The Aftermarket in High-Tech IPOs

The Aftermarket in High-Tech IPOs By Sanjiv Jaggia Satish Thosar* Department of Economics School of Finance and Economics Suffolk University University of Technology, Sydney 8 Ashburton Place P.O. Box 123 Boston, MA 02108 Broadway, NSW 2007 U.S.A. Australia * Corresponding author Email: satish.thosar@uts.edu.au Phone: +612 9514 7761 Fax: +612 9514 7711

The Aftermarket in High-Tech IPOs Abstract We examine the medium-term (six-month) aftermarket in high-tech IPOs launched during the late 1990s. We assume the perspective of an investor who has no preferential allotment and access only to virtually costless information in the public domain. Using an ordered logistic regression approach, we demonstrate the potential to earn marketadjusted returns in excess of 100 percent with an optimal exit strategy. Our model indicates that momentum variables are important while fundamental variables have either no, or at best weak, explanatory power. We discuss our results in light of the minimally rational standard for market rationality articulated by Rubinstein (2000). 1

The Aftermarket in High-Tech IPOs Introduction Most empirical studies relating to IPOs focus on two persistent so-called anomalies: the initial underpricing and the long-run underperformance of IPO firms. These patterns have been documented in various markets and sample periods. 1 The theoretical work has mainly attempted to explain the initial underpricing phenomenon. 2 Our paper sidesteps these issues and examines the medium-term (six-month) aftermarket in high-tech IPOs launched in the late 1990s, arguably a significant hot issue period. We assume the perspective of an investor who has no preferential allotment and access only to easily available and virtually costless information in the public domain. Our objective is to study whether such an investor can earn significant market-adjusted excess returns in an environment characterized by high uncertainty. Our sample is deliberately narrowly drawn. Our high-tech IPO firms fall primarily in the following sectors: computer hardware/software, e-commerce, telecommunications and biotechnology. Clearly, these sectors have huge potential for future growth and profitability but individual firms and their investors face considerable uncertainty about the viability of their technology and/or business models. At one level therefore, our study is about devising and implementing an optimal investment strategy to exploit the turbulent IPO aftermarket. Indeed, using an ordered logistic regression approach, we show that it is possible to earn market-adjusted returns in excess of 100 percent with an appropriate exit strategy. Our model indicates that a long 1 See Loughran et al (1994) for a survey of the international evidence. Also see Jain and Kini (1994), Lee et al (1996) and Loughran and Ritter (1995). 2

position in a high-tech IPO stock entered into at the close of day 1 after the IPO should be cashed out 14 weeks after the IPO date. It appears that momentum variables are important while fundamental variables have either no, or at best weak, explanatory power. In a broader context, we believe that the IPO aftermarket may well be a laboratory for the examination of some of the issues raised in the ongoing debate about whether markets are rational. Rubinstein (2000) in arguing the affirmative case for rational markets concedes that the weight of paper in academic journals supporting anomalies is now much heavier than evidence to the contrary. He points out that while anomalies may exist, there may still not be abnormal profit opportunities in which case markets are at least minimally rational. We discuss our findings in light of Rubinstein s arguments. The next section outlines our data and methodology. The third section describes and discusses our results. The final section contains concluding comments. Data and Methodology In carrying out this study, we assume the perspective of an investor with no preferential allotment in the IPO, access only to freely available information but with some expertise in statistical modeling. Our self-imposed constraint is that the investor should be able to implement her strategy without access to large research departments or subscriptions to expensive databases. 3 2 See Rock (1986), Welch (1989), Grinblatt and Hwang (1989), Benveniste and Spindt (1989) and Loughran and Ritter (2000). 3 It is worth noting that most academic studies ignore potentially large research and information gathering costs in generating and reporting excess returns. 3

