新闻传播学相关参考资料合集：在互联网上媒体扮演的角色

THE ROLE OF THE MEDIA IN THE INTERNET IPO BUBBLE* Utpal Bhattacharya Kelley School of Business Indiana University ubhattac@indiana.edu Neal Galpin Kelley School of Business Indiana University ngalpin@indiana.edu Rina Ray Kelley School of Business Indiana University riray@indiana.edu Xiaoyun Yu Carlson School of Management University of Minnesota xyu@csom.umn.edu September 15, 2005 Key words: initial public offerings, media, internet bubble JEL number: E32, G14, G30 * We are grateful for suggestions from Laura Field, Murray Frank and seminar participants at Penn State University, Singapore Management University, University of New Orleans and at the following conferences: European Finance Association Annual Meeting 2005 and China International Finance Conference 2005. All mistakes remain our own responsibility. THE ROLE OF THE MEDIA IN THE INTERNET IPO BUBBLE Abstract The first part of this paper explores whether aggregate media coverage was different for internet IPOs as opposed to a matching sample of non-internet IPOs in the late 1990s. So we read all news items that came out between 1996 through 2000 on 458 internet IPOs and a matching sample of 458 non-internet IPOs – a total of 171,488 news items – and classify each news item as good news, neutral news or bad news. We find, not surprisingly, that the overall media coverage was more intense for internet IPOs. Further, we document that the overall media coverage was much more positive for internet IPOs in the period of dramatic rise in share prices and was much more negative for internet IPOs in the period of dramatic fall in share prices. The second part of the paper explores whether this differential media coverage had any effect on the difference in risk-adjusted returns between internet stocks and non-internet stocks. We find, surprisingly, that the answer is no: the market somewhat discounted the media sentiment – though net news (number of good news minus number of bad news) Granger caused risk-adjusted returns for both samples of IPOs, the effect was lower for internet IPOs, especially in the bubble period. So our conclusion is that the media was not a significant factor in the dramatic rise and fall of internet shares in the late 1990s. 1 THE ROLE OF THE MEDIA IN THE INTERNET IPO BUBBLE I. INTRODUCTION The offer price of eToys’ initial public offering (IPO) on May 20, 1999 was $20. It shot up to $76.5625 at the end of the first trading day. Later, in a special report, Wall Street Journal’s technology editor Jason Fry noticed that “driving the company’s success is a cleanly designed, easy-to-use site designed to soothe adults rattled by the pitfall of toy shopping in the real world”. He wrote: “…for many an overeager Web outfit, they [eToys] have proved hard to duplicate. The company is fanatical about providing top-rate customer service, including actual human beings who answer the telephone. It didn’t open for business until founder and Chief Executive Officer Toby Lenk believed he understood the toy industry and how it should work online. And eToys is dedicated to moving at Web speed. It has stayed ahead of its deep-pocketed competitors by constantly transforming itself, for example, by diversifying its inventory to include books, music, videos and baby products.” By the end of 2000, the firm traded at $ 0.1875 per share, which was 0.24% of its first day closing price, and 0.94% of its offer price. Today, eToys does not trade. The company went bankrupt and its stock was de-listed from Nasdaq on February 26, 2001. In this paper, we use the phrase “bubble/post-bubble” to reflect the dramatic rise and dramatic fall of stock prices in the period of 1996-2000. We ask and answer the following two questions: was the overall media coverage for internet IPOs in the years 1996 through 2000 different from a matching sample of non-internet IPOs and, if yes, did this difference in the media coverage have any effect on the difference in risk-adjusted returns between internet stocks and non-internet stocks. 1 There are many reasons for believing that the media coverage and/or its effect were more pronounced for internet IPOs than for non-internet IPOs in the late 1990s. First, the media were more 1 In this paper we look at stock returns after the first day of trading. In a companion paper, we look at stock returns at the first day of trading. The reason we separated our analysis is because the information dissemination during the pre- IPO book-building stage is very different from the information dissemination during the post-IPO stage. In the pre- IPO stage, institutions disseminate information and, therefore, the natures of these institutions are the significant control variables. A rich literature exists on what affects the first day’s return. In the post-IPO stage, on the other hand, the main source of information is the traded price itself. This, therefore, becomes the paramount control variable in this paper. 2 likely to be interested in internet IPOs, because in this period there were so many of them, and many of them had dramatic first-day returns (Benveniste, Ljungqvist, Wilhelm and Yu 2003, Ljungqvist and Wilhelm 2003 and Ritter and Welch 2003). Second, as the internet industry was new, there was no history of cash flows of comparable firms that had gone public. This made valuations difficult, and so expectations of future cash flows for internet IPOs were more likely to be sensitive to media news (Blanchard and Watson 1982; Hirota and Sunder 2002 provide experimental evidence.) Third, the limits to arbitrage were more binding for internet IPOs during the period of 1996-2000 (Lamont and Thaler 2003 and Ofek and Richardson 2003), so the divergence of stock prices from their fundamental value was likely to be greater for internet IPOs. Further, institutional investors did not attempt to trade against market movements, but actively rode with both the run-up and run-down of the stock (Griffin, Harris and Topaloglu 2003 and Brunnermeier and Nagel 2004). Fourth, and finally, there is now increasing evidence that the spectacular rise and spectacular fall of internet IPOs in the late 1990s can not be reconciled with fundamentals (Cooper, Dimitrov and Rau 2001, Ofek and Richardson 2002, Loughran and Ritter 2004). 2 A good place to begin looking for other explanations is to formally explore the economic role of the media in this period. 3 Was the overall media coverage different for internet IPOs? We read all news items that came out between 1996 through 2000 on 458 internet IPOs and a matching sample of 458 non-internet IPOs – a total of 171,488 news items – and classify each news item as good news, neutral news or bad news. We find, not surprisingly, that the media coverage was more intense for internet IPOs. All types of news – good, bad, or neutral – were more for internet IPOs than for non-internet IPOs in both the bubble period and the post-bubble period. Second, we use net news (good news minus bad news) to proxy media sentiment. 4 We find that compared to the matching sample, the net news was more positive for internet IPOs in the bubble 2 Cutler, Poterba and Summers (1989) is one of the early papers that show stock return variances cannot be caused by fundamental news such as macroeconomic news. 3 Shiller (2000) believes that the stock price increases in the late 1990s were driven by irrational euphoria among individual investors, fed by an emphatic media, which maximized TV ratings and catered to investor demand for pseudo-news. Professional investors “are not immune from the effects of the popular investing culture that we observe in individual investors” (p.18). 4 The issue – whether media sentiment reflects public sentiment or is different from public sentiment – though interesting in its own right, is beyond the scope of this paper. 3 period, and more negative for internet IPOs in the post-bubble period. Third, we document that net news increased after a positive stock return, and decreased after a negative stock return for internet firms. This provides some evidence in favor of Shiller’s (2000) positive feedback hypothesis. Interestingly, the positive feedback was asymmetric. During the bubble period, net news increases regardless of whether the previous stock return is positive or negative. However, during this bubble period, the increase after a positive stock return was larger for internet IPOs than for non-internet IPOs. These findings reverse in the post-bubble period. In the post-bubble period, net news decreases regardless of whether the previous stock return is positive or negative; however, the decrease in net news after a negative stock return was larger for internet IPOs than for non-internet IPOs. Our results are robust to whether we follow the traditional definition of the bubble in calendar time (the period that ended March 24, 2000) or in event time (the firm- specific period that ended on the day the firm’s stock price peaked). We, therefore, draw the following conclusion about media coverage of IPOs in the late 1990s: it seems that the overall media coverage was more positive about internet IPOs in the bubble period and more negative about internet IPOs in the post- bubble period. Did the differential media coverage have any effect on the difference in risk-adjusted returns between internet stocks and non-internet stocks during this period? We check whether news in the media, measured by numbers and type, Granger caused abnormal returns, where abnormal returns is the error term of a Fama-French (1993) three factor model. We incorporate the previous period’s abnormal returns as a control when we examined the effect of the media in our analysis. This buys us a few advantages. First, and most important, it corrects for feedback effects: does the media affect returns or do returns affect the media. Both are possible. As a matter of fact, our analysis on aggregate media coverage showed that the latter effect exists. Second, it corrects for possible momentum or reversal in daily stock returns. Third, as the last period’s abnormal return contains the cumulative effect of past media reports, it corrects for the possibility that the media effect lasts longer than a day. Finally, because lagged