An IPO s Impact on Rival Firms

Transcription

1 An IPO s Impact on Rval Frms By Matthew Spegel and Heather Tookes June 5, 2014 Yale School of Management, P.O. Box , New Haven CT Emal: matthew.spegel@yale.edu and heather.tookes@yale.edu. We would lke to thank semnar partcpants at Yeshva Unversty for helpful dscussons.

2 Abstract There s a long lterature documentng the process by whch frms conduct ntal publc offerngs (IPOs). However, there has been a relatve paucty of research nto how one frm s decson to swtch from beng prvate to publc mpacts ts rvals. Ths paper uses a structural approach to address that queston. We develop a contnuous tme model n whch heterogeneous frms producng heterogeneous goods compete for consumers. Because the model takes place n real tme t produces a structure wth parameters that can be estmated emprcally. Importantly, ths allows the emprcal work to make meanngful statements about parameter magntudes as well as sgns. In general, when frms conduct IPOs, they are relatvely small and ther mpact on rval values amounts to a few bass ponts one way or the other. However, we do fnd that new ssues presage a more homogenous product envronment whch ultmately cuts down on ndustry profts. The paper s structural model also offers a new way to test for whether an IPO causes these ndustry changes or smply presages them. Roughly, f the IPO makes the newly publc frm stronger, the IPO should beneft the IPO frm and negatvely mpact ts rvals. By comparng forecasts of changes n fundamentals and out-of-sample forecast errors over tme, one can see f the IPO frm s future looks fundamentally dfferent from ts rvals. Our tests ndcate they look the same, mplyng that IPOs forecast future ndustry changes but do not cause them.

3 An ntal publc offerng (IPO) s a major event n the lfe of any frm. But what does an IPO mply for the ndustry s future? Does gong publc sgnal that compettve pressures wll ncrease or decrease? Pressures may ncrease f the newly publc frm s now a more formdable rval. Alternatvely, the IPO frm s now mandatory dsclosures may prove useful to rvals that can now better copy ts strategy. It s also possble that the frm may decde to go publc to take advantage of new opportuntes facng the ndustry, mplyng greater proftablty n the future for t and ts rvals. Typcally, papers n corporate fnance focus on one dmenson of problems lke ths, or they attempt to characterze a domnant effect, to be appled across all ndustres. Ths paper takes a structural approach that allows dfferent ndustres to progress n dfferent ways post IPO. We uncover a great deal of heterogenety n the data, whch mproves our understandng of the range of economc forces that are assocated wth IPO actvty. If one s forced to make a sweepng generalzaton, then ths paper fnds an IPO augurs n an era of reduced profts and greater consumer moblty wthn an ndustry. However, n the upper tal of the nterquartle range, ndustres are forecasted to see hgher profts and lower consumer moblty. If the goal s to see what an IPO may portend for an ndustry then t seems natural to defne an ndustry as a set of frms that compete wth each other for the same customers as a practcal matter those based on 4-dgt SIC codes. 1 However, the fact that there are hundreds of 4 dgt ndustres makes t dffcult to know what varables should or should not be ncluded. A structural model offers a potental soluton. Usng one that descrbes ndustry dynamcs wth a relatvely small set of varables makes estmaton on an ndustry-by-ndustry bass feasble. Here, we begn wth the Spegel and Tookes (2013) contnuous tme model of competton n a heterogeneous product olgopoly. Ther model s then modfed to allow for a gradual change n the compettve envronment over tme post IPO. The settng s general enough that the change can be due to a number of factors that play out n a varety of ways. For 1 It may be that an IPO mples one thng about the IPO frm s rvals. Then another for frms wthn ndustres downstream and upstream from t. There s, however, no reason to beleve the mplcatons should be the same across all three groups. For example, suppose an IPO leads ncreased competton n the frm s own ndustry. That should produce hgher profts for ts downstream customers along wth hgher unt sales but overall lower profts for the IPO frm s own ndustry.

4 example, nformaton dssemnated by the IPO frm may ncrease or decrease compettve pressures wthn the ndustry, leadng to changes n spendng on customer acquston. Unlke a purely emprcal model, a structural one can be tested on the bass of ts dynamc forecasts. If those prove accurate then t at least sets a benchmark aganst whch other explanatons (and ultmately forecasts) can be measured. Several tests show that ths paper s dynamc model does qute well relatve to current alternatves n ths regard. The model generally does well at capturng changes to a frm s profts and values after a rval s IPO. For example, the model yelds an n-sample R 2 statstc of nearly 17% when 3 years of proftablty data are used and 34% wth 5 years of data. To see what ths means, compare these values to those found n Hsu, Reed and Rocholl (2010). They have a purely emprcal model that seeks to explan operatng proftablty for rval frms post IPO and report n-sample R 2 statstcs of approxmately 4%. The data ft from ths paper s structural model comes about despte the fact that the emprcal model requres only four estmated ndustry parameters and three frm-specfc parameters. When compared wth the results typcally seen n emprcal corporate fnance, where far more ndependent varables are used, the ft produced here s qute good. Beyond ths general n-sample comparson, the paper also presents predctve regressons n whch the model s used to forecast future changes n proftablty and value. Whle the R 2 naturally declnes out-of-sample, the model stll performs well compared to the other varables from the lterature. We present tables from a horse race between our model s perod-ahead forecast and the perod-ahead forecasts based on other ndependent varables seen n the lterature. All of them show that, f you have to restrct yourself to one or two ndependent varables, then our model s predcted values and profts should always be among them. Most of the tme, f you are restrcted to just one, t would be our model s forecast. Overall, the model ndcates that an IPO s generally bad news for an ndustry s future profts per unt of market share. Dependng on the estmaton wndow used (3, 5 or 10 years of data) the medan ndustry wll see a long term drop of between 10% and 25%. However, the estmated heterogenety 2

