbent, is not necessrily better thn seller rted only by the incumbent. In generl, the entry strtegy of new gency depends on the rting scheme of the incumbent, nd ultimtely, on the mrginl vlue of informtion to sellers. We restrict ttention to the vlues of tht result in one or two rtings ssigned by the incumbent, < 2: Then the entrnt cn trget t most three groups - sellers with high rting, sellers with low rting, nd sellers with no rting by the incumbent. Denote x; y nd z the lowest types rted by the entrnt in ech of these ctegories, with v M + b M x ; v M y v M + b M nd 0 z v M. When 6 nd ll types re pooled to the sme rting grde by the incumbent, x is set to zero. For ll rted types the entrnt sets fee t e. An optiml entry strtegy of new rting gency solves mx (x;y;z;t e) (( x) + (v M + b M y) + (v M z))t e u(r; R e ; v) t e t u(r; v) t; u(r; R e ; v) t e t 0: The totl demnd for the services of the entrnt is the sum of coverge in ech rting ctegory. The rst constrint gurntees tht seller type v bene ts from obtining second rting. The second constrint is prticiption constrint. The entrnt hs leewy to decide which ctegories of sellers to trget. Like for monopolist, incresing coverge reduces the fee becuse the rting of the entrnt cretes more surplus for better types. The di erence is tht the potentil mrket of the entrnt consists of discrete segments in ech rting ctegory of the incumbent. As result, the entrnt hs two decisions to mke. First it selects the segments of the mrket to be rted; second, it speci es the stringency of the rting in ech trgeted segment. There is no closed form solution for this problem for > 2, but the problem cn be solved numericlly. The solution is summrized in Figure 3. 5 Empiricl Anlysis Our theoreticl model yields severl empiricl predictions regrding the optiml strtegies of pro t-mximizing entrnt. We seek to test these predictions tking dvntge of dt on the U.S. property-libility insurnce industry during the yers The insurnce industry during the time period of the lte 980 s nd through the decde of the 990 s is uniquely suited to test our hypotheses s Stndrd & Poor s, the well-respected bond rting gency, invested signi cnt resources to expnd their in uence nd entered the mrket for insurnce rtings. Prior to this time period, the mrket for insurnce rtings ws lrgely dominted by the A.M. Best Compny. Incorported in 899, A.M. Best hs published rtings on virtully ll U.S. 5

16 .2 Figure : Optiml entry strtegy of the entrnt x; y nd z insurers nd, for mjority of their history, they were the only gency doing so. The monopoly position Best s enjoyed, however, begn to erode fter Best s ws criticized following the libility insurnce crisis of the mid 980 s nd fter severl nturl ctstrophes in the erly 990 s tht bnkrupted numerous insurers. The most ggressive gency to enter the mrket ws Stndrd & Poor s (S&P) who begn publishing rtings on property-libility insurers in 983 nd then expnded coverge once in 987 nd then gin in 99 (Stndrd & Poor s 987; A.M. Best 992). Tody, S&P provides rtings on insurers tht represent in excess of 80 percent of the ssets of the industry - more thn ny other new entrnt except Weiss Reserch Hypothesis Development We test four distinct hypotheses tht re derived either directly from the theoreticl model presented in this pper or from the prior literture. The rst hypothesis comes from Proposition 5 where we predict Hypothesis The new entrnt gency will nd the gretest demnd for its services from the high qulity insurers seeking to di erentite themselves from other insurers tht hve rting 5 Like A.M Best, Weiss Reserch provides rtings on lmost every insurer tht opertes in the U.S. mrketplce. However, the process Weiss uses in the ssignment of their rtings is fundmentlly di erent thn the process used by Bests nd S&P. We consider S&P to be the more in uentil new entrnt into the mrket for property-libility insurnce rtings given their estblished reputtion in the bond rting mrket. 6

17 similr to their own. More speci clly, we predict tht higher-thn -verge qulity insurers within rting clss who re bundled together with lower-thn-verge insurers in tht sme clss will self-select nd seek to di erentite themselves by obtining new rting. 6 In relted hypothesis, we predict Hypothesis 2 The new entrnt rting gency will require higher stndrds, on verge, in order for rm to receive rting similr to the one they received from the incumbent gency. Thus rms tht seek rting from the new gency my not receive higher rtings but insted, for ech rting clss, the newly rted insurers should hve higher verge nncil qulity. The next hypothesis is relted both to the mount of informtion tht is vilble to mrket prticipnts regrding the nncil qulity of the insurer nd how vluble tht informtion is to mrket prticipnts. First, consistent with Rmkrishnn nd Thkor (984) nd Millon nd Thkor (985), we expect Hypothesis 3 More opque insurers or insurers for which mrket prticipnts hve more di cult time ssessing the true nncil strength of the rm will be more likely to seek n dditionl rting. 