return occurs on the same day as the news reports used for tests, lagged return controls for contemporaneous relationships between stock prices and news coverage. We also control for firm-fixed effects to mitigate the variations in news 4 coverage related to the specific nature of the firm and to remove the author bias effect from news classification, if we assume this effect is firm-specific. After controlling for previous abnormal returns, firm fixed effects, market trading conditions and news volumes, we find, not surprisingly, that good news increases risk-adjusted returns the next period, and bad news decreases risk-adjusted returns the next period, and so net news (good news minus bad news) increases risk-adjusted returns the next period. We find, surprisingly, that the effect of net news on next period’s risk-adjusted return was lower for internet IPOs, especially during the bubble period. In addition, during the post-bubble period, the effect of good news matters more on next period’s risk-adjusted return for internet IPOs than for non-internet IPOs. Our results are robust to whether we risk-adjust individual stocks, or whether we risk-adjust a portfolio consisting of either internet or non-internet stocks. We, therefore, draw the following conclusion: though the media coverage was much more positive about internet IPOs in the bubble period and was much more negative about internet IPOs in the post-bubble period, the market somewhat discounted the media sentiment, especially during the bubble period. Although there could be many other rational or irrational factors that contributed to the dramatic rise and fall of the stock market in the period 1996 through 2000, our results suggest that media sentiment, being discounted by the market, was not a major determinant of the bubble and its crash. 5 Our paper is organized as follows. In Section II, we discuss the related literature on media and its relation with changes in stock prices. Section III discusses how we obtained our data. Section IV gives our results on differential media coverage of internet IPOs as opposed to a matching sample of non-internet IPOs. Section V answers whether the differential media coverage affected the difference in risk-adjusted returns between the two samples. Section VI covers various tests for robustness that we conducted, 5 Though we are tempted, we want to be cautious to draw definitive conclusions about market efficiency from our findings. This is because of the following reason. Differences in net media news (good news minus bad news) will positively affect differences in net returns between internet firms and non-internet firms if media news reflect fundamentals and the market is efficient with respect to media coverage (i.e. only media news about fundamentals moves prices), or, if media news reflect sentiment and the market is inefficient with respect to media coverage (i.e. only media news about sentiment moves prices). On the other hand, differences in net media news will not affect differences in net returns between internet firms and non-internet firms if media news reflect sentiment and the market is efficient with respect to media news, or, if media news reflect fundamentals and the market is inefficient with respect to media news. 5 including an experiment to check the validity of our technique of classifying news. We conclude in Section VII. II. THE RELATED LITERATURE The first question of this paper – was the overall media coverage for internet IPOs different from the overall media coverage of non-internet IPOs – belongs to a growing literature on bias in the financial media. How do the financial media choose which stories to cover? Of the stories they choose to cover, what is the slant given? And why is there a slant? Shiller (2000) writes: “The role of the news media in the stock market is not, as commonly believed, simply as a convenient tool for investors who are reacting directly to the economically significant news itself. The media actively shape public attention and categories of thought, and they create the environment within which the stock market events we see are played out.” He believes that the financial media strive to enhance interest by attaching news stories to stock price movements that the public has already observed, thereby creating a positive feedback effect. Dyck and Zingales (2003a) note that there is a pro-company bias in the financial media, which is stronger during a boom, and is weaker and is sometimes reversed during a bust. They argue that this is because of incentives. Reporting good news during booms allows media access to the company, but this access is not important during busts because the company does not want to share news. Dyck and Zingales (2003b) find empirical support in that media spin affects the stock market response to earnings announcements. Mullainathan and Shleifer (2003) demonstrate that the media can slant the presentation of the news to cater to the preferences of their audience. Baron (2004) explains why persistent media bias can exist in a competitive equilibrium; in his hypothesis, bias originates with journalists who have a preference for influence and are willing to sacrifice wages to exercise it. Our second research question – did the differential media coverage have any effect on the difference in risk-adjusted returns between the two samples – extends from a large literature on how media news affects returns. According to classical asset pricing models, news will affect returns if it affects expectations of future cash flows and/or expectations of the discount rate. By filtering, aggregating and 6 repackaging information into news items, the media reduce the cost of collecting and certifying relevant information, and therefore can have significant impact on financial markets. In an early paper, Niederhoffer (1971) observes large price movements following world event headlines; the market appears to overreact to bad news. Mitchell and Mulherin (1994) document a weak relationship between the amount of publicly reported information, approximated by the number of daily Dow Jones news stories, and the aggregate trading activity and the price movements in securities markets. Antweiler and Frank (2004a) find statistically significant return momentum for many days after the news is made public and a bigger and more prolonged impact of average news on returns during a recession than during an expansion. Chan (2003) shows stocks with large price movements, but no identifiable news, show reversal in the next month, and prices are slow to reflect bad public news. Alternatively, Merton (1987) argued that investors will buy and hold only those securities which they are aware of. The most common way to facilitate investors’ awareness is to promote the visibility of the firm through media. Falkenstein (1996) documents that mutual funds avoid stocks with low media exposure. Barber and Odean (2003) provide direct evidence that individual investors tend to buy stocks that are in the news. Antweiler and Frank (2004b) and Wysocki (1999) find that the volume of stock messages posted on internet stock message boards predicts subsequent stock returns and market volatility. Tetlock (2003) provides evidence that media coverage affects market index returns and aggregate trading volume. Huberman and Regev (2001) document that old news repackaged as new news can also affect returns. Antunovich and Sarkar (2003) find that stocks with higher media exposure have bigger liquidity gains and lower excess returns on the pick day. Chen, Noronha, and Singal (2002) show that media exposure increases following additions to the S&P 500 index, and price changes around S&P 500 index additions are consistent with greater investor awareness of the added stocks. A literature focusing on the relation between media and IPO firms has already started to emerge. Examining the post-offer performance of a sample of IPOs, Loughran and Marietta-Westberg (2002) find that investors over-react to positive-return news events and under-react to negative news events. Johnson and Marietta-Westberg (2004) show that the increase in idiosyncratic volatility for IPO firms over time is 7 Granger-caused by the increase in news in recent decades. The extent of pre-IPO media exposure is found to be positively related to IPO underpricing both in US (Reese 1998 and Ducharme, Rajgopal and Sefcik 2001a) and in other countries (Ho, Taher, Lee and Fargher 2001). The pre-IPO media hype is related to the IPO’s short-term and long-term volume (Reese 1998) and to their post-offer return performances (Ducharme, Rajgopal and Sefcik 2001b). Initial return is shown to have a positive influence on the subsequent media coverage (Demers and Lewellen 2003), suggesting IPO underpricing publicizes stocks to investors who buy the stock in the after-market. Our paper differs from this literature in that we do not focus on what drives the media coverage of IPO firms. Nor do we focus on the pre-IPO stage or on the first day’s return. Instead, we take the bubble period as given and follow IPOs over their rise and fall in the period 1996 through 2000. Second, unlike most of the above papers, we look at both the numbers of news items and their type (good, bad or neutral) from all media sources to capture the aggregate media effect. III. DATA A. The IPO sample We start with a large sample of firms that went public between January 1996 and December 2000. After excluding unit offers, rights offers, closed-end mutual funds, REITs, and ADRs, our search of the Thomson Financial’s SDC database yielded 2,603 completed issues. We identify and extract 461 internet companies from this sample using the reference list from Loughran and Ritter (2004). We cross-check our internet IPO issues with Loughran and Ritter (2002) and Ljungqvist and Wilhelm (2003) to correct errors in the SDC data. We remove one issue that went public twice and was, therefore, counted twice during our sample period. That leaves us with 459 internet IPOs. For the remaining 2,142 issues in this SDC sample, we first manually check for misclassification, and exclude 9 issues which are in fact ADRs, 1 belonging to unit trusts, 2 misclassified as IPOs, 2 without filing, offer or trading price information in SEC, news sources and CRSP, and 1 foreign offer with a minor tranche in the US. Then, we extract a matching set of non-internet IPOs from the rest of the 2,127 issues based on offer size and offer date as follows: for each of the internet IPOs, we impose a 20% band on its 8 offer size, and choose the matching firm with the closest offer date among candidates. Matches are formed without replacement. 