5 Bn Percentage across ndustres s qute large wth an nterquartle range between 60% and +40%. If forced to provde a broad characterzaton of what happens, the hypothess that the nformaton released from an IPO leads frms to a more homogenous form of product competton (and thus lower profts per unt sold) appears to domnate. The parameter estmates ndcate that post-ipo t becomes 3 to 4 tmes easer to lure away a rval s customers. An example of ths type of market evoluton can be seen n the cell phone ndustry. As a number of artcles have noted, unt sales are up but profts are down. 2 The generally accepted reason s that, as tme has passed, the product offerngs have become more homogenous, whch has ncreased prce pressure. Unlke a statc model or ts emprcal analog a structural model s parameters produce mplcatons about magntudes rather than just sgns. Ths permts one to ask a number of questons that are otherwse dffcult to address. For example, are the estmates economcally reasonable? In the case of IPOs t s partcularly mportant to retan some perspectve on what s occurrng and what role the IPO frm may be playng. As Fgure 1 shows, when frms go publc they typcally have very small market shares Dgt 4-Dgt Market Share Fgure 1: IPO frm's market share. 2 See, for example, Elmer-DeWtt (2014) and Mller (2014). 3

6 Bn Percentage The hstogram dsplays the market shares of frms as of ther IPO dates usng both the 2- and 4-dgt Standard Industral Codes (SIC), snce both defntons appear n pror studes. Frms wth market shares of more than 10% are relatvely rare, comprsng less than 1% of all IPO frms usng 2-dgt ndustres and fewer than 12% of all IPO frms usng 4-dgt ndustres. In fact, the vast majorty of IPO frms have market shares that are smaller than 1%. In ndustres defned at the 2-dgt level, over 90% of the frms undertakng an IPO have market shares under 1%. Usng 4-dgt ndustres, ths fgure naturally drops, but stll remans at just over 50%. IPO frms are not only small, but over the three years followng the IPO, ther market shares do not change very much. As shown n Fgure 2, over 90% of the IPO frms see ther market shares change n absolute value by less than 1% usng 2-dgt ndustres. Even at the 4-dgt level, more than 60% see changes n ther absolute market share of less than 1% n the subsequent 3 years. Ths s consstent wth the results n Chemmanur, He and Nandy (2010). They fnd that post-ipo, the newly publc frm s sales growth and productvty declne and that ts market share changes very lttle over the next few years Dgt 4-Dgt 3 Year Absolute Change n Market Share Fgure 2: IPO frm growth n the followng 3 years. 3 Ths rather mundane productvty performance for the IPO frm s also accompaned by long run stock return underperformance (Rtter and Welch, 2002). 4

7 The fact that IPO frms are small and grow rather lttle over the next few years suggests that f one s lookng to measure the extent to whch a frm s decson to go publc changes ts compettve stance and mpacts other frms then the magntude should be rather small. 4 Alternatvely, f the IPO frm s ether a condut through whch the ndustry changes (va newly released nformaton about the reasons behnd the newly publc frm s success) or smply a portent (lke a canary n a coal mne) then the mpact can potentally be much larger. The nterquartle range of ndustry value changes n our sample ranges from 4% to +3% over the long run post-ipo. These results ndcate that overall an IPO s mpact on an ndustry s value s qute modest, especally when contrasted wth some of the pror results n the lterature. However, gven the evdence n Fgure 1 and Fgure 2, these values seem to be more n lne wth what economc ntuton mght lead one to expect. Whle any model can potentally be judged by whether or not ts results are reasonable n magntude; reasonable may le n the eye of the beholder. A more concrete test s to determne whether or not the model s estmated parameters tell us anythng useful about how an ndustry evolves post-ipo. That s, can the model forecast ndustry dynamcs out of sample? The short answer, here, s yes. Furthermore, our tests show that forecastng wth the model and ts estmated parameters does a better job out-of-sample than other emprcal varables that have been used n the IPO lterature. The out-of-sample performance of the model s partcularly mportant, gven the recent crtcsm that dynamc structural models n corporate fnance are not held to hgh test standards, makng them dffcult to falsfy (Welch, 2013). 5 These forecasts are dffcult to structure wthn emprcal work based on statc models and offer a unque hurdle that dynamc structural models can potentally clear. Focusng on predcton also offers a natural way to rank varous explanatons; better ones presumably produce superor forecasts. 4 Market shares for IPO frms that we observe through year t+3 do ncrease on average; however, we observe 3,290 IPO frms at date 0 and only 2,569 by the end of year 3 followng the IPO. Whle some of the frms dsappear from the sample due to M&A actvty, many also dsappear because of falure (.e., market shares approach zero). 5 Strebulaev and Whted (2013) argue aganst ths clam, pontng out that structural models allow us to glean relatonshps that are mpossble to observe wth statc reduced form models and that some have predctve power. Hennessy (2013) emphaszes the ablty of dynamc structural models to provde quanttatve predctons. 5

8 Forecasts not only test a model aganst the data, but also offer a wndow nto causalty. In general, good news for one frm should be bad news for ts compettors and vce versa. For example, f gong publc makes a frm stronger, ts forecasted profts per unt of market share should ncrease and ts rvals proftablty should decrease. Alternatvely, f the IPO leads to the transmsson of formerly prvate nformaton useful to the frm s rvals, then the opposte should be true. Instead, the paper fnds that estmated parameter changes pre- and post-ipo look smlar for both the newly publc frm and ts rvals, ndcatng the IPO s best descrbed as a canary n the coal mne rather than a causal compettve event. Another way to test for causalty s to look out of sample. If the IPO causes future changes to the ssuer s compettve prospects ts forecast errors should be negatvely related to those of ts rvals. However, we fnd that ths correlaton s postve, provdng further evdence that an IPO presages events rather than causes them. To our knowledge, ths s the frst paper to employ structural parameter change estmates and forecasts as a way to test an event s causal relatonshp to future changes wthn an ndustry. One can thnk of these tests as structural model analogs to a dfferences-n-dfferences (DID) analyss. In a DID test the target or event frm s matched another frm that has smlar characterstcs along a few dmensons but s not assocated wth the event. Matchng crtera often nclude selectng from the same ndustry as the target frm. Unfortunately, matches wthn 4-dgt ndustres are often mpossble frequently leadng to the use of 2-dgt ndustry groups. 6 However, as noted earler, 2-dgt ndustres are rather broad. If the event n queston arses from ndustry specfc changes among frms competng wth each other, then a match at the 2-dgt ndustry level may not correct for that. The structural test proposed here has the potental advantage of usng the frm s own ndustry as ts benchmark. Changes wthn the set of compettors are thus pcked up whch can help address the concern that observed effects are due to changes n a specfc ndustry rather than from the event beng studed. 6 DID tests have appeared n the IPO lterature snce there s lttle or no pre-event data on the ssung frm. However, t has proven useful n other contexts where there s both pre-event data and researchers have been wllng to use broad ndustry defntons n order to fnd reasonably close matches along varous dmensons. Recent papers usng 2-dgt SIC codes nclude Agrawal and Nasser (2012), Almeda, et al. (2011) and Gormley and Matsa (2011). 6