Accordingly, our primary sample of high-tech firms was drawn from ipo.com, which lists the universe of U.S. IPOs with dates; offer prices etc. broken down in a number of categories. We chose all IPOs from January 1, 1998 through October 30, 1999 in the following sectors: biotechnology, computer hardware, computer software, electronics, Internet services, Internet software and telecommunications. This resulted in a sample of 316 high-tech IPO firms. Daily closing prices for each firm in the sample and the corresponding NASDAQ index level are downloaded for 125 trading days (approximately six months) beyond the IPO date from yahoo finance. 4 We execute an ordered logistic regression in which the dependent variable, the market adjusted excess return, belongs to one of five categories. The market adjusted return for firm i, where i = 1,2,, N is defined as: R it Pit P i1 PMt P M 1 = 2. Pi 1 PM1 For each firm, P it represents the price t days after the IPO and P i1 is the initial price at the close of day 1 after the IPO. Similarly P Mt and P M1 are the corresponding levels for the market index (Nasdaq). The sample is created so as to make each firm fall into one of the five categories defined in the following table. The dependent variable is captured in terms of the dummy variables Y 1, Y 2, Y 3, Y 4, Y 5 where Yj = 1 if R it exceeds some threshold value; 0 otherwise. For instance, Y 5 = 1 if R it > 1.0 and Y 4 = 1 if Y5 1 and R it > 0.5. Categories Dummy Variables Description Category 1 Y 1 =1, Y 2 =0, Y 3 =0, Market adjusted returns are in the < 0 range Y 4 =0, Y 5 =0 Category 2 Y 1 =0, Y 2 =1, Y 3 =0, Market adjusted returns are in the (0, 0.25] range 4 Price data were spot-checked for validity from alternate sources. 4

Y 4 =0, Y 5 =0 Category 3 Y 1 =0, Y 2 =0, Y 3 =1, Y 4 =0, Y 5 =0 Category 4 Y 1 =0, Y 2 =0, Y 3 =0, Y 4 =1, Y 5 =0 Category 5 Y 1 =0, Y 2 =0, Y 3 =0, Y 4 =0, Y 5 =1 Market adjusted returns are in the (0.25, 0.50] range Market adjusted returns are in the (0.50, 1] range Market adjusted returns are in the >1 range The above categories are defined when the threshold is first reached. At this point, the time variable defined as the number of weeks (one week is represented by five trading days) after the IPO date is also recorded. For instance if the firm's market adjusted return becomes 100 percent in 4 weeks, Y 5 equals 1 and the time variable takes on value 4. For the worst category (Y 1 =1), the time variable equals zero. In many economic applications the dependent variable is discrete and represents an outcome of a decision between a finite set of alternatives. Sometimes there are multinomial choice variables that are naturally ordered (Greene (2000)). Examples include opinion surveys (strongly agree, agree, disagree and strongly disagree), insurance coverage (full, partial, none), bond ratings, etc. In this application, a firm's aftermarket IPO performance falls into one of the five ordered categories defined above. We need a model that explains the influence of variables on the probability of the firm falling into these categories. In the estimation process, although the underlying performance variable (Z) is continuous, only the discrete responses are observed. Consider the following grid that puts firms in the various categories: Y 1 =1 Y 2 =1 Y 3 =1 Y 4 =1 Y 5 =1 γ 0 γ 1 γ 2 γ 3 γ 4 Z P(Y 1 =1) = P(Z < γ 0 ), P(Y 2 =1) = P(γ 0 Z<γ 1 ), etc. For an ordered logit model, 5

( 1 ) 1 PZ ( < γ j ) = 1 + exp( β' X γ ), j where ( 2 ) β X β0 β1x1 β2x2 β k X k = + + + +. The coefficient β j measures the influence of the explanatory variable X j on the probability of falling into a particular category. The γ j s are the unknown parameters to be estimated along with the βs. These probabilities are used to specify the following loglikelihood function that is maximized to obtain the parameter estimates: ( 3 ) = N 5 i= 1 j= 1 ( Yj = 1)ln P( Yj = 1) Further, given a constant term in X, γ 0 is set equal to zero without any loss of generality in the estimation. 5 Potential explanatory variables, X, are selected with guidance from previous literature and researcher intuition, the latter coming into play mostly in constructing the so-called momentum variables. These variables and their sources are as follows: P1i Oi! Underpricing at t=1: O i where P 1i is the closing price at the close of the first day s trading and O i is the offer price. Offer price data were obtained from ipo.com. This represents the extent of initial underpricing (or overpricing). Pi2 P i1 PM2 P M1! Momentum at t=2: 2 Pi 1 PM1 5 Maximum likelihood estimates are obtained using the MAXLIK module of the GAUSS programming language. 6