5.2 Methodology We conduct severl tests using two di erent econometric methodologies to investigte the hypotheses stted bove. In the rst set of tests we seek to empiriclly compre the stringency of the rtings ssigned by the incumbent rm (A.M. Best) reltive to the entrnt (S&P). More stringent rtings stndrds re sid to exist when the verge/medin probbility of defult for insurers in prticulr rting clss is lower for one gency thn the other. The tests re designed to nlyze the following questions. How do the rtings ssigned by S&P compre to the rtings of A.M. Best for rms tht re jointly rted by both compnies? Wht is the verge nncil qulity of the insurers in ech rting ctegory cross the two gencies? Answering these questions requires us to develop summry sttistic of the nncil qulity of the insurers nd then to use tht sttistic to compre the rtings systems of the two gencies. The benchmrk we use is the one-yer probbility of defult for ech rm in our dt set. We rgue the one-yer probbility of defult is resonble benchmrk since both gencies stte the primry objective of their rting systems is to provide n opinion bout the insurer s bility to meet its contrctul obligtions to policyholders. We use these probbilities to exmine the stringency of the rting system by compring either the medin or men probbility of defult for given rting clss. A shortcoming of nlyzing the one-yer defult probbilities using the hzrd model is 6 Our s is not the rst to investigte this hypothesis s Cntor nd Pcker (997) empiriclly investigte similr hypothesis using dt on rms tht obtin rtings from gencies other thn the dominnt bond rting rms Moody s nd S&P. Unlike Cntor nd Pcker, however, we nd strong evidence of self-selection using dt from the insurnce mrket. 7

18 tht the only dt vilble for us to estimte the model is publicly vilble informtion. By de nition, rting gencies should only exist if their preferentil ccess to privte informtion llows them to better di erentite the qulity of insurers. Thus, our second empiricl test investigtes the determinnts of di erences in the rtings tht re ssigned by the incumbent versus the new entrnt rting gency while lso controlling for the privte informtion the gencies lern through the rting process. The methodology we employ lrgely follows the work of Cntor nd Pcker (997) s we use Heckmn-style smple selection model to investigte di erences in the rtings when n insurer chooses to be rted by both gencies. This second methodology llows us to control for privte informtion s we include vribles to proxy for the rting ssigned the A.M. Best gency in the rst stge probit regression. In ddition, we dd to the literture since, unlike Cntor nd Pcker, the theory developed in this pper provides strong guidnce for the control vribles tht we should use to explin di erences in rting ssigned by the two gencies. 5.3 Estimting Defult Probbilities A vriety of methods cn be used to forecst the likelihood of bnkruptcy for n insurnce compny. U.S. regultory uthorities use three univrite models to forecst bnkruptcy. The Insurnce Regultory Informtion System (IRIS), the oldest system, utilizes series of twelve udit rtios bsed upon nncil sttement dt led with the regultors. The newer Finncil Anlysis nd Surveillnce Trcking (FAST) system uses n expnded set of udit rtions, pproximtely thirty, where ech rtio is given corresponding score. Regultors multiply ech individul rtio by its corresponding score nd then sum over ll rtios to produce FAST score. Insurers with higher FAST scores re more likely to become nncilly distressed nd re subject to greter regultory scrutiny. Finlly the risk-bsed cpitl system de nes minimum mount of cpitl insurers must hold. The individul cpitl chrges depend on the riskiness of the ssets nd the businesses in which the insurer prticiptes. In ddition to the univrite regultory models discussed bove, economists hve developed nd implemented vriety of solvency prediction models bsed upon multivrite sttisticl techniques. Insolvency forecsting models bsed upon multiple discriminnt nlysis (Trieschmnn nd Pinches (973)), or logistic regression (Cummins, Grce nd Klein (999)) re common in the literture. In ddition, bnkruptcy prediction models bsed upon neurl networks (Brockett et l., 994) nd dynmic csh ow simultion models (Cummins, Grce nd Phillips 999) hve lso been discussed. The limittion of these methods is tht they re bsed on sttic models implemented using dt tht spns only one or just few yers. As result, these sttic models re indequte for the long-term pnel dt tht we ssembled for the study 7. 7 In ddition, Theodossiou (993) suggests tht rbitrrily choosing when to observe ech rm s chrcteristics leds to unnecessry selection bis problems nd reduced forecsting bility. 8