9 Whle the IPO lterature s volumnous (see Rtter and Welch (2002) and Ljungqvst (2008) for excellent surveys), we are only aware of three other artcles that explore emprcally how a frm s decson to go publc mpacts the values of other frms n ts ndustry. One s Hsu, Reed and Rocholl (HRR) (2010). Ther paper looks at 2-dgt SIC ndustres and asks how well they perform after a very large IPO n the ndustry. We are nterested n IPO events more generally, so we nclude both large and small IPO frms n our sample. As dscussed and shown n Fgure 1 above, the vast majorty of ssuers are relatvely small, especally at the 2-dgt level. A second paper that examnes an IPO s mpact on ts rvals s Chod and Lyandres (2011). 7 Lke us, they examne 4-dgt SIC ndustres. The Chod and Lyandres (CL) paper begns wth the development of a statc model that they use to motvate a subsequent regresson analyss. In ther model, when frms go publc the founders can dversfy ther portfolos and then take a more aggressve (rsker) stand n the product market. As wth any statc model, the best one can do s verfy whether or not the regresson parameters are consstent wth t. CL fnd that a measure developed n Sundaram, John and John (1996) that examnes cross frm demand elastctes produces estmates n the drecton ther model ndcates t should. Because the Sundaram, John and John (1996) measure s related to some of deas developed here, the text contans fuller dscusson and ncludes a comparson of t wth the procedures employed here. The thrd paper to look at how an IPO mpacts ts ndustry s Chemmanur and He (CH) (2011). They begn wth a statc model n whch gong publc allows a frm to obtan lower cost external fnancng. Ths becomes preferable to fnancng growth nternally f there s a productvty shock. The goal of ther paper s to help explan why we see IPOs waves wthn ndustres. As part of ther analyss, they look at market share growth post IPO across frms n the ndustry and fnd that those that go publc gan relatve to ther prvate rvals. 7 Maksmovc and Pchler (2001) also examne IPOs n compettve settngs, but ther focus s on explanng the tmng of offerngs and IPO waves wthn ndustres. An emprcal paper n ths area s De Jong, Hujgen, Marra and Roosenboom (2012). They fnd that those n ndustres wth lower entry barrers (as measured by captal ntensty) tend to go publc earler. A fully dynamc model of when a sngle frm should go publc can be found n Pastor and Verones (2005). Even though they do not explctly model a frm s compettors, we menton t here because they dscuss how partcular elements of an ndustry s structure mght affect the economc envronment they model. From ths, they then draw some conclusons regardng IPO decsons across ndustres. An explct dynamc olgopoly model wthn the IPO lterature can be found n Kang and Lowery (2014). Ther paper looks at the prcng of IPO servces. Here the focus s on how the IPO affects the IPO frm s own ndustry. 7

10 In addton to ts contrbuton to the IPO lterature, ths paper adds to the growng body of structural corporate fnance research. Promnent examples nclude Hennessy and Whted (2005, 2007), Strebulaev (2007), and Rddck and Whted (2009), whch focus on captal structure and nvestment dynamcs. 8 These papers provde tests of quanttatve predctons, n addton to qualtatve analyses (.e., tests of domnant effects) found n more common reduced form estmaton. Lke these papers, ours s clear about the objectve functons of frms and the ways n whch the frms choces over tme mpact future dynamcs. Our contrbutons le not only n the IPO applcaton, but also n the model s ablty to characterze the value dynamcs of entre ndustres. The paper s structured as follows: Secton I presents the structural model. Secton II contans the emprcal estmates. Secton III concludes. I. The Model The model begns wth the basc structure found n Spegel-Tookes (2013). Consder an ndustry contanng n frms that produce a heterogeneous product. Frm competes wth ts rvals for market share (m ) by spendng funds u (t) at tme t on customer acquston. Here customer acquston should be construed qute broadly. Advertsng s perhaps the most obvous way frms acqure market share. But so s research and development on mproved product desgn. In some ndustres customer acquston may nclude captal expendtures that create outlets closer to where customers shop: McDonalds and Starbucks are two examples of frms that compete for market share n ths way. The model assumes that market share evolves over tme va: sut dmt m n t dt s ju jt j1 (1) 8 See Strebulaev and Whted (2012) for a survey. 8

11 Of course, some frms are more effcent at customer acquston than others. Ths heterogenety s captured through the varable s. Hgher values of s mply a frm acheves a greater bang for ts buck when tryng to gan market share. The varable represents customer loyalty wth hgh values ndcatng that t s relatvely easy to lure away a rval s customers. To allow for a rcher comparson wth the pror IPO lterature t s useful to add a stochastc ndustry sze component to the basc Spegel-Tookes (2013) model. Let ς represent an ndustry s sze measure and assume t follows the law of moton d gdt dz (2) where dz s a standard Brownan moton. In the model, all else equal, nstantaneous corporate profts are proportonal to ndustry sze and are gven by t t e m t u t f (3) The parameter α translates a unt of market share nto corporate profts gross of ts spendng on customer acquston and ts fxed operatng costs f. Frms seek to maxmze ther expected present dscounted value rt E e tdt, where r (assumed to be greater than ς) s the dscount rate. Lettng δ equal t0 r g ς the HJB equaton for ths problem can be wrtten as 2 max m f u V m V V 0 n. u j1 us us j j 1 2 (4) The formulaton of (4) ncorporates the soluton to the Spegel-Tookes (2013) model along wth the guess that the addton of the stochastc term n (3) wll produce a value functon V(ς) of the form e ς V 9