This represents the market adjusted return on day 2. This is a momentum variable in the sense that it is a purely technical indicator to indicate price direction net of market movement after the first day s trading.! Net Income/Revenue in the pre-ipo year. (Source: ipo.com)! UW Reputation. The lead underwriter s identity was obtained from ipo.com and the reputation proxy is the Carter-Manaster measure reported in Carter et al (1998).! Offer Size. This represents the Offer price * Number of shares sold in the IPO (Source: ipo.com). In the regression, we use the log value of the offer size.! Age. This represents the number of years from the date the firm was incorporated to the IPO date. (Source: FISOnline)! Green Shoe provision. This dummy variable takes value 1 if there is a green shoe provision in the IPO contract, 0 otherwise (Source: ipo.com). Briefly, a green shoe provision gives the underwriter the option to purchase additional shares at the offer price to cover overallotments.! Ind23. This dummy variable takes value 1 if the firm belongs either to the computer hardware or software sectors, 0 otherwise.! Ind56. This dummy variable takes value 1 if the firm belongs either to the Internet services or Internet software sectors, 0 otherwise.! Ind7. This dummy variable takes value 1 if the firm belongs to the telecommunications sector, 0 otherwise. 6 6 No industry dummy variables were set up for the biotechnology or electronic sectors because there were relatively few firms in these sectors in our sample. 7

Table 1 provides a comparison between the distribution of firms across the five categories if the investor always cashes out at the optimal point defined as when the threshold is reached versus the distribution that results if a simple buy and hold strategy is adopted. It should be clear that the optimal distribution is based on perfect hindsight and the buy and hold is a naïve strategy that involves buying every IPO stock at the Day 1 closing price and selling it after six months. Nonetheless, the contrast is striking. With the optimal strategy, 106 (33.5 percent of the total) firms end up in category 5 (market adjusted return exceeding 100 percent) whereas only 47 firms (14.8 percent of the total) achieve the same result under buy and hold. Also, the numbers for category 1 (negative market adjusted return) are 35 (11 percent of total) for the optimal versus 215 (68 percent of total) for the buy and hold. It s obvious that timing the sell decision correctly is of paramount importance and also that time has a non-linear influence. Accordingly, we construct two additional momentum variables before executing the regression. As indicated above, the first variable (Time) represents the time in weeks from the IPO date to the optimal sell date (defined as the date the threshold is reached, not necessarily the peak stock price) and the second variable (Time-Sqd) is simply Time squared to capture the implicit non-linearity. We fully realize that these two variables involve ex-post look back and cannot be used in a pure predictive model. However, as we discuss in greater detail in the results section, we will show that it is possible to simulate the effect of the time variable to predict the distribution of firms across the five categories by evaluating the other non-look back variables. It may appear to the reader that there is a degree of arbitrariness in the way that we construct our market adjusted excess returns or the thresholds that define the ordered 8

categories or even the number of categories. This is quite true but far from being a shortcoming, this arbitrariness is actually an advantage in the context of our application. It allows the individual investor to set her own bar in terms of excess returns while the model itself is flexible enough to accommodate (within reason) any number of categories. Since, there is no a-priori beta type risk measure available for high-tech IPO firms, our measure of excess return (actual return minus twice the Nasdaq return over the comparable period) is in our opinion a reasonably challenging hurdle. Also, the cut points to define the categories are chosen to reflect ambitious but still reasonable thresholds for a speculative investment strategy. Results The results of the ordered logistic regression are reported in Table 2. An interesting but not necessarily surprising finding is the unimportance or comparative weakness in the explanatory power of fundamental variables. For instance, the profitability (or lack thereof) of the firm in the pre-ipo year plays no role in the mediumterm IPO aftermarket. Similarly, the number of years that the firm has been in business prior to the IPO does not seem to matter. These are classic old-economy variables that supposedly enable formation of expectations about future cash flows and/or risk. Our sample is designed to reflect high-tech so-called new-economy IPO firms in an arguably hot-issue period. Therefore, the lack of significance associated with these fundamental variables validates the observation by market watchers that technology sector valuations in the late 1990s did not conform to traditional pricing models. Other variables related to the IPO contract such as the offer size, the presence of a green-shoe 9