6 So we have a matching sample of 459 non-internet IPOs. Since we study the effect of media on returns of IPO firms during the boom and bust of the internet bubble period, we expect each of our sample firms to have some degree of news coverage. There is one firm in our non-internet sample where we cannot identify any news report from Factiva using various combination of search. We therefore removed this firm from our analysis. Excluding or including this firm in our sample does not change our results. Our final sample contains 458 internet IPOs and a matching 458 non-internet IPOs. Offer characteristics such as offer size, venture-capital backing, and the stock exchange in which the IPO first traded, are from SDC. Stock prices and daily returns are from CRSP. Fama-French factors are obtained from French’s website. We manually collect missing founding date for 193 issues within the non- internet sample and 222 issues within the internet sample from SEC filing prospectuses, subsequent 10-Ks, or news sources. B. The news sample We define the media to be the Dow Jones Interactive Publications Library (DJI) of past newspapers, periodicals, and newswires. After DJI’s conversion to Factiva in June, 2003, we create a customized list that includes major news and business publication sources worldwide. 7 This list is consistent with the news sources in DJI prior to its conversion. We choose Dow Jones Interactive and Factiva because they provide by far the most complete sources of media coverage across time and stocks. As pointed out by Chan (2003), this source does not suffer from gaps in coverage, and is the best approximation of public news for general investors. We do not include magazines, since it is difficult for us to pin down precisely when the information is publicly available. We also exclude investment newsletters, analyst reports and other sources 6 We use the matched firms in two ways. For the main analysis, we ignore the individual matches, drawing conclusions based only on whether internet firms differ from non-internet firms. For robustness (see Section VI, C2), we consider the individual matches directly, drawing conclusions based on whether internet firms differ from their particular match. 7 The resulting list of data sources includes Dow Jones Asia, Europe, Africa, North America, South America, Australia and New Zealand and contains all the English language sources of daily news. 9 that are not available to the general audience. 8 There are more sources in Factiva towards the end of our sample period. However, the difference will not be crucial to our results, as our sample period is relatively short, and all the econometric analyses are benchmarked with the non-internet sample during the same period. For each IPO in our sample, we “search by name” in Factiva for the period between 90 days prior to the public offer and the end of December, 2000. Since the book building period of most IPOs in our sample lasts 13 weeks, this includes the majority of media coverage during the pre-IPO stage for each firm. Occasionally, we collect news reports from Factiva using “search by keyword” instead of “search by name.” 9 This occurs mainly for firms involved in mergers during or after our sample period (Factiva drops all indexing after a merger, even if it just happened this year). 10 We hand-collect all the news articles in which the IPO firm was mentioned. In particular, we do not limit our news articles only to those news items where the firm is only mentioned in the headline or in the lead paragraph, because this could potentially exclude a large volume of news reports that actually cover the firm. There are a total of 171,488 news items. Two coauthors start news coding in the fall of 2002 and the entire sample is completed at the end of 2004. We classify each news item into one of three categories: “good”, “bad” or “neutral”. Good news items (bad news items) are defined as news items which carry positive (negative) statements or implications about the firm. Neutral news items are news items that cannot be classified as good or bad. We do not classify news based on previous returns as in Chan (2003), because doing so automatically assumes the signed direction of causality from returns to news. 8 So we also could not include the following: Factiva Aviation Insurance Digest, Factiva Marine Insurance Digest, Dow Jones Emerging Market Reports, Dow Jones Commodities Service, Dow Jones Money Management Alert, and Dow Jones Professional Investor Report. 9 There is a subtle distinction between “search by keyword” and “search by name” of an IPO firm in Factiva. When “search by keyword” is used, Factiva returns virtually all news articles that at least mentioned the name of the firm once, which could be noisy. On the other hand, news articles generated by “search by name” are more related to the firm, and therefore more focused. 10 Prior to late March of 2004, Factiva re-indexed all the previous news reports about a target firm involved in a merger to the acquirer. A search by a firm's name in this case only returns news items where both the target and the acquirer are reported. After late March 2004, this particular situation was solved as Factiva introduced an updated version of its database. 10 There are two ways of classifying news items: mechanically using Content Analysis software or using human judgment. The advantage of the former method over the latter method is that it is less expensive, it is faster, and it is consistent. The disadvantage of the former method over the latter method is that it is prone to making serious mistakes. If software is programmed to classify a news item as good if it detects a number of positive words, it is bound to misclassify a news item as good if the news contains many good words about the competition, and few bad words about the firm. The software may also misclassify a news item as neutral if there are no obvious positive or negative words in the article, whereas a human will judge correctly from the context that the article is good news or bad news about a firm. So we chose human judgment for classification. We read each of the 171,488 news items individually, and classify each news item as either “good”, “bad” or “neutral” using our judgment. Our judgment is based on the content of each individual news item, without forming a new expectation after each piece of news. This method of human judgment has obvious drawbacks, the most important of which is lack of consistency. To reduce possible time-varying judgment errors, we have one author start from the last firm that went public in the internet IPO sample and read the news in reverse chronological order, while the second author starts from the first firm that went public in the non-internet IPO sample and read the news in chronological order. Later on, when we conduct the regression analysis, we difference news variables to remove the firm effect. This also removes the author bias effect, if we assume that the author’s bias is firm-specific. However, even with the above two approaches, we could still face possible judgment error as the same piece of news may be categorized differently by different human beings over time. So we conducted an experiment to verify the consistency of our judgments. Though we will delay the discussion of this experiment to the “robustness tests” section of this paper, it should be pointed out here that we did exhibit consistency in our judgment. Correlations between our classification of news items and the classification of the same news items made by other participants in the experiment were strongly positive. We define the degree of media coverage as the number of news items about a sample IPO firm during a specific period. For the pre-IPO period (up to the offer day), news items are counted and 11 classified for the whole period, as there is no price information during this period. For the post-IPO period, news items are classified and counted on a daily basis. For any given day, we aggregate news items about the same firm from multiple media sources and do not distinguish between “real news” and “spin-news”. This research design is created with the intent to investigate the impact of the intensity of the media coverage, and is based on the fact that different types of media may reach different types of investors. In addition, the criterion for estimating the influence of individual media is ambiguous, and very often the same contents will be covered by various media sources. 11 C. Summary IPO Statistics Table 1 reports summary statistics of offer characteristics obtained from SDC and CRSP, broken down by internet and non-internet IPOs. As the internet industry was new, the internet firms, not surprisingly, are younger. The average IPO is over 9 years old at the time of its offering for the non-internet sample, but is less than 5 years old for the internet sample. However, the difference in age between the two samples drops substantially when we examine the median instead of the mean firm age. During the book- building period (from the registration date to the offer date), the average expected offer price, reflected in the mean of the indicative price range included in the issuer’s S-1 filing, is significantly higher for the non- internet sample ($13.24 compared to $12.14 for the internet sample). This is in sharp contrast with the final offer price, where the internet issues on average are set a higher price ($14.76 versus $13.67 for the non- internet sample). Accordingly, in terms of price revisions, measured as the percentage change between the final offer price and the expected offer price, the average internet firm revises its price much more than the average non-internet firms, 23% versus 4.15%, and this difference is highly significant. The most distinguishing feature of the sample is the first-day returns, calculated as the percentage change between the final offer price and the first-day closing price, which we take from CRSP tapes if 11 Surprisingly, when we disaggregated the media into the top ten by circulation (Associated Press News Wire, Chicago Tribune, Daily News, NY, Dow Jones News Service, Houston Chronicle, LA Times, Reuters News, New York Times, Wall Street Journal, USA Today) and the wire services only (Associated Press News Wire, Dow Jones News Service, and Reuters News), we found that the former covered internet firms with more statistically significant intensity (the top ten media sources account for 58.74% of the media coverage for internet firms and only 55.66% of the media coverage for non-internet firms), but the latter showed no statistically differential preferences (53.28% for internet firms vs. 52.17% for non-internet firms). 