|
COPYRIGHT 2001 MIT Press Journals
I. Introduction
EVER since Ritter's seminal empirical study (Ritter, 1991), the post-issue performance of IPOs has been considered to be a puzzle. In the long run, IPOs significantly underperform standard benchmarks or equity in appropriately matched firms. The puzzle has been confirmed in numerous follow-up studies. (See, for example, the Spring 1993 issue of Financial Management, Jain and Kini (1994), and Loughran and Ritter, 1995.) The evidence is now generally interpreted as suggesting that the market is too optimistic when pricing young issues. It realizes its mistakes slowly, adjusting prices as the issues mature.
Although some have argued that the biases in the market's prior at the issue date are a natural consequence of shortsale restrictions (Miller, 1977; Morris, undated), it could also be a mere sign of the beliefs at a particular point in time. Indeed, most studies focus on IPOs executed during the 1970 and 1980s. That priors over this period were biased does not necessarily imply irrationality, because the bias was demonstrated to be there only ex post, that is, with the benefit of hindsight.
Instead, it seems much more fruitful to ask whether subsequent changes in the market's beliefs were rational. If beliefs can be expressed in terms of the chance numbers of classical probability theory, we know precisely what this means: changes should obey the rules of conditional probability (Bayes' law).(1) We will also take this to mean that the market knows the likelihood of the signals it receives given the eventual fate of an issue (will it default?). We set out to test this weaker restriction on market beliefs.
Although the market will be assumed to know the likelihood function, we do allow for biases in the market's prior at the IPO date, which means that it need not be confirmed in subsequent realizations. Therefore, we deviate from the standard view that market beliefs are unbiased in all respects (the efficient markets hypothesis (EMH)). But the deviation is minimal: we let the market have biased priors only at the issue date.
Both Bayesian updating and the use of correct-likelihood functions are an integral part of the learning that is assumed under EMH. Therefore, we will refer to our model of market beliefs as the hypothesis of an efficiently learning market (ELM). ELM differs from EMH only in that prior beliefs may be biased.(2)
This article tests ELM in the IPO aftermarket with a methodology that requires little or no information about the actual market prior at the issue date and how priors varied across issues. The methodology was originally developed by Bossaerts (1996, 1998) and successfully applied to experimental winner-take-all markets by Bondarenko and Bossaerts (2000), to digital option prices implied by index call and put option prices by Bondarenko (1997), and to straight index call options by Bossaerts (1998). The applications have one thing in common: they concern securities with a clear bankruptcy state, like the equity contracts studied here.
The specific framework of analysis is the following. At the launch date, it is known that a certain number of IPOs eventually fail (default), but it is not known exactly how many and, if the issue at hand does fail, at what time. For simplicity, the recovery rate conditional on bankruptcy is set equal to zero.
Initial priors about the probability of bankruptcy are arbitrary and may vary across IPOs. Price changes in the aftermarket reflect rational updating of these priors from news about the fate of the company. The market is supposed to understand how the news relates to bankruptcy. (It knows the likelihood of the signals given the bankruptcy status.) Likewise, the market correctly predicts the expected value of equity in the company conditional on no default. (If default occurs, the market of course knows that this value will be zero.)
The tests that we use to verify this belief model (an example of ELM) are simple and powerful. A novelty is that they require the empiricist to split the available sample in a winner category (companies that did not default) and a loser category (companies that defaulted). Standard returns have to be modified slightly and weighted appropriately. If rational updating is rejected, the sign of the statistic provides information on the nature of the inefficiency: whether the market overreacts or underreacts to new information.
The need to split the sample into winner and loser categories (in this article, we exclusively investigate the former) turns sample-selection bias into a virtue: the results will not suffer from survivorship bias by construction. That is, even if one does not know the exact proportion of winners and losers (perhaps because some histories of losers became unavailable), our tests remain valid.