provision and even the offer price do not impact the aftermarket. In our sample, the mean of the underpricing variable is 47.3 percent, which seems broadly in line with other recent studies. 7 We were particularly surprised that the extent of initial underpricing (or overpricing) had no carry through effect. However, in a sense this too is a fundamental variable that presumably captures information about the IPO firm implicit in the first day s trading. The only fundamental variables that do appear to have some influence in aftermarket price behavior are underwriter reputation and the industry dummy (Ind56) relating to specifically Internet firms. Both these variables have positive and significant coefficients though the t-statistics are not dramatic in magnitude. 8 It should be noted that the interpretation of the coefficients in an ordered logistic regression is not straightforward. However, in our application, a significantly positive coefficient implies that the variable positively influences the probability of a good outcome, i.e., the probability that the market adjusted excess return will end up in category 4 (50 to 100 percent range) or 5 (>100 percent range). It turns out that momentum variables are basically driving most of the aftermarket action. For instance, the market adjusted return on day 2 (Momentum) is highly significant. According to our analysis, if this variable is positive it strongly favors the probability of the preferred outcomes. However, it is one thing to identify variables that influence the outcome and quite another to devise and implement an appropriate investment strategy. It was intuitively clear to us based on the analysis reported in Table 1 that timing the sell decision was crucial. In other words, in the largely momentum 7 See Loughran and Ritter (2000) and Arosio et al (2000). 8 The original analysis was done using the Carter-Manaster reputation rankings. We found that almost all the lead underwriters in our sample were at the high end (>7). We ultimately substituted the Carter- 10

driven IPO aftermarket, time itself is a key non-linear variable. The results bear this out. Time has a positive impact while Time-Sqd has a negative one and both are highly significant variables with t-statistics of 11.01 and 8.81 respectively. This means that in the initial period after the IPO, it pays to hold the stock for a while because the probability of landing in the higher categories is improving but as more time passes, this probability wanes. In Table 3, we report the actual proportions across categories versus those predicted by our ordered logistic regression model. While recognizing that this is an insample analysis, the predicted proportions are virtually identical to the actual ones; this gives us confidence that the overall model is reasonably well specified. Obviously, an investor cannot employ our model in a traditional predictive sense because the time variable is constructed using ex-post look back. However, it is possible to simulate the probabilities of various outcomes for different values of time. In Table 4, we report the results of the simulation analysis that generated predicted proportions across the five categories with respect to time with other variables at their actual ex-ante values for each firm. The simulated proportions (or probabilities) are averaged across our sample of 301 high-tech IPO firms. For instance, we can see that the probability of ending up in category 5 (market adjusted return > 100 percent) is maximized at 14 weeks after the IPO date. In fact the probability of a category 4 or 5 outcome at 14 weeks is over 90 percent. 9 The same probability at 28 weeks is considerably lower. Manaster number with a dummy variable which takes value 1 if the underwriter ranking was greater than 7, 0 otherwise. 9 It is important to note that the probabilities reported in Table 3 are conditional on the time variable taking on a positive value. As indicated in the methodology section, the value of the time variable is determined when the ordered threshold is crossed but takes on a value of zero if the firm never does cross any positive market adjusted return threshold, i.e., ends up being a category 1 firm. Therefore, the probabilities in Table 11

It appears that an investor with access to the kind of virtually costless and easily accessible information we use in this study can buy a high-tech IPO stock at the close of the first day s trading and earn market adjusted returns in excess of 100 percent with high probability if she employs the appropriate exit strategy for our sample a cash-out point of approximately 14 weeks after the IPO date. So the broad conclusion appears to be that a purely technical trading strategy is viable and potentially extremely profitable in the medium term IPO aftermarket. Does this mean that this particularly segment of the market is not even minimally rational as defined by Rubinstein (2000)? Do our results imply abnormal profit opportunities? While suggestive of such a conclusion, there are many caveats. First, our unorthodox (and somewhat arbitrary) approach to generating market adjusted excess returns may be inadequate, though no obvious alternative suggests itself. Second, and perhaps more importantly, our study could be classified as an exercise in data mining, even though we have relied on virtually costless data sources and certainly a rather narrow sample size and period. We have no evidence that our findings can be replicated outside our sample period or in other hot issue markets. Therefore, we remain agnostic as to the implications of our findings for the rational markets debate. We do believe, however, that the market setting examined by us could serve as a laboratory for the investigation of a possible source of market uncertainty. Specifically, Rubinstein (2000) proposes that there may be a rational explanation for the excess volatility anomaly: much of the volatility in prices derives from changes in beliefs about the demand curves of other investors, a form of endogenous uncertainty this 3 are intended to be illustrative and not definitive. However, an independent mean-reversion analysis also indicated that the mean reversion effect is strongest at around 14 weeks. 12