12 along wth a term arsng from the Brownan moton process. Not too surprsngly, ths guess wll ndeed work and produce a soluton to (4). A. Gradual Incorporaton of Informaton from an IPO The prmary objectve of ths paper s to apply the basc Spegel-Tookes (2013) model gven above to the queston of how an IPO mpacts ts ndustry. When a frm conducts an IPO, t s forced to release qute a bt of nformaton about ts sales and operatons. Whle ths nformaton may help nvestors, t also may prove to be of use to rval frms that can to work towards ncorporatng the newly gleaned nformaton nto ther own strateges and product offerngs. (Ths, presumably, s why the nformaton was hdden pror to the IPO.) Alternatvely, antcpated changes n ndustry structure (due to new producton technologes, consumer preferences or other factors) may nduce successful frms to go publc. In ths case, the IPO contans nformaton about the ndustry rather than frm-specfc compettve nformaton. The Spegel-Tookes (2013) framework allows for both of these possbltes. Assume that, pror to the IPO, frms are endowed wth a set of parameter values for ther profts per unt of market share,, 1 n, spendng effcency s * s *,, * 1 sn and fxed operatng costs f * f *,, * 1 fn * * *. For smplcty assume frm n conducts the IPO. Over tme, the ndustry wll adapt to the nformaton the IPO frm s forced to release eventually leadng to new parameter sets, s s s 1,, n,, 1 n and f f,, 1 fn. These new parameters may be favorable relatve to exstng parameters. They may also mply losses. The overall mpact on value wll depend on consumer demand and on the compettve responses of all frms n the ndustry. Assume that the transton to the new parameter occurs at an exponental rate ψ so that at tme t a frm s actual profts per unt of market share, spendng effcency, and fxed costs are gven by, s and f : t * t e 1 e t e, (5) 10

13 t * t s e s 1 e s t s e s, (6) and * e t t f f 1 e f t f e f (7) where wrtten as *, * s s s and * f f f. Under ths specfcaton the HJB equaton can be su 0 max e m f e f u V m n V V u su j j j1 t t m t (8) where superscrpts on the value functon V ndcate partal dervatves wth respect to market share (m ) and tme (t). To solve (8) assume V has the form: V m, t a t b t m t. (9) Under ths assumpton, the frst order condton for the optmal u and the resultng HJB equaton has the same form as n Spegel-Tookes (2013). After defnng t ˆ (10) e and n 1 zˆ (11) ˆ s j1 j j Followng the steps n Spegel-Tookes (2013) one can wrte the soluton to u as 11

14 u ˆ n 1 b s zˆ n 1 ˆ szˆ 2. (12) Usng (12) and (9) n (8) and the applyng some addtonal algebra, the problem comes down to solvng two equatons; one for a and one for b. In Spegel-Tookes (2013), these equatons are algebrac. In ths case, they are the ODEs for a k 1 1 ˆ 1 b n b s z n t 0 e f f n 1 b szˆ n 1 bs 2 n 1 ˆ 1 b n b s z n n b s zˆ n ˆ 1 n b s zˆ n 1 1 k k 1 1 b n b ksk z n a a t (13) and for b t t 0 e b b. (14) Equaton (14) yelds a soluton for b of e t b ke t (15) where k s an arbtrary constant of ntegraton. Snce the soluton to b at t goes to nfnty must approach the soluton where Δα equals 0, ths mples k=0. Thus the soluton for b s t e b. (16) In general, there does not appear to be a closed form soluton for the a s. However, after pluggng (16) n for the b n (13) one can wrte the non-lnear ODE for solvng a as 12

15 2 ˆ ˆ s zˆ n 1 ˆ szˆ t 0 e f f a a. t 2 (17) The above does not, n general, admt a closed form soluton because the ˆ szˆ terms vary over tme. However, f one assumes that t and s s 1 kse e ˆ t t k e for ndustry wde constants k α and k s (mplyng each frm experences a proportonal change to ts profts per unt of market share and marketng capabltes post-ipo) then t s easy to show that the ˆ szˆ terms reduce to the constant ˆ s zˆ s z (18) where z * s defned as s j 1 j j. Pluggng ths nto (17) t becomes t k e s z n 1 2 sz t t 0 e f f a a. 2 (19) The soluton to (19) has the same form as that for (16) s z s z 2 2 f s z n 1 k s z n 1 t f a e. 2 2 (20) The solutons gven by (9), (16) and (20) can be readly appled to ndustry data near IPO events. Dong so can help nform us about the economc forces drvng the value and proft dynamcs of rval frms that follow publc offerngs. II. EMPIRICAL ANALYSIS OF THE COMPETITIVE EFFECTS OF IPOS Dealng wth data coverng a wde varety of 4-dgt SIC ndustres can potentally requre an emprcal model that ncludes so many controls the results become nearly meanngless. One potental 13