Our methodology tests for correct updating of priors about the likely default of each company separately. The methodology is of the event-study type: each history is mapped in event time, with a common event time zero (the IPO date); one loses potential information from the knowledge that two histories occurred sequentially in calendar time. This implies, in particular, that our methodology does not investigate whether the default history of companies floated earlier in calendar time were correctly reflected in the priors at the issue date of subsequent IPOs. Our methodology allows there to be such updating (priors can vary arbitrarily across IPO histories), but it does not study its rationality.(3)
In a rigorous and comprehensive way, this article tests the conjecture made by Jegadeesh (1998), namely, that negative aftermarket IPO excess returns should be rationalizable in terms of the relatively negative news that the market received about IPOs. Jegadeesh documented that a large fraction of the aftermarket underperformance can be explained by three-day price changes around earnings announcements. However, he was not able to determine to what extent these price changes were "correct"; moreover, he focused on price reactions to well-identifiable events (earnings announcements). The methodology of this article tests whether price changes following any news event are rational, in the sense that they reflect Bayesian updating with a correct-likelihood function.
Unlike in Ritter (1991), we do not compare post-issue IPO returns with contemporaneous returns on benchmark portfolios (value-weighted or equally weighted market indices, and size-based portfolios) or matched securities (similar market capitalization and industry). Instead, we use an explicit intertemporal asset-pricing model that has been shown to be successful in other contexts, namely, Rubinstein's model (Rubinstein, 1976). His model uses a pricing kernel that is a simple nonlinear transformation of the return on a value-weighted market index.
Our data set is approximately triple the size of Ritter's, and covers 4,848 U.S. IPOs in the period 1975 to 1995. Although risk adjustment appears not to be necessary in the earlier part of our sample (covering Ritter's period of 1975-1984, when the risk premium recorded for U.S. stock markets was historically exceptionally low), it becomes important in the second part.(4)
It is also crucial that a value-weighted market index is used in Rubinstein's pricing kernel, in full consistency with the model: we will demonstrate that an equally weighted index is misspecified. In particular, as IPOs mature, their prices converge to Rubinstein's model with a value-weighted portfolio as benchmark. In contrast, the fit of Rubinstein's model with an equally weighted index does not improve with the age of IPOs.
The remainder of the article is organized as follows. The next section discusses the data and summarizes past evidence. Section III presents our approach and methodology. Section IV discusses the aggregate empirical results. Section V reports tests conditional on issue information. Section VI corroborates the findings by documenting how fast the market learns, and, hence, eliminates biases in the pricing. Section VII concludes.
II. The Evidence
Our results are based on a merging of Ritter's and van Bommel's sample(5) of IPOs in the 1975-1995 period and the CRSP 1998 NYSE/AMEX/NASDAQ monthly return tape. The two data sets were merged on the basis of PERM number, or CUSIP when the PERM number was not available.
The following three IPOs were eliminated.
* All the IPOs for which the EXCHANGE variable in Ritter's dataset is equal to 4, corresponding to non-Nasdaq OTC issues.
* IPOs for which the CRSP variable SHRCD, the share code, differs from 10 or 11. These are certificates, ADRs, SBI (shares of beneficial interest), units, companies incorporated outside the United States, Americus primes and scores, closed-end funds, closed-end fund companies incorporated outside the United States, and REITS.
* IPOs of which the first digit of the SIC code equals 6 or 9, corresponding to financial institutions, insurance, savings and loans (6), and utilities (9).
The latter exclusion was decided on because many IPOs in that category were in fact well-established firms that issued stock on the NYSE/AMEX/NASDAQ because of regulatory changes. The nature of these IPOs differs dramatically from that of the typical one, wherein a young company is floated, usually in a new area of industrial activity.(6)
The final sample contained 4,848 IPOs, about triple Ritter's data set. For each IPO, trading and delisting information, as well as a vector of 120 post-issue monthly returns were extracted from the CRSP tape (together with 120 synchronous CRSP equally weighted and value-weighted index returns). This means that ten-year post-issue performance histories for each IPO were available, except of course for the issues from 1989 on, for which...
Read the full article for free courtesy of your local library.
|
Bookmark this article
Print this article
Link to this article
Email this article
Digg It!
Add to del.icio.us
RSS