may also explain that while stock prices typically react to news about fundamentals, they also seem to change when there is no news. An IPO aftermarket may be the ideal arena to generate and test hypotheses in this emerging strand of literature. Finally, we would like to propose that the ordered logistic regression approach adopted by us is a potentially powerful methodological tool in the study or practice of financial economics. As alluded to above, a number of responses or outcomes that constitute dependent variables are either already naturally ordered or can be modeled in that manner. In particular, this approach could be utilized in the formulation of trading strategies involving technical indicators. Conclusion In this study, we examine the medium-term (six-month) aftermarket in high-tech IPOs launched during the late 1990s. We assume the perspective of an investor who has no preferential allotment and access only to virtually costless information in the public domain. Using an ordered logistic regression approach, we demonstrate the potential to earn market-adjusted returns in excess of 100 percent with an optimal exit strategy, which essentially involves cashing out 14 weeks after the IPO date. Our model indicates that momentum variables are important while fundamental variables have virtually no explanatory power. We discuss our results in light of the minimally rational standard for market rationality articulated by Rubinstein (2000) and suggest that the IPO aftermarket is an ideal setting for studying a possible source of excess volatility: endogenous uncertainty arising from changes in investor beliefs about other investors demand curves. 13

Table 1 Distribution of IPO firms across categories based on Naïve versus Optimal Timing Strategies. The Naïve strategy involves buying at the end of first day s trading (t=1) price and selling exactly six months later. The Optimal strategy assumes perfect hindsight and requires buying at the end of first day s trading price and selling at the point the threshold (not necessarily price peak) is reached. Categories Number of Firms (Naïve Strategy) Number of Firms (Optimal Strategy) Category 1 215 35 Category 2 20 81 Category 3 10 44 Category 4 24 50 Category 5 47 106 Definition of Categories Categories Category 1 Category 2 Category 3 Category 4 Category 5 Description Market adjusted returns in the < 0 range Market adjusted returns in the (0, 0.25] range Market adjusted returns in the (0.25, 0.50] range Market adjusted returns in the (0.50, 1] range Market adjusted returns in the >1 range 14

Table 2 Results of Ordered Logistic Regression; Sample Size 301 (original sample is 316; 15 firms are dropped due to incomplete information on one or more explanatory variable) The dependent variable is captured in terms of the dummy variables Y 1, Y 2, Y 3, Y 4, Y 5 where Yj = 1 if the market adjusted return exceeds some threshold value; 0 otherwise. Categories Dummy Variables Description Category 1 Y 1 =1, Y 2 =0, Y 3 =0, Market adjusted returns are in the < 0 range Y 4 =0, Y 5 =0 Category 2 Y 1 =0, Y 2 =1, Y 3 =0, Market adjusted returns are in the (0, 0.25] range Y 4 =0, Y 5 =0 Category 3 Y 1 =0, Y 2 =0, Y 3 =1, Market adjusted returns are in the (0.25, 0.50] range Y 4 =0, Y 5 =0 Category 4 Y 1 =0, Y 2 =0, Y 3 =0, Market adjusted returns are in the (0.50, 1] range Y 4 =1, Y 5 =0 Category 5 Y 1 =0, Y 2 =0, Y 3 =0, Y 4 =0, Y 5 =1 Market adjusted returns are in the >1 range Model Estimates Parameters Mean of the Variable Estimates t-statistics Constant 1.000 2.2268 1.059 Time 6.781 0.7975* 11.074 Time-Sqd 45.978-0.0270* -8.808 Underpricing 47.310-0.0028-1.570 Momentum -0.738 0.0816* 8.879 Net Income/Revenue -1.767-0.0353-0.841 UW Reputation 0.801 0.6813* 2.133 Offer Size 17.943-0.1268-1.008 Green Shoe Dummy 0.575 0.1805 0.692 Ind23 Dummy 0.166 0.2254 0.604 Ind56 Dummy 0.605 0.8060* 2.202 Ind7 Dummy 0.179 0.1415 0.357 Age 5.327-0.0003-0.017 * Indicates significance at 5 percent or lower level in a two-tailed test. 15