19 In this study we use the discrete-time hzrd model suggested by Shumwy (200) to overcome the bises of the sttic models nd to tke dvntge of our pnel dt. The hzrd model pproch hs t lest two primry dvntges over the more trditionl sttic models. First, hzrd models llow for time-vrying covrites tht explicitly recognize tht the nncil helth of some rms will deteriorte over time even though the rm my not declre bnkruptcy for mny yers. Sttic models only mke comprisons between rms tht re clssi ed s helthy or not helthy t just one point in time nd they therefore ignore rms tht re t risk of bnkruptcy even though they hve not yet become bnkrupt. Shumwy (200) shows tht ignoring this informtion cretes selection bis which leds to inconsistent prmeter estimtes. Intuitively, hzrd models correct this problem by llowing to extrct useful informtion from the times series dt on ech individul rm. In ddition, it cn be shown tht the prmeter estimtes from hzrd models re unbised nd consistent. The second reson the hzrd model is preferred to sttic models is becuse doing so llows us exploit ll vilble informtion bout the condition of the rm rther thn just the lst yer s observtions. Thus, the incresed mount of dt increses the e ciency of the model which yields more relible prmeter estimtes nd better out-of-smple forecsting results. Implementing the discrete-time hzrd model is rther strightforwrd. Shumwy (200) shows tht the likelihood function of discrete time hzrd model is identicl to the likelihood function for multiperiod logit model. Thus, estimting the hzrd model is equivlent to estimting the trditionl sttic logistic model except the coding of the dependent vrible is slightly di erent. Speci clly, the dependent vrible for the hzrd model, y it, is binry indictor set equl to if rm i is declred bnkrupt in yer t + nd equls 0 otherwise. In other words, the dependent vrible equls 0 for ech yer the rm does not exit the system nd ech bnkrupt rm contributes only one filure observtion, i.e., y it =, in the lst yer the rm hs dt. Time vrying covrites re esily incorported by using ech rm s nnul dt: 5.3. Dt The dt to estimte the hzrd model comes from the nnul regultory sttements of ll property-libility insurers mintined in electronic form by the Ntionl Assocition of Insurnce Commissioners (NAIC). We include ll rms tht meet our dt requirements (discussed below) over the yers Consistent with the literture, we de ne the yer of insolvency s the yer tht the rst forml regultory ction is tken ginst troubled insurer. We identify the yer of rst regultory ginst insurers through vriety of sources including the NAIC s Report on Receiverships (vrious yers) nd the Sttus of Single-Stte nd Multi-Stte Insolvencies (vrious yers). We lso obtin the list of insolvent insurers provided in report by A.M. Best Compny which lists ll property-libility insurers tht filed from (A.M. Best, 2002). From these sources we identify 300 property-libility insurers tht fil between 990 nd

20 The explntory vribles we use to estimte the hzrd model re the nineteen blnce sheet nd income sttement rtios tht mke up the NAIC s FAST solvency trcking system. Grce, Hrrington nd Klein (995) conclude tht there re diminishing mrginl returns to incorporting dditionl blnce sheet nd income sttement rtios not lredy included in the FAST system. Thus, the FAST system seems to cpture s much predictive power s cn be glened from nncil sttement rtios lone. We lso include controls for rm size equl to the nturl logrithm of the rel ssets of the rm where the price de tor we use is the Consumer Price Index; nd n orgniztion form control vrible which is n indictor set equl to if the insurer belongs to mutul or reciprocl group of insurers. As discussed bove, we estimte the hzrd model using ll insurers for which we hve dt to clculte the FAST rtios. Thus, not only do we include insurers rted by A.M. Best nd/or S&P, but we lso include insurer rm-yer observtions tht do not receive rtings from either of these two gencies. The only insurers we delete from the nlysis re those with insu cient dt needed to clculte the nineteen FAST vribles or those who do not hve dt vilble in the yer prior to their rst event yer. In n e ort to include s mny insolvent observtions in the nlysis, we include insurers who report dt two yers prior to their rst event yer but who do not report in the yer prior to their rst event yer. We delete ny bnkrupt rms for which we were unble to locte dt within 2 yers of their rst event yer. The nl dt set contins 24,062 solvent rm-yer observtions nd 24 insolvent rm-yer observtions Empiricl Results Summry sttistics for the solvent nd insolvent compny observtions re shown in Tble. Not surprisingly, tests between the mens of the solvent nd insolvent smples suggest the two groups of insurers di er signi cntly cross number of dimensions. Insolvent insurers crry signi cntly higher leverge rtios (the Kenney Rtio nd the reserves to policyholder surplus rtio) thn do solvent insurers. 