12 available with seven days of the offer date (as in Lowry and Schwert 2002.) The internet firms averaged a stunning 83.72% first-day return during our sample period, which is similar to the 89% first-day return documented by Ljungqvist and Wilhelm (2003) for internet IPOs during 1999 and 2000. This first-day return is more than twice in size compared to the non-internet sample (41%). Surprisingly, the two samples do share many similarities. Because of our method of construction, the average gross proceeds are around $88 million for both samples. The width of the filing price range, defined as the difference between the high and low prices suggested in the preliminary prospectus and often viewed in the IPO literature as a proxy for ex ante uncertainty about a firm’s value, is virtually the same between the two samples. About sixty-seven percent of internet firms operate in what we refer to as “high-tech” industries (three-digit SIC codes 283, 357, 366, 367, 381, 382, 383, 384, 737, 873, and 874). This definition follows Benveniste, Ljungqvist, Wilhelm and Yu (2003) and “Hi Tech Industry Group” defined by SDC, and covers industries such as pharmaceuticals, computing, computer equipment, electronics, medical and measurement equipment, software and biotech industries. Interestingly, sixty-three percent of non-internet firms belong to the “high-tech” category as well, and the difference between the two samples is not significant either economically or statistically. This reflects the “high-tech” industry clustering in the sample period of 1996-2000. Correlated with this feature, most of the firms in the two samples trade on Nasdaq. IV. MEDIA COVERAGE A necessary condition for the media to have an economic role in the internet IPO bubble is that the overall media coverage for internet IPOs is different from that of non-internet IPOs. We investigate this issue in the current section. We explore this question for the entire sample period of 1996-2000, as well as for two sub-periods: before and after the price peaked. The first peak definition is a market-wide definition. On March 24th, 2000, the Nasdaq 100 index reached its highest point in our 1996 to 2000 sample period. We follow the 13 traditional literature and take this date as given to be the market peak. The second peak definition is intended to capture shifts in individual stocks, and is calculated as the date at which the firm’s market capitalization reaches the highest point in the sample period. The sub-periods defined using the first definition of a peak (March 24, 2000) are, therefore, in calendar time, whereas the sub-periods defined using the second definition of a peak (firm peak) are in event time. Throughout this paper, we use before the peak and bubble period interchangeably, and after the peak and post-bubble period interchangeably. A. Unconditional Media Coverage First, we examine the unconditional media coverage of the internet sample and the non-internet sample, without taking into account the impact of previous price movements on this coverage. Figures 1-a through 1-e and Figures 2-a through 2-e provide a visual presentation of these patterns over various periods of time. In Figures 1-a through 1-e, news items for the two samples are aggregated over time and firms, while in Figures 2-a through 2-e, the news items are per day per firm. Figures 1-a and 2-a cover the entire sample period. Figures 1-b (1-c) and figures 2-b (2-c) report the degree of media coverage before (after) the peak, where the peak is the day in which the firm’s stock price peaked. Figures 1-d (1-e) and figures 2-d (2-e) report the degree of media coverage before (after) the peak, where the peak is March 24, 2000. Compared to the non-internet sample, the internet sample had significantly higher media coverage in all the three measures (total number of news, good news, and bad news), during all the three time periods (entire sample period, before and after the peak, however you define the peak). This was true in the aggregate, as well as per firm per day basis. Next, we use net news, defined as the difference between the number of good news and bad news, to proxy media sentiment. Media sentiment, regardless whether it reflects public opinion or is different from public opinion, is considered optimistic if net news is positive and is considered pessimistic if net news is negative. Interestingly, as shown in Figures 1-b, 1-d, 2-b and 2-d, during the bubble period, internet IPO firms have more positive net news than their matching sample. This indicates that media reported relatively more good news than bad news for the internet firms than they did for the non-internet 14 firms in the bubble period, suggesting media generally not only provided more coverage but also had a more optimistic view, whether rational or not, about internet firms in the bubble period. However, Figures 1-c, 1-e, 2-c and 2-e reveal that post peak, there was a dramatic shift in media sentiment. Internet IPO firms have more negative net news than their matching sample after the bubble burst. This indicates that media reported relatively more bad news than good news for the internet firms than they did for the control firms in the post-bubble period, suggesting media had a more pessimistic view, whether rational or not, about internet firms in the post-bubble period. Finally, note the following asymmetry: the relative pessimism on internet firms over non-internet firms in the post-bubble period was higher than the relative optimism on internet firms over non-internet firms in the bubble period. Table 2 documents results from formal statistical tests that corroborate our observations from the above informal ocular tests. Panel A shows prior to market peak defined as March 24, 2000, an internet firm receives on average 1.04 pieces of news reports per day and 21.04 per month. Among these news items, 0.39 per day and 7.98 per month are good news – more than twice the amount of good news a non- internet firm receives. The difference is statistically significant. The net news in the bubble period averages 0.10 per day and 2.17 per month for an internet firm, significantly higher than the 0.07 per day and 1.42 per month coverage for a non-internet firm. After the market peak, with an average of 0.65 news items per day and 14.01 news items per month, an internet firm still receives over twice the media attention than a non-internet firm. However, the media coverage is more negative about internet firms during this sub-period. The net news is -0.12 per day and - 2.28 per month for an internet firm, compared to -0.01 per day and -0.19 per month for a non-internet firm. Panel B of Table 2 shows the same pattern of media coverage under a different measure of peak: the day a firm’s price peaked in the period 1996 to 2000. So Table 2 leads us to conclude that during the bubble period the media coverage was more positive about internet IPOs than it was about non-internet IPOs, but during the post-bubble period, the media coverage was more negative about internet IPOs than it was about non-internet IPOs. 15 Since large offers tend to attract more public attention, we further break down our sample into two sub-samples, large and small IPOs, based on the median gross proceeds of the combined two samples. In two alternative and separate classifications, we break down our sample into tech and non-tech IPOs, as well as VC-backed and non VC-backed IPOs. Our previous observations of the difference in media coverage between the two samples hold, regardless of the size of the offer, the technological nature of the firm, and whether or not the issue is backed by venture capitalists. Finally, we confirm the asymmetry we had noticed before: the relative pessimism on internet firms over non-internet firms in the post-bubble period (net news was -0.12 per day for internet firms and -0.01 per day for non-internet firms) was higher than the relative optimism on internet firms over non-internet firms in the bubble period (net news was 0.10 per day for internet firms and 0.07 per day for non-internet firms.) B. Conditional Media Coverage We now explore conditional news coverage between the two groups of firms during our sample period. Specifically, do the media report more good news than bad news when the previous period experiences a price increase, and do the media report more bad news than good news when the previous period was a price decrease? If yes, this would be consistent with the positive feedback hypothesis discussed in Shiller (2000). Graphic illustrations of this test are given in Figures 3-a to 3-h in both calendar time and event time, and in aggregate and per-firm basis. We notice that the optimism of the media, captured by the net news items per month, moves along with market capitalization for both internet and non-internet firms in calendar time (Figure 3-b) and in event time (Figure 3-f). This effect is about the same for internet firms and non-internet firms in the pre-peak period, but it is much stronger for internet firms in the post-peak period. The post-peak difference between the two samples suggests the media are more pessimistic about falling prices for internet firms than they are about falling prices for non-internet firms in this period. Not only is media sentiment positively