16 soluton s to reduce the number of ndustres by smply swtchng to 2-dgt groupngs. 9 Whle 2-dgt codes reduce the problem s dmensonalty, the resultng estmates are now based on groupngs comprsng relatvely broad sectors. Often ths s harmless and, better yet, such an analyss can be unquely nformatve. But, f the goal s to see how one frm s actons affect ts compettors then t s mportant to recognze that 2-dgt ndustres nclude frms that are not n fact compettors. As an example, consder the 2-dgt SIC ndustry labeled 28. Ths represents frms n chemcal and alled products. It ncludes frms that make plastc materals and resns (2821), dagnostc substances (2835) and perfumes and cosmetcs (2844). It s dffcult to magne that frms n these three areas consder each other compettors. To get a better dea of just what types of frms populate the SIC 28 two dgt ndustry, consder Advanced Polymer Systems 10, whch s n ndustry It K descrbes ts busness as, Advanced Polymer Systems, Inc. and subsdares ("APS" or the "Company") s usng ts patented Mcrosponge(R) delvery systems and related propretary technologes to enhance the safety, effectveness and aesthetc qualty of topcal prescrpton, over-the-counter ("OTC") and personal care products. Compare APS wth Guest Supply (2844) whose K flng ncluded the statement that, The Company operates prncpally as a manufacturer, packager and dstrbutor of personal care guest amentes, housekeepng supples, room accessores and textles to the lodgng ndustry. The Company also manufactures and packages personal care products for major consumer products and retal companes. It seems extremely unlkely that n 1997 APS and Guest Supply saw each other as compettors. 11 If they had any relatonshp, t lkely nvolved Guest Supply as a customer of APS. Whle large aggregatons of frms nto 2-dgt ndustres may be useful for some applcatons, f we hope to explore how IPOs mpact a frm s rvals then one needs to use at least 4-dgt ndustres. 9 Ths s lkely why HRR used 2-dgt Standard Industral Codes n ther analyss. 10 Whch became AP Pharma and s now Heron Theraputcs. 11 Both Advanced Polymer Systems and Guest Supply are part of the paper s IPO database. 14

17 A. Estmaton An mportant advantage of the model s that t characterzes the value dynamcs resultng from changes n the compettve structure of the ndustry n a way that s amenable to emprcal estmaton and testng. In ths secton, we present results from estmates of the key model parameters. The paper focuses on nnovatons n proftablty per unt market share (α) and changes n consumer loyalty ( ), but, as shown n the prevous secton, the model can easly be extended to nvestgate other potental shocks (e.g., to the fxed costs of operatons) as well. The proftablty equaton provdes a structure for estmatng both frm-specfc and ndustry-wde gt parameter values. Recall that the basc proft functon s gven by: ( t) e ( m ( t) u ( t) f ). gt Followng Spegel and Tookes (2013) let ( t) e u ( t ) ˆ () t = (Revenue Cost of Goods Sold). Ths can then be adapted to the pre-spendng proftablty equaton to ncorporate the slow nformaton revelaton underlyng the model. To keep the emprcal problem manageable, wth the data at hand, the focus here s on the transton from * to (for smplcty, set * f f ). Under the assumpton that, for every frm, k *, the model estmates: gt gt t * * t u t k k e m t f ( ) e ( ) e (( ( 1) ) ) ( ). (21) Usng only data on revenue and costs of goods sold, we can use non-lnear least squares to obtan estmates for * *, g,,, and f. Recall that defnes the transton rate from the pre-ipo * to the new. Call ths the transton rate and the assocated nformaton revelaton n broad terms. For example, the IPO decson could be the result of IPO-frm s desre to obtan publc fnancng n order to take advantage of an mpendng postve ndustry-wde shock (that wll take tme to fully mpact the frms n the ndustry). The IPO decson could also be the result of the IPO frm s managers assessment of an mpendng negatve shock that wll decrease the benefts assocated wth remanng prvate. 15

18 In order to estmate model-mpled values, we also need estmates for consumer responsveness and compettve strength ( and sz, respectvely). Recall that, because proftablty evolves at the common rate, then the transton from * to has no mpact on sz. Lettng m equal frm s steady state market share, one can follow Spegel and Tookes (2013) and use the equaton dm ( m m ( t )) dt, whch has a soluton for m (t) of : t m ( t) m ( m (0) m ) e. (22) Equaton (22) can be estmated va non-lnear least squares and provdes estmates for each frm s well as the ndustry parameter. It also gves frm-specfc compettve strength sz snce, n steady m as state, m 1 n 1 s nz. The emprcal model estmates pre and post IPO consumer responsveness parameters (ϕ 0 and ϕ respectvely) by allowng t to change as of the IPO date. Usng data from the pre and post IPO perod then allows equaton (22) to generate estmates of ϕ 0 and ϕ. Followng Spegel and Tookes (2013), the only restrcton that we mpose s that s non negatve and less than 25 (n our quarterly estmaton, ths would correspond to a customer half-lfe of just 2.5 days). B. Data and Summary Statstcs The model s estmated for each IPO event usng quarterly Compustat and CRSP data for all rval frms that share the IPO frms 4-dgt SIC codes as recorded by Compustat. 12 The ntal sample of IPO events are from Securtes Data Corporaton New Issues Database and ncludes the IPOs of U.S. publcly lsted stocks from 1983 through Because we are nterested n olgopolstc competton, only the 3,290 IPO events that occurred n ndustres wth 50 or fewer compettors are ncluded n the ntal sample. All publcly traded rval frms for whch we have market share data at the begnnng of the estmaton perod are ncluded n the estmaton. We begn the estmaton at the IPO quarter and estmate 12 Pror work ndcates Compustat s SIC codes do a better job than those generated by CRSP when t comes to ndcatng whch frms vew themselves as belongng to the same ndustry. For example see Guenther and Rosman (1994). 13 We exclude fnancals and utltes (SIC codes and ). 16