8 Insolvent insurers re signi cntly smller in terms of sset size thn re solvent insurers nd re less likely to be members of mutul. Insolvent insurers py out signi cntly more csh reltive to premiums collected thn do solvent insurers nd they much more relint on reinsurnce (see the surplus id to policyholder surplus rtio). 9 The results of the discrete-time hzrd model re shown in Pnel A of Tble 2. Overll the explntory power of the model is resonble s the pseudo R 2 sttistic is 26 percent. The results re consistent with mny of the inferences tht were discussed fter reviewing the summry sttistics shown in Tble. For exmple, the estimted coe cients suggest highly levered rms, rpidly growing rms, nd rms tht rely more hevily upon reinsurnce to support their 8 The Kenney Rtio equls the net premiums written by the insurer divided by the insurer s policyholder surplus. Policyholder surplus is the trditionl nme used in the insurnce industry to represent the equity cpitl of the insurer under sttutory (i.e., not GAAP) ccounting rules. 9 Surplus id is sttutory ccounting item which equls the increse in the mount the book vlue equity cpitl of the insurer due to the purchse of reinsurnce. 20

21 cpitl positions re ssocited with higher filure rtes. Lrger rms nd insurers tht re prt of mutul orgniztions re reltively less likely to defult. Finlly, rms tht hve high csh out ows reltive to in ows nd who experience dverse reserve development re more likely to fil. Pnel B of Tble 2 shows summry sttistics of the estimted one-yer probbilities of defult for solvent nd insolvent insurers. The verge/medin probbility of defult for the helthy rms is 0.8/0.2 percent while the verge/medin sttistics for the rms in the yer before they become bnkrupt is 9.4/4.5 percent. Thus, the verge estimted one-yer probbility of defult for rms tht become bnkrupt in the next yer is over 0 times lrger thn the verge probbility for helthy rms. Clerly the model does resonble job ssigning high defult probbilities to rms tht ultimtely fil nd low probbilities to helthy rms. We lso note here A.M. Best reports the verge nnul probbility of defult for property-libility insurers from ws 0.95 percent - result very consistent with the probbilities produced by our model (A.M. Best, 2004). With the one-yer probbilities of defult estimted, we cn now begin to investigte our hypotheses. However, before we do so we rst need to de ne mpping between the di erent rting ctegories used by the two gencies. Unfortuntely single one-to-one mpping between two systems does not exist nd prior reserch investigting insurnce rtings cross gencies hve used di erent de nitions. 0 For this study we reviewed the verbl descriptions ech gency scribes to their individul rting clsses nd decided to use the ve rting ctegories shown in Tble 3. Numericl vlues, lso shown in the tble, were ssigned to ech rting ctegory to fcilitte comprisons cross gencies nd over time. Tble 4 shows the extent of the coverge ech gency provided of the property-libility insurnce industry over the time of this study. The totl number of insurnce compnies in the NAIC dt bse rnged from low of 897 rms in yer 990 to high of 200 rms in yer 996. The totl ssets of the industry grew from $534 billion in 989 to lmost $940 billion by the end of Of these compnies, A.M. Best ssigned rtings to pproximtely percent of the rms where these rms held pproximtely 93 percent of the ssets of the industry. Obviously during the period of the 990 s, A.M. Best ws providing lmost complete coverge of the property-libility insurnce industry. By comprison, S&P provided rtings on only 8 percent of the rms in the industry in insurers - nd the number grew to 590 insurers by the end of Bsed on ssets, S&P does provide greter coverge s they were rting rms tht represented lmost 70 percent of the ssets of the industry by the end of the 2000 up from low of 24 percent in For exmple, Pottier nd Sommer (999) use four ctegory system to mp the individul rtings ssigned by ech gency. Both the GAO (994) nd Doherty nd Phillips (2002) use ve level system but the individul rtings ssigned to the ve ctegories di er slightly. In work not shown here, we compred the results of the ve level ctegoriztion system we dopted with the four level system suggested by Pottier nd Sommer. The primry conclusions re similr regrdless of which ctegoriztion is used. 2