linked with price levels, but also is its interest. Notice that media coverage, as captured by the total news per month, moves along with market capitalization for both 16 internet and non-internet firms in calendar time (Figure 3-a) and in event time (Figure 3-e). The effect is stronger for internet firms both in the pre-peak period as well as in the post-peak period. Finally, observe that the net news per firm spiked before March 24, 2000 (Figure 3-d), or before the firm reached its maximum value (Figure 3-h), suggesting that media sentiment turned before the market peaked. We explore this tantalizing result formally in the next section, where we ask whether media sentiment Granger-causes returns. The findings from Figures 3-a to 3-h are corroborated formally in Table 3. We report the results based on two arbitrarily selected cutoff points about the degree of price movement: price increases or decreases more than 0% and 1% from previous day for daily study, and 0% and 10% from previous month for monthly analysis. Alternative cutoff points do not change our results qualitatively. Using abnormal returns instead of raw returns yields virtually identical results and hence these results are not reported. Panel A of Table 3 reveals that if prices increased in the previous period, net news was positive this period for both internet stocks (0.07 per day) and for non-internet stocks (0.05 per day). If prices decreased in the previous period, net news was negative this period for internet stocks (-0.06 per day), but still positive this period for non-internet stocks (0.01 per day). Examining alternative cutoff points for previous price movements yields the same pattern. This means that Shiller’s (2000) positive feedback hypothesis works especially for internet shares. The analysis of the bubble and the post-bubble stages in Panels B and C of Table 3 leads to more interesting results. It seems that only one leg of the positive feedback hypothesis worked in the bubble stage, and another leg of the positive feedback hypothesis worked in the post-bubble stage. In the bubble stage, if prices increased in the previous period, net news was positive this period, but if prices decreased in the previous period, net news was not negative this period. In the post-bubble stage, if prices decreased in the previous period, net news was negative this period, but if prices increased in the previous period, net news was not positive this period. Interestingly, during the bubble period, this asymmetry was both economically and statistically stronger for internet stocks when we examine net news per month. In the post-bubble period, this 17 asymmetry is stronger for internet stocks whether we use net news per day or net news per month. This suggests that in the bubble stage, the media coverage is non-negative in the event of price falls, especially for internet stocks; and in the post-bubble stage, the media coverage is non-positive in the event of price rises, especially for internet stocks. These results remain whether the peak is defined in calendar time or in event time. V. THE MARKET REACTION TO MEDIA COVERAGE In the previous section we documented the differences in aggregate media coverage between internet firms and non-internet firms. This is a necessary, but not sufficient, condition to conclude that the media had a role in the meteoric rise and fall of internet stocks in the late 1990s. In this section, we explicitly examine the impact of this differential media coverage on stock prices. We first conduct the analysis at the firm level, and then conduct the analysis at the portfolio level. A. Firm-Level Analysis The dependent variable is the abnormal return of a firm’s stock estimated by fitting a Fama-French (1993) 3-factor model for each firm. We use contemporaneous Fama-French factors to control for the most recent market-wide information to ensure the conservativeness of our news analysis, though our results are almost identical with respect to both economic and statistical significance if lagged factors are selected. We then examine the impact of different types of news on daily and monthly abnormal returns, respectively, by including number of net news (NN), good news (GN) and bad news (BN) per firm from the previous period (either day or month) as independent variables in our regression model. 12 Net news is simply number of good news minus number of bad news per firm from the previous period (either day or month). 12 Busse and Green (2002) use a sample of 322 news reports in the Morning Call and Midday Call segments on CNBC between June 12 and October 27, 2000, and find that prices respond within seconds. In contrast, we focus on aggregate media coverage and its effect on stocks over the entire bubble and post-bubble period. The nature of the print media, the availability of news publications, and the use of market closing prices in calculating returns generate only a daily frequency dataset. So we cannot say whether prices respond within seconds. However, as we will see, the effect lasts at least for a day. Antweiler and Frank (2004a) and Huberman and Regev (2001) also find a prolonged effect of news on returns. 18 Our control variables are the following. First, we include lagged abnormal returns from the previous period to be consistent with the Granger causality test. Lagged abnormal return also serves as a control variable for liquidity (Pastor and Stambaugh 2003), for possible previous information leakage (Antweiler and Frank 2004a), for any potential bid-ask bounce (Roll 1984). And, as the last period’s abnormal return contains the cumulative effect of past media reports, it corrects for the possibility that the media effect lasts longer than a day. In particular, because the lagged return occurs on the same day as the news reports used for tests, the lagged return controls for contemporaneous relationships between stock prices and news coverage. Analyses with up to five lagged abnormal returns as control variables yield virtually same results, so only the results incorporating one lag of abnormal returns are tabulated. Second, to control for transaction costs, we include the time t – 1 bid-ask spread Bid-Ask t-1 (closing bid-ask spread scaled by the price at the end of the day) and trading volume log(1+Vol t-1 ) for each firm, where Vol t-1 is the number of shares traded at time t-1. As the number of news reports is much higher for internet stocks than non-internet stocks, in regression analyses after A.1, we also control for total number of news reports for the purpose of normalization. A.1. The explanatory power of the media in explaining abnormal returns Considering that Busse and Green (2002) found that stock prices react within seconds to CNBC news, we first check whether the media have any effect on stock prices the next day. To quantify the explanatory power of the media, we estimate our regression model with only the control variables but without any news variables and note the R 2 . We then include the net news variable NN t-1 in this regression (later we break down net news into good news and bad news) and note the new R 2 . The increase in R 2 gives us the overall improvement in the fitness of the model. Except for the abnormal returns and the lagged abnormal returns, which already reflect the differences in log stock prices, both news variables and control variables are differenced in the regressions to remove firm fixed effects. Table 4 shows that adding news variables improves the overall fit of the model as R 2 increases significantly. For example, by including net news NN t-1 , R 2 increases 12% for the internet sample during the bubble period (from 0.34% to 0.38%), and 22% post bubble (from 0.51% to 0.62%). For the non- 19 internet sample, R 2 increases 58% during the bubble period, and 44% post bubble. This suggests that even at low frequency daily data, media coverage has impact on the next day’s abnormal returns. Interestingly, decomposing net news into good news and bad news does not help to improve the overall statistical fit for internet firms in each sub-sample period. A.2. The marginal impact of the media on abnormal returns We now formally estimate the marginal impact of media on risk-adjusted stock returns for the entire sample period of 1996-2000. We isolate the bubble period by interacting the news variables with PostPeak, a dummy which equals 1 if it is after the market peak of March 24, 2000, and zero otherwise. The firm-level regression model is assumed to be as follows: ( )iablesControlVarPostPeakNNNNfABRET ttt ,, 11 ×= ?? The regression analysis is conducted for the internet IPO sample and the non-internet IPO sample separately. Later, we re-estimate the model by breaking down net news variable NN t-1 into GN t-1 and BN t-1 to examine the separate impact from positive and negative news reports. Except for the abnormal returns and the lagged abnormal returns, which already reflect the differences in log stock prices, both news variables and control variables are differenced to remove firm fixed effects. In addition to lagged abnormal returns, bid-ask spread and trading volume, we now include log(1 + TN t-1 ) as a control variable, where TN t-1 is the total news per firm from the previous period (either day or month). This is because our regression model focuses on the impact of daily aggregate media coverage per firm on firm abnormal stock returns in a dynamic panel. It controls for firm fixed effects. Consequently, news variables frequently contain zero values, which prevent us from using any ratio transformations. To take into account the vast differences in the amount of news for internet and non- internet firms, we add log(1 + TN t-1 ) as a scaling variable to normalize the key news variables of interest. Since larger firms are commonly believed to receive more media attention than smaller firms, log(1 + TN t-1 ) can also be viewed as a control for size. Table 5 reports the daily returns results from the least-square estimation of the regression models. p-values are based on Newey-West standard errors (one lag). Although not reported, our results remain 20 unchanged when we re-estimate Table 5 for various lagged Newey-West standard errors (up to five