19 the model over horzons that nclude data through 3, 5 and 10 years post-ipo (.e., usng a total of 24, 32 and 52 quarters of data for each frm respectvely). Parameter estmates are shown n Table 1. We obtan estmates for between 726 and 843 events, dependng on the estmaton horzon. 14 The medan estmated ranges between to Ths mples that, followng the medan IPO event, between 17% and 32% of the transton from the old to new proftablty regme occurs n the frst quarter. At the end of four quarters the transton s between 52% and 78% complete. By the end of the second year between 77% and 95% has occurred. There s however substantal cross-sectonal varaton n. For example, the nterquartle range of estmates usng data for the 10-year horzon s to Ths mples that, for slower transton ndustres, only 9% of the proftablty change occurs n the frst quarter. For fast ones, ths value s 38%. Table 1 also provdes estmates of the value of the proftablty shock k, where k *. Across all of the estmaton horzons, the medan estmated value of the proftablty shock k s less than 1. Ths mples that IPO events are more often than not followed by a reducton n ndustry-wde profts per unt of market share. However, as n the case of, there s substantal cross-sectonal varaton. The nterquartle range for the estmated value of k usng data over the 10-years followng the IPO s to Ths mples that rval frms experence between a 59.5% drop and a 33.3% ncrease n proft per unt market share followng the IPO. There are some IPO events for whch estmated parameters are not reasonable (for example, the maxmum estmate for k for the 3-year estmaton horzon s 740,827); however, most are qute plausble. As noted earler, no model can be expected to ft every ndustry and the one n ths paper s no excepton. Whle ndustry proftablty tends to decrease followng IPOs, the ndustres are stll growng. The medan estmated quarterly real ndustry growth rate s near 0.5% per quarter under all specfcatons. 14 The model convergence rate s approxmately 20-25%. As mentoned n the ntroducton, we do not clam that the model s sutable for all ndustres. For example, the model s not ntended for ndustres for whch r<g. 17

20 Ths quarterly growth parameter vares wthn a reasonable range (for example, based on the estmates usng data for the 10-year wndow, we obtan estmates wth an nterquartle range of 0.6% to 2.1%). The medan consumer responsveness parameter pre-ipo (stll usng the 10-year estmaton horzon as an example) of mples that, for the medan ndustry, t would take a compettor about 33vquarters to lose half of ts customers f t completely stopped spendng to attract them. (For expostonal purposes, call ths the market share half-lfe.) Post-IPO t appears the medan ndustry transtons to a state where consumers are much more wllng to swtch brands. The medan ϕ s whch mples a market share half-lfe of only 12 quarters. Medan pre-ipo consumer responsveness n the IPO sample s more than 50% slower than n the broad ndustry sample n Spegel and Tookes (2013). Post-IPO the medan values n the two studes look very smlar. Economcally, ths seems to mply that ndustres wth mpendng IPOs contan frms producng products that consumers vew as relatvely unque, at least when compared wth the typcal ndustry n the overall economy. Post-IPO, the ndustry transtons to a state where consumers are about as loyal to a partcular product as elsewhere n the economy. The ϕ parameter n ths model has a smlar nterpretaton to the Compettve Strategy Measure (CSM) developed by Sundaram, John and John (1996) and employed by Chod and Lyandres (2011). In Chod and Lyandres (CL) the CSM s used to proxy for the degree of compettve nteracton among frms n an ndustry. Usng ths paper s notaton, CSM for frm can be wrtten as CSM corr, S S (23) where S equals the change n frm 's sales and S the change n the sales of ts rvals. The dea s to provde a smple way to capture the degree to whch frms n an ndustry pull away each other s customers. Whle the ntuton s useful, one mmedate ssue that needs to be dealt wth when comparng the CSM to our ϕ parameter s that the frst varable n Equaton (23) s unt free whle the second s not. 18

21 Nevertheless, snce sales map nto market shares Equaton (23) can be reformulated to ft ths paper s model. The algebra s carred out n the appendx but the end result s the followng expresson g e 1u f CSM corr m e m g m S e mt 1 e 1 mt 1 mt S 0 g gt t, t1 1 t 1 t 0 (24) where S 0 s ntal aggregate sales. Gven the complexty of the above measure, t s perhaps not surprsng that the correlaton between the estmated values of CSM and ϕ are near zero. Some of the problem les n the denomnator of (23) and thus (24). The change n sales s not strctly postve or negatve. Values near 0 therefore blow up the left hand term and, when sales flp sgn, the change n the left hand sde can be dramatc (as llustrated n Fgure 3). In contrast, the estmates of ϕ usng (22) do not have smlar problems. Ths comparson shows how a structural model can help produce parameter estmates of nterest that may be more stable than those usng an estmator based on economc ntuton. The fnal nput to the frm value functon s, the cost of captal mnus the growth rate. We defne as the long-run (1926 through perod t) hstorcal market rsk premum plus the rsk-free rate mnus the long-run GDP growth rate. Whle the changes n profts per unt of market share (k) and consumer responsveness (φ 0 versus φ) n Table 1 may seem large, the actual estmated long-run change n market value for the ndustry post- IPO s actually qute modest. To see ths, we begn by pluggng the estmated parameters nto Equatons (9), (16) and (20) wth t set to zero and nfnty to get the pre and post IPO frm values, respectvely. We then calculate the rato of the post to pre IPO ndustry market values. Frms whch have an estmated value change n excess of a factor of 100 are dropped (post/pre of less than 0.01 or greater than 100). Next, for each ndustry and date, the frm medan, value weghted mean and equally weghted mean ratos are calculated. Fnally, the mean or medan value across all dates s calculated for each IPO event. The results from ths exercse are n Table 2. For the medan ndustry, the value change post-ipo s near zero. However, the cross secton exhbts consderable varablty. The medan frm n the medan ndustry on 19