Table 3 Actual versus Estimated Average Proportions Categories Actual Proportions Predicted Proportions C1 (<0) 0.116 0.113 C2 (0,0.25] 0.246 0.271 C3 (0.25,0.5] 0.136 0.142 C4 (0.5,1] 0.159 0.135 C5 (>1) 0.342 0.338 16

Table 4 Results of a simulation exercise with respect to the time variable where all other variables are evaluated based on their ex-ante actual values. Proportions for each category simulated for each firm and then averaged across the sample of 301 firms. Time in Weeks C1 C2 C3 C4 C5 2 0.105 0.447 0.224 0.125 0.099 4 0.034 0.262 0.251 0.215 0.238 6 0.013 0.130 0.188 0.240 0.428 8 0.006 0.066 0.121 0.204 0.602 10 0.003 0.038 0.079 0.159 0.720 12 0.002 0.026 0.058 0.128 0.786 14 0.002 0.022 0.050 0.114 0.812 16 0.002 0.023 0.051 0.116 0.809 18 0.002 0.028 0.062 0.134 0.774 20 0.003 0.043 0.088 0.169 0.696 22 0.007 0.078 0.136 0.216 0.564 24 0.016 0.156 0.207 0.241 0.380 Definition of Categories Categories Category 1 Category 2 Category 3 Category 4 Category 5 Description Market adjusted returns in the < 0 range Market adjusted returns in the (0, 0.25] range Market adjusted returns in the (0.25, 0.50] range Market adjusted returns in the (0.50, 1] range Market adjusted returns in the >1 range 17

References Arosio, R., Giudici, G., and Paleari, S., Why Do (or Did?) Internet-Stock IPOs Leave So much Money on the Table?, Working Paper, December 2000. Benveniste, L. M. and Spindt, P. A., How Investment Bankers Determine the Offer Price and Allocation of New Issues, Journal of Financial Economics, 24(2), 1989, 343-361. Carter, R. B., Dark, F. H., and Singh, A. K., Underwriter Reputation, Initial Returns and the Long-Run Performance of IPO Stocks, Journal of Finance, 53(1), 1998, 285-311. Greene, W. H., Econometric Analysis, 4th Edition, 2000, Prentice-Hall Inc. Grinblatt, M. and Hwang, C. Y., Signaling and the Pricing of New Issues, Journal of Finance, 44(2), 1989, 393-420. Jain, B. A. and Kini, O., The Post-Issue Operating Performance of IPO Firms, Journal of Finance, 1994, 1699-1726. Lee, P. J., Taylor, S. L., and Walter, S.W., Australian IPO pricing in the short and long run, Journal of Banking and Finance, 20, 1996, 1189-1210. Loughran, T., Ritter, J.R., and Rydqvist, K., Initial Public Offerings: International Insights, Pacific-Basin Finance Journal, 2, 1994, 165-199. Loughran, T., and Ritter, J.R., The new issues puzzle, Journal of Finance, 50, 1995, 23-51. Loughran, T., and Ritter, J.R., Why Don t Issuers Get Upset About Leaving Money on the Table in IPOs?, Working Paper, August 2000. Rock, K., Why New Issues Are Underpriced, Journal of Financial Economics, 15, 1986, 187-212. Rubinstein, M., Rational Markets: Yes or No? The Affirmative Case, forthcoming in Financial Analysts Journal. Welch, I., Seasoned Offerings, Imitation Costs, and the Underpricing of Initial Public Offerings, Journal of Finance, 44, 1989, 421-449. 18