lags). Winsorizing at the 1 st and 99 th percentile yields similar results. From Columns (1) and (2) in Table 5, the number of net news from previous period is positively and significantly related to the current abnormal returns. This indicates that conditional on previous abnormal returns, market liquidity, trading volumes and the amount of total news reports, net news leads to higher post-issue abnormal returns for both internet and non-internet IPO firms. We then explore media impact by dividing our analysis between internet sample and non-internet sample at a given period (a cross- sectional perspective) and between bubble period and post-bubble period for a given sample (a time-series perspective) respectively. From a cross-sectional perspective, Table 2 shows that there are more net news items for an internet firm than for a non-internet firm during the bubble period. However, Table 5 reports that marginal net news generates an extra 9.0 and 22.8 basis points in returns for an internet firm and a non-internet firm, respectively. Controlling for total news coverage, market trading conditions and the previous day’s price movements, the marginal impact of net news is much smaller (13.8 basis points) for internet firms than for non-internet firms. This difference is also statistically significant (p = 0.00). During the post-bubble period, the difference in coefficients associated with the number of net news between internet IPO sample and non- internet IPO sample remains negative but is not significant. This suggests that even though net news leads to higher abnormal returns, this effect is lower for internet firms, especially during the bubble period. 13 From a time-series perspective, notice that the coefficient associated with the interaction term between net news and post-bubble dummy (PostPeak) is positive for both samples (0.125 and 0.066 respectively). However, it is highly significant (p = 0.00) only for internet firms. This suggests that net news was more credible in the post-bubble period than in the bubble period, especially for internet firms. 13 Table 5 reveals the direct effect of the media. What about the indirect effect, the effect that is caused by positive feedback? A “back of the envelope” calculation reveals this indirect effect. From Table 3, a positive return on day t generates 0.19 net news items for internet firms and 0.09 for non internet firms on day t + 1. Coefficients shown in Table 5 then translate the effect of this news coverage to 0.017% (0.19 × 0.09%) abnormal returns for internet IPOs and 0.021% (0.09 × 0.228%) returns for non-internet IPOs. So, though positive feedback is higher for internet firms than non-internet firms (Table 3), its overall effect is lower for internet firms. 21 To summarize, the first two columns of Table 5 indicate that, after controlling for firm-fixed effects, previous abnormal returns, total volume of news, and market trading conditions, net news positively Granger-causes returns for internet IPOs, but the effect is lower in all periods compared to non-internet IPOs, especially in the bubble period. 14 We next explore the impact of the nature of the news by breaking down net news into good and bad news. This is shown in Columns (3) and (4) in Table 5. We find that conditional on previous abnormal returns, market liquidity, trading volumes and the amount of total news reports, good news leads to higher abnormal returns while bad news leads to lower abnormal returns for both internet and non-internet IPO firms. This result also suggests that the human judgment we used when classifying news as good, bad or neutral was prudent ex post. However, the effect of bad news is not statistically significant for internet firms. From a cross-sectional perspective, note that the difference in coefficients associated with good news between internet IPO sample and non-internet IPO sample is negative (-0.100) and marginally significant (p = 0.08) during the bubble period, suggesting that even though good news leads to higher abnormal returns, this effect is lower for internet firms during the bubble period. These results suggest the market relatively downplays the good news of internet firms in the bubble period. From a time-series perspective, notice that the coefficient associated with the interaction term between good news and post-bubble dummy (PostPeak) is positive (0.122) and highly significant (p = 0.01) only for internet firms. These results suggest that good news was more credible in the post-bubble period, especially for internet firms. Although the results are not reported in Table 5, we also analyze the impact of media coverage on monthly returns. Unlike the results on daily returns, we find no impact of monthly news on monthly abnormal returns. Including or excluding a fourth factor, the momentum factor, changes neither the statistical significances nor our conclusions. We interpret these monthly non-results simply as evidence in 14 In a study whose results compliment our results, Agrawal and Chen (2005) find that investors appear to have discounted analyst opinions more during the late-1990s stock market bubble. 22 favor of market efficiency. If markets are efficient, the impact of news on returns should be immediate; it should not take a month for prices to incorporate information. A.3. The overall impact of the media on abnormal returns The previous subsection focuses on the marginal impact of the media on risk-adjusted returns. Though the marginal effect of the media on the returns of internet firms was lower than their effect on the returns of non-internet firms, as there were more news items about the former than the latter (see Figures 2- a through 2-e), it could be argued that the media had a higher overall effect on the returns of internet firms than they had on non-internet firms. We now turn to the analysis of this overall news effect. We compute the cumulative abnormal returns predicted by news for an average internet firm and an average non-internet firm for the 1996 to 2000 sample period using the coefficient estimates in Columns (1) and (2) in Table 5. 15 Figure 4 compares this predicted cumulative abnormal return to the actual cumulative abnormal return for the average internet firm and for the average non-internet firm. Note that the difference in predicted returns between the internet firms and the non-internet firms is much less than the difference in actual returns. This suggests that the media coverage cannot account for the value difference between the two samples, which implies that the media is unlikely to be the primary cause of the internet bubble. B. Portfolio Analysis Tables 4 and 5 examine the impact of media on individual firm returns. To capture the effect of media on the entire internet group, we now extend our analysis from firm level to portfolio level. We replicate Table 5 in daily and monthly analyses for the following four portfolios: equally-weighted and value-weighted internet portfolios and equally-weighted and value-weighted non-internet portfolios. 15 This is done as follows. We regress the abnormal returns from the Fama-French 3-factor model on net news for internet and non-internet samples during the bubble period and the post-bubble period. We multiply the coefficient associated with the net news variable by the average number of net news items in each sub-sample in each sub-period to obtain a series of daily average predicted abnormal return for each sub-sample in each sub-period. We average across firms to get the average abnormal return predicted by news. We then cumulate these over time to obtain a predicted cumulative abnormal return. To make comparisons easy, we normalize the starting point of the predicted CAR to 100 for both sub-samples. 23 Table 6 reports the results from estimating the portfolio returns using a regression model similar to the one presented in Table 5. Bid-ask spreads and trading volumes are excluded from our analysis as these are control variables for firm-level liquidity measures, and cannot be averaged to capture the portfolio level of liquidity. Again, both alternative specification of Newey-West standard errors (up to five lags) and winsorizing at the 1 st and 99 th percentile yield virtually the same results and these results are therefore omitted from the table. From a time-series perspective, for the internet portfolio, previous average net news per firm in the internet portfolio has no impact on risk-adjusted portfolio returns during the bubble period, but leads to significant and positive abnormal portfolio returns during the post-bubble period for both equal-weighted and value-weighted portfolios. Decomposing net news into good news and bad news shows that good news was responsible for this result: previous average good news per firm in the portfolio has no impact on risk- adjusted portfolio returns during the bubble period, but leads to positive abnormal portfolio returns during the post-bubble period. This is not seen in the non-internet portfolio. As a matter of fact, from a cross-sectional perspective, the difference in coefficients associated with average net news per firm between internet IPO portfolio and non-internet IPO portfolio is only significant for the post-bubble period. This result is further confirmed by the difference in coefficients associated with average good news per firm. We make the following conclusion from the portfolio analysis: when examined at the portfolio level, net news about internet firms was less credible in the bubble period than net news about non-internet firms (economically, but not statistically different). Further, net news about internet firms was more credible in the post-bubble period than net news about non-internet firms (economically and statistically different). This conclusion is weaker than our conclusion from the individual firm analysis in Table 5, where we had an economically and statistically significant difference in both the bubble and the post- bubble periods. Although results are omitted from Table 6, we again find monthly news per firm in the portfolio does not impact monthly returns, which is another evidence of market efficiency. 