22 the medan date has nterquartle range between 96% and 101% usng a 3 year estmaton horzon to a range of 96% and 113% usng a 10 year estmaton horzon. Gven how small most IPOs frms are, these are fgures n the range one mght expect. They also show that there s qute a bt of varablty across ndustres. Although the HRR paper focuses on 134 large IPOs at the 2-dgt ndustry level, t s worth usng ther results as a benchmark. They report that, wthn days of an IPO, the ndustry sees a loss n market value of somewhere between 0.5% and 1.0%. Ths s broadly consstent wth the average value changes that we fnd n Table 2 usng a broader sample of IPOs and analyzng rvals at the 4-dgt ndustry level. HRR credt ths change to the compettve advantages the IPO frm sees from gong publc. However, as noted n the ntroducton, even for large IPOs, the market shares of IPO frms are generally small, especally at the 2-dgt level (the 63 frms wth the largest market shares at the 2-dgt level n our sample have a medan market share of 5.5%). Gven ther small sze, the observed value effect on rval frms may be due to forces other than the IPO frms nducng large expected losses on a broad group of compettors. Ths s especally true when usng 2-dgt SIC codes snce ndustres wll nclude any combnaton of an IPO frm s supplers, customers and others wth whch t has no meanngful busness relatonshp. C. In-Sample Estmates Table 1 and Table 2 ndcate that an emprcal model that smply says ndustres lose value post IPO may be mssng some mportant heterogenety n the data, even f that s the medan result. Of course, lke any structural model, the one estmated here was clearly unable to ft some ndustres. Not too surprsngly, ths resulted n some very mplausble forecasted value changes. Because our focus s on ndustres for whch the model s relevant and because t s unlkely that a forecaster would use extreme or unreasonable estmates for predcton, we remove extreme value observatons from our sample. These are defned as observatons n whch: (1) IPO events wth estmated ψ, k, g or φ that are less than the 1st percentle of all estmates or greater than the 99th percentle; (2) model mpled or actual changes n frm 20

23 value (log rato of values) that are less than 1 or greater than +1; and (3) model-mpled or actual changes n proftablty that s greater (n absolute value) than the value of begnnng-of-perod assets. The parameter estmates * *, g,,, f, m and and market share data ( m t ) can be plugged nto the value and proftablty equatons to generate model-mpled changes n them. The model-mpled change n value s calculated as the log rato: V ( mt, t) ln V ( mt 1, t 1), where V ( mt, t ) s the value functon defned earler. 15 The actual value V t s n 2011 dollars and s defned as the market value of equty, plus the book value of assets, mnus the book value of equty and deferred taxes at the end of quarter t. Actual value change s calculated as: Vt ln Vt 1. To test the model, actual changes n frm value are regressed on the model-mpled changes. Test statstcs are calculated usng pooled data (all IPO events and all rval frms) and clustered standard errors at the IPO event level. Results are shown n Table 3. In all cases, the model mpled value changes predct those n the actual market. The coeffcents on the model mpled changes are all statstcally sgnfcant and range between and Roughly, ths mples that a 10% ncrease n model-mpled value s assocated wth an ncrease n actual value of between 0.57% and 0.35% n those ndustres for whch reasonable estmates were obtaned. The adjusted R 2 values range between and 0.003, whch s expected gven returns are the dependent varable. Smlar to the value change regressons, the table also ncludes tests to see whether model-mpled changes n proftablty explan actual changes. Let ( m ) equal model-mpled proftablty as gven n Equaton (21), dvded by total assets at tme t zero. Next, defne model-mpled change n proftablty as a( mt ) a( mt 1). Actual proftablty s calculated as revenue mnus cost of goods sold (n 2011 a t dollars) durng quarter t, dvded by t zero assets. Actual change n proftablty s the frst dfference of 15 Observatons n whch model-mpled (, ) t V m t are less than or equal to zero are mssng. The model assumes that the ndustry and frm parameters are such that there s no ext. Thus, these observatons are analogous to cases n whch the model does not converge. We do not clam that the model s approprate for all ndustres and these are examples of the cases n whch the model does not do a good job n characterzng ndustry and frm dynamcs. 21

24 quarterly proftablty. Table 3 presents the results from regressng actual proftablty changes on model mpled changes. An mmedate observaton from the table s that the explanatory power of the model s even greater for profts than t s for frm value. The R 2 statstc ranges from 17% to 34%. Ths may not seem surprsng at frst, gven that the proft functon s used to estmate key model parameters. However, note that the parameters are estmated n a regresson based on proft levels, not changes. 16 Estmated coeffcents on model-mpled changes n proftablty are all hghly sgnfcant and range from and These mply that a 10% ncrease n model-mpled proftablty s assocated wth between a 4.55% and 8.45% ncrease n actual proftablty. Under the model s assumptons, the model-mpled changes n value and proftablty are the only relevant explanatory varables n the value and proftablty regressons, respectvely. The fndngs n Table 3 confrm that these are mportant; however, n order to assess margnal mpact of the model, t s useful to study other varables from the lterature as well. Table 4 adds the explanatory varables from HRR. These are the lagged change n value (and proftablty, for the proftablty equaton), the natural log of total assets, ndustry market-to-book value, the annual level of IPO underprcng, frm age, an IPO dummy equal to one f perod t occurs durng years 0, 1, 2, or 3 relatve to the IPO, and IPO event fxed effects. Snce HRR s analyss was not proscrbed by an underlyng model, presumably they selected the varables for ther regresson equatons because they seemed lkely to offer the greatest chance of emprcally descrbng the data. Ths naturally leads to the queston of whether the model estmates produced here add anythng to what they have already documented. One potental advantage of a dynamc structural model s that t can pont to partcular specfcatons that are not obvous from the ntuton one mght get from a statc model. Table 4 looks nto ths ssue wth regard to ths paper s model. The results ndcate that the structural model s mpled changes help explan post-ipo changes to an ndustry beyond what can be sad wth the varables used n other papers. Note that the estmated coeffcents on the model mpled changes are smlar n magntude and statstcal sgnfcance to those n the analyss from Table 3, whch excluded the HRR controls. 16 To sharpen the nterpretaton, we also perform out-of-sample tests, as wll be dscussed later. 22