24 VI. ROBUSTNESS OF RESULTS A. Change of information environment within the sample period On October 23, 2000, the SEC implemented the Regulation Fair Disclosure Law requiring that material disclosures by publicly traded companies be disseminated so that the disclosures are simultaneously accessible to all concerned. Prior to the adoption of this law, selective disclosure such as disclosing important nonpublic information to securities analysts or selected institutional investors or both was permissible. By enforcing this law, the SEC intended to eliminate differential informational advantages. We take into account this regime change in our data set by including a dummy variable in our regression analysis of Table 5. This dummy variable is 1 after October 23, 2000, and zero otherwise. The coefficient associated with the post-regulation dummy is negative and significant, suggesting that abnormal returns are lower after the implementation of the law. However, our qualitative results in Table 5 do not change. B. Alternative variable specifications and additional control variables We further check the robustness of our results by replacing the number of good news with the sum of good news and neutral news. In another check, we replace the number of bad news with the sum of bad news and neutral news. We obtain similar results from the first alternate measure of good news, and even stronger results from the second alternative measure. To take into account the process of information dissemination, we re-define NN t-1 as the average of NN t-1 and NN t-2 (GN t-1 , BN t-1 and TN t-1 are redefined analogously), and re-estimate our model in Table 5 and Table 6. Our results are robust to these alternative definitions. For Table 3, we substitute the raw returns with abnormal returns to capture the effect of risk- adjusted returns on media reporting. For Table 5, we replace the abnormal return at t-1 by the sum of two- day abnormal returns at t-1 and t-2. Our results are virtually the same. For example, the difference in coefficients associated with net news between internet IPO sample and non-internet IPO sample remains negative (-0.415) and highly significant (p = 0.01) during the bubble period. To take into account possible 25 impact from reversal and momentum, we also re-estimate our model in Table 5 and Table 6 by including additional control variables of lagged abnormal returns (up to 3 lags). Again, our results remain qualitatively unchanged. C. Alternative sample specifications C.1. Exclusion of price-driven news items Most of our news is about economic fundamentals, but some of our news is about the previous period’s price movement. Through the process of news collection and classification, we have observed that the second type of news occurs most frequently during the first month after a firm goes public. So we re- estimate Tables 5 and 6 by excluding the first-month data of each firm. Again, our results remain qualitatively unchanged for both daily and monthly returns. The difference in coefficients associated with net news between internet IPO sample and non-internet IPO sample is -0.187 (p = 0.00) during the bubble period, and remains negative (-0.104) but becomes marginally significant (p = 0.08) during the post-bubble stage. C.2. Paired matching: Early termination due to mergers, liquidations, bankruptcy and delisting In this paper, we investigate the impact of the media on IPO firms’ returns during the bubble and post-bubble periods by comparing the media coverage between internet IPO firms and a matching sample of non-internet IPO firms. There are 31 internet issues acquired later during the sample period, 19 of which are acquired by firms outside the internet sample. In addition, there are 2 internet firms liquidated, 1 bankrupted, and 2 de-listed. It is possible that an internet IPO firm is terminated earlier from our internet sample while the returns of its matched non-internet firm are still included in the non-internet sample. To check the robustness of our results, we first re-estimate Table 5 using sub-samples in which all the firms terminated prior to the end of the sample period are dropped. The signs of the key coefficients do not change, but the magnitude of the difference in coefficient associated with net news between internet sample and non-internet sample decreases (from -0.137 in the original model to -0.097). The coefficient does remain statistically significant (p = 0.02). Second, to ensure a perfect paired-matching in our analysis, we construct another two sub-samples which require both the individual internet IPO firm and its matching 26 firm have non-missing values in returns on any given day within the same period. Again, our key results do not change. D. Boredom: Are investors over-exposed to the news? The key result in this paper, that the market discounts the media coverage for internet IPO firms, especially during the bubble period, may also be explained by investors’ limited attention (Daniel, Hirshleifer and Teoh, 2002, and Hirshleifer and Teoh, 2003) and over-exposure to the news reports at that time. Given the substantially high volume of media coverage on internet firms during the bubble period, investors who have been surrounded by the news about the same type of firms in the past may eventually “grow tired” of any reports about internet stocks, and hence may discount the impact of the news. D.1. The cumulative effect To check this claim, we first examine the cumulative news exposure by re-estimating our regression model in Table 5 with additional four control variables. For each firm in our sample, we multiply cumulative total number of news of the firm up to date to each of the following independent variables of interest in Table 5: net news and the interaction term between net news and the post-bubble dummy. These new control variables allow the impact of news on abnormal returns to decrease as boredom sets in, where boredom is defined as the cumulative total number of news items till then. Surprisingly, we do not find evidence of boredom. The difference in coefficients associated with net news between internet sample and non-internet sample remains negative (-0.125) and significant (p = 0.01), but the coefficients associated with the two new control variables are not significant (p = 0.61 and 0.11 respectively). D.2. The non-linear effect Instead of number of net news, good news and bad news, we take a log transformation of these key variables and re-estimate Table 5. The idea behind this is an appreciation of the fact that the effect of news decreases as the number of news items increase. However, our key result is robust even to this non-linear transformation. The difference in coefficients associated with net news between internet sample and non- internet sample remains negative (-0.180) and significant (p = 0.03) for the bubble period. 27 E. Lockup expiration Most IPOs feature lockup agreements, which prevent insiders from selling their shares to the market over a specified period, typically 180 days. Field and Hanka (2001) show that the popular press has interest in lockup expiration and find significant negative abnormal return around the scheduled unlock day. To correct for the possible unusual impact of news coverage around lockup expiration, we re- estimate the regression model in Table 5 by removing the returns and news data of each firm in our sample in its sixth month after the offer date. We find no evidence that the event of lockup expiration affects our results during this sample period. The economical interpretation of all coefficients remain qualitatively unchanged, although the difference in coefficients associated with net news between internet IPO sample and non-internet IPO sample becomes significant (p = 0.02) for the post-bubble period. F. Learning It can be argued that our main result – the market “discounts” news coverage for internet firms – is driven entirely by the learning process of investors about them. We know that learning curves flatten out over time, and as many internet firms went public later in the sample period than earlier, the media impact on the returns of internet firms may have been lower than the media impact on the returns of non-internet firms. We re-estimate our regressions using only the data between 1996 and 1998. If the argument about learning holds, then the documented “market-discounting” activities should be lower or less significant at this early stage of the sample period. Surprisingly, we find the market discounts the net news about internet firms more during this early sub-sample period. The difference in coefficients associated with net news between the two samples is -0.213 and significant (p = 0.01). G. Experiment with news classification To ensure that our categorization is consistent, we conducted a small experiment (Human Subject Study # 04-9087, approved on April 22, 2004). To utilize news from all stages of the internet boom and the subsequent bust, we selected two firms among the earliest IPOs in the combined sample, Yahoo! from internet stocks and Sapient Inc. from non-internet stocks. We then recruited seven undergraduate students 28 to participate in the experiment and divided them into two groups. 16 The three students in the first group were each given 100 random news items about Yahoo! and the four students in the second group were each given 100 random news items about Sapient Inc. The undergraduates were instructed to use their own judgment to categorize each news article into good, bad, or neutral, except in cases of news about insider trading (sells are automatically bad, buys are automatically good) or news about analyst recommendations. The experiment occurred on April 23, 2004 and lasted about two hours. Each student received a payment of $50 for his participation in this experiment. The resulting number of instances of agreement between the authors and each undergraduate is presented in Table 7. From Panel A, the control firm news results in relatively few disagreements. The authors agree in 71% of cases. Though unreported, that number jumps to over 97% if neutral classifications are ignored. Undergraduates 1, 2, and 4, agree to a similar degree, suggesting that anyone reading news for non-internet firms comes to roughly the same conclusion. Only undergraduate number 3 appears to classify news differently. While individuals may disagree, these disagreements appear to be random and cancel out on average. Panel B demonstrates that news on internet firms is harder to interpret. The authors agree in 65% of cases, though that number jumps to 90% when neither chooses a neutral classification. This differential arises mainly from the fact that Author 1 is less conservative in assigning classifications. Interestingly, since Author 2 was responsible for much of the internet data collection, this fact suggests that the effect of news on internet returns is actually slightly upward biased. Even with this potential bias, we are able to determine that internet firms have a significantly lower marginal response to news items than do control firms. The undergraduates appear to have considerably different opinions about the impact of news. Again, though the individuals differ in their opinions, an average constructed from individual classifications shows disagreements canceling out in the aggregate. 17 16 We actually recruited eight undergraduate students, but one did not show up at the time of the experiment. 17 Our consistency results are in line with Niederhoffer (1971) who documents an 80% agreement between the average of two news coders and ten college students when classifying thirty randomly picked New York Times headline news per person. In a parallel comparison, the maximum precision level of electronic filters in the artificial intelligence 29 VII. CONCLUSION ISDEX, an authoritative and widely cited internet stock index, rose from 100 in January 1996 to 1100 in February 2000 – an incredible increase of about 1000% in four years – only to fall down to 600 in May 2000 – an incredible decrease of about 45% in four months. Of all the bubbles in history, this internet bubble ranks amongst the most spectacular. Though there is much agreement that such a spectacular rise and fall of internet stock prices cannot be explained by fundamentals, there is less agreement of what can explain it. In this paper, we explore the role of the media in this internet bubble. We ask and answer the following two questions: was the media coverage for internet IPOs in the years 1996 through 2000 different from a matching sample of non-internet IPOs and, if yes, whether this differential media coverage had any effect on the difference in risk-adjusted returns between internet stocks and non-internet stocks. Our answer to the first question is the following. The media coverage was more intense for internet IPOs: there were more total news, more good news and more bad news for internet IPOs than for a matching sample of non-internet IPOs. We further document that media coverage was more positive for internet IPOs in the bubble period, and more negative for internet IPOs in the post bubble period. Moreover, we note that positive feedback – positive (negative) price changes leads to increase (decrease) in net news – was particularly pronounced for internet stocks. Our answer to the second question is the following. We find that the market somewhat downplayed the intensive media coverage on internet firms relative to the non-internet firms: though net news Granger caused risk-adjusted returns for both groups of IPOs, the effect was lower for internet IPOs, especially in the bubble period. So our conclusion is that the media was not a significant factor in the dramatic rise and fall of internet shares in the late 1990s. literature when deployed for mainstream news has been achieved at 50% (see, for example, Shepherd, Watters and Marath 2002 for a detailed discussion). 30 REFERENCE Agrawal, Anup and Mark Chen, 2005, Do analyst conflicts matter? Evidence from stock recommendations, Working Paper, University of Alabama and University of Maryland. Antunovich, Peter, and Asani Sarkar, 2003, Fifteen minutes of fame? The market impact of internet stock picks, Working Paper, The Federal Reserve Bank of New York. Antweiler, Werner, and Murray Z. Frank, 2004a, The market impact of corporate news stories, Working Paper, University of Minnesota. Antweiler, Werner, and Murray Z. Frank, 2004b, Is all that talk just noise? The information content of internet stock message boards, Journal of Finance, forthcoming. Barber, Brad, M., and Terrance Odean, 2003, All that glitters: The effect of attention and news on the buying behavior of individual and institutional investors, Working Paper, University of California, Davis and University of California, Berkeley. Baron, David, P., 2004, Persistent media bias, Working Paper, Stanford University. Benveniste, Lawrence, Alexander P. Ljungqvist, William J. Wilhelm Jr., and Xiaoyun Yu, 2003, Evidence of information spillovers in the production of investment banking services, Journal of Finance 58, 577-608. Blanchard, Olivier, J., and Mark W. Watson, 1982, Bubbles, rational expectations and financial markets, in Crises in the Economic and Financial System, ed. by P. Wechtel. Lexington, Lexington Books. Brunnermeier, Markus, and Stefan Nagel, 2004, Hedge funds and the technology bubble, Journal of Finance 59, 2013-2040. Busse, Jeffrey, and T. Clifton Green, 2002, Market efficiency in real time, Journal of Financial Economics 65, 415-437. Chan, Wesley, 2003, Stock price reaction to news and no-news: drift and reversal after headlines, Journal of Financial Economics 70, 223-260. Chen, Honghui, Gregory Noronha, and Vijay Singal, 2002, Investor awareness and market segmentation: Evidence from S&P 500 Index changes, Working Paper, University of Baltimore. 31 Cooper, Michael J., Orlin Dimitrov, and P. Raghavendra Rau, 2001, A Rose.com by any other name, Journal of Finance 56, 2371-2388. Cutler, David M., James M. Poterba and Lawrence H. Summers, 1989, What moves stock prices? Journal of Portfolio Management 15, 4-12. Daniel, Kent, David Hirshleifer, and Siew Hong Teoh, 2002, Investor psychology in capital markets: evidence and policy implications, Journal of Monetary Economics 49, 139-209. Demers, Elizabeth, and Katharina Lewellen, 2003, The marketing role of IPOs: evidence from internet stocks, Journal of Financial Economics 68, 413-437. Ducharme, Larry, Shivaram Rajgopal, and Stephan Sefcik, 2001a, Lowballing for 'Pop': The case of internet IPO underpricing, Working Paper, University of Washington. Ducharme, Larry, Shivaram Rajgopal, and Stephan Sefcik, 2001b, Why was internet IPO underpricing so severe? Working Paper, University of Washington. Dyck, Alexander, I. J., and Luigi Zingales, 2003a, The bubble and the media, in Peter Cornelius, and Bruce Kogut, eds.: Corporate Governance and Capital Flows in a Global Economy, Chapter 7 (Oxford University Press, New York). Dyck, Alexander, I. J., and Luigi Zingales, 2003b, The media and asset prices, Working Paper, Harvard University and University of Chicago. Falkenstein, Eric, G., 1996, Preferences for stock characteristics as revealed by mutual fund portfolio holdings, Journal of Finance 51, 111-135. Fama G, and Kenneth French, 1993, Common risk factors in the returns on stocks and bonds, Journal of Financial Economics 33, 3-56. Field, Laura, and Gordon Hanka, 2001, The expiration of IPO shares lockups, Journal of Finance 56, 471- 500. Griffin, John, Jeffrey H. Harris and Selim Topaloglu, 2003, Investor behavior over the rise and fall of Nasdaq. Working Paper, University of Texas. 32 Hirota, Shinichi, and Shyam Sunder, 2002, Stock market as a “beauty contest”: Investor beliefs and price bubbles sans dividend anchors. Working Paper, Waseda University and Yale University. Hirshleifer, David, and Siew Hong Teoh, 2003, Limited attention, information disclosure, and financial reporting, Journal of Accounting and Economics 36, 337-386. Ho, Beauty, Maya Taher, Robert Lee, and Neil L. Fargher, 2001, Market sentiment, media hype and the underpricing of initial public offerings: The case of Australian technology IPOs, Working Paper, University of New South Wales. Huberman, Gur, and Tomer Regev, 2001, Contagious speculation and a cure for cancer: A nonevent that made stock prices soar, Journal of Finance 56, 387-396. Johnson, William, C., and Jennifer Marietta-Westberg, 2004, The effect of news on volatility: A study of IPOs, Working Paper, Michigan State University. Lamont, Owen and Richard Thaler, 2003, Can the market add and subtract? Mispricing in tech stock carve- outs, Journal of Political Economy 111, 227-268. Loughran, Tim, and Jennifer Marietta-Westberg, 2002, When any news is bad news: extreme events following an initial public offering. Working Paper, Notre Dame University. Loughran, Tim, and Jay Ritter, 2002, Why don’t issuers get upset about leaving money on the table in IPOs? Review of Financial Studies 15, 413-443. Loughran, Tim, and Jay Ritter, 2004, Why has IPO underpricing changed over time? Financial Management 33, 5-37. Lowry, Michelle, and G. William Schwert, 2002, IPO market cycles: Bubbles or sequential learning? Journal of Finance 57, 1171-1200. Ljungqvist, Alexander, and William J. Wilhelm, Jr., 2003, IPO pricing in the dot-com bubble, Journal of Finance 58, 723-752. Merton, Robert, 1987, Presidential address: A simple model of capital market equilibrium with incomplete information, Journal of Finance 42, 483-510. 33 Mitchell, Mark, L., and J. Harold Mulherin, 1994, The impact of public information on the stock market, Journal of Finance 49, 923-950. Mullainathan, Sendhil, and Andrei Shleifer, 2003, The market for news, Working Paper, MIT and Harvard University. Niederhoffer, Victor, 1971, The analysis of world events and stock prices, Journal of Business 44, 193-219. Ofek, Eli, and Matthew Richardson, 2002, The valuation and market rationality of internet stock prices, Oxford Review of Economic Policy 18, 265-287. Ofek, Eli, and Matthew Richardson, 2003, DotCom mania: The rise and fall of internet stock prices, Journal of Finance 58, 1113-1137. Pastor, Lubos, and Robert Stambaugh, 2003, Liquidity risk and expected stock returns, Journal of Political Economy 111, 642-685. Reese, William, Jr., 1998, IPO underpricing, trading volume, and investor interest, Working Paper, Tulane University. Ritter, Jay, and Ivo Welch, 2003, A review of IPO activity, pricing, and allocations, Journal of Finance