25 However, the HRR controls add nsghts of ther own: (1) age s generally negatvely related to changes n proftablty, (2) recent IPO underprcng s postvely related to value changes and profts and (3) ndustry market to book s negatvely related to value and postvely to proft changes. 17 None of the above three results from the HRR controls would have been predcted by the structural model. Ths ndcates that, even though the structural model captures qute a bt of the data s varablty, t does not explan all of t. CL predct that gong publc ncreases an IPO frms rsk-takng ncentves. To test ths hypothess, the authors lnk rvals returns near IPOs to the compettve strategy (CSM) measure defned n Equaton (23), ndustry demand uncertanty and the systematc porton of demand uncertanty. Because competton n ther model s characterzed by strategc substtutes, they nclude only those ndustres n whch ndustry CSM s negatve and they use the absolute value of CSM n ther regressons. 18 In Table 5, we repeat the extended HRR regressons from Table 4, and we add the CL varables. The sample sze s substantally smaller than n prevous tables because, for comparablty wth CL, we lmt our attenton to those ndustres n whch estmated CSM s negatve. The results ndcate that the absolute value of CSM, demand uncertanty and the systematc component of demand uncertanty are all postvely related to rvals value changes. 19 CL do not estmate proftablty regressons; however, we nclude them n Table 5 to mantan consstency wth the earler tables. In the case of proftablty, the coeffcents on the CL varables are more mxed, but overall they appear to be negatvely related to changes n proftablty. Importantly, the model-mpled changes n value and proftablty reman hghly sgnfcant, both statstcally and economcally, n all regressons. 17 The other control varables are not as consstent n ther sgns and statstcal sgnfcance. 18 Chod and Lyandres (2011) use 20 rollng quarters of hstorcal data to generate all three of these measures. They calculate CSM for each frm accordng to Equaton 22. Industry CSM s defned as the medan of the frm-by-frm estmates. Demand uncertanty n quarter t s the standard devaton of seasonally adjusted ndustry sales growth durng the pror 20 quarters. The systematc porton of demand uncertanty s the rato of the varance of the predcted values from a regresson of seasonally adjusted ndustry sales growth on the seasonally adjusted sales growth of all Compustat frms. See Chod and Lyandres (2011) for seasonal adjustment and further estmaton detals. In our sample, the dstrbutons of all 3 of these varables are comparable to those reported n ther paper. 19 Chod and Lyandres (2011) fnd that rvals value changes are postvely related to systematc uncertanty, negatvely related to total uncertanty and nsgnfcantly related to CSM. The postve and sgnfcant estmated coeffcent on the demand uncertanty n Table 5 s nconsstent wth ther fndngs; however, f we repeat the analyss usng only the Chod and Lyandres (2011) varables, ths coeffcent becomes statstcally nsgnfcant. 23

26 D. Out-of-Sample Tests The results n Table 3, Table 4 and Table 5 provde strong evdence of the model s emprcal valdty. Insample tests lke these are the standard assessment tools n the emprcal corporate fnance lterature. They help us to understand the degree to whch varous varables and models ft the hstorcal data and potentally explan what occurred. Whle these tests are valuable, t s also useful to know how well a model handles data out-of-sample. Ths not only allows one to see f over fttng has occurred, but also offers another avenue for assessng each model s relatve explanatory power. Comparng a dynamc model s ablty to forecast out-of-sample changes wth the statc models others have estmated s naturally problematc. Nevertheless, ths s an mportant ssue. Dynamc structural models have the potental to advance beyond the lmts of statc models by, n part, offerng a way to predct future events. However ths does beg the queston of how to mplement a forecast usng a statc model n order to compare the approaches. The followng sectons employ two methodologes towards ths end. Secton II.D.1 examnes what mght be called pseudo forecasts. In t, a perod t+1 projecton comes from the parameter set estmated wth data up to tme t, as n standard out-of-sample tests. However, data from tme t+1 s then used to forecast the dependent varable s perod t+1 value. The procedure s desgned to gve the statc model ts best chance of producng a superor forecast to the dynamc one developed here. Absent the use of perod t+1 data the statc model forecasts a constant value for the dependent varable gong forward; due to the use of constant parameter values along wth fxed ndependent varables. The pseudo forecast thus lets the statc model produce a more dynamc projecton, albet one that cannot be conducted n real tme. To see how the dynamc and statc models do n real tme. Secton II.D.2 conducts a set of true out-of-sample tests. In t, a perod t+1 forecast s made solely on the bass of data avalable as of perod t. Unlke the pseudo forecasts, these can be created n real tme. 1. Pseudo Out-of-Sample Tests 24

27 The left-hand-sde panel n Table 6 repeats the unvarate tests shown n Table 3, but nstead of n-sample regressons, t uses the parameters estmated over the ntal 3 and 5 year horzons, along wth real-tme market share data to predct quarterly value and proftablty changes over the next 3 and 5 years, respectvely. As noted above, whle the parameter estmates are based only on hstorcal data the forecasts then employ them wth data concurrent n tme wth the dependent varable. From Table 6 t s clear that the model performs well, even when out-of-sample parameter estmates are used n the forecasts. 20 All of the regressons produce postve and sgnfcant coeffcents on model-mpled value changes. Not too surprsngly, the model does a better job of explanng how profts evolve over tme than how market values change. The mddle and rght-hand-sde panels of Table 6 compare the model s out-of-sample performance relatve to the real tme HRR and CL varables. The model contnues to perform well, even after the ncluson of these addtonal explanatory varables. In all cases, the dynamc model s forecasts reman statstcally sgnfcant. For the value change estmates, the coeffcents on the model-mpled forecasts actually ncrease n magntude wth varables from HRR and CL are added, mplyng t s not smply a proxy for the nformaton they contan. Whle the HRR and CL varables add consderably to the R 2 statstcs when explanng value changes, they add lttle to the proft change regressons. In the latter, by tself, the dynamc model yelds an R 2 of 0.14 when usng a 3-year model and 0.26 when usng a 5 year model. Addng n both the HRR and CL varables only ncreases these values by less than These modest ncreases occur despte the fact that the dynamc model alone has just one ndependent varable n the regresson whle the HRR and CL models combned have Predctve Regressons 20 We use the term pseudo because the tests use real tme market share data. In the next secton, we conduct out-of-sample analyss n whch all explanatory varables are out-of-sample forecasts. The pseudo out-of-sample analyss s partcularly useful when we compare the model s performance to the HRR and CL varables whch come from statc models and are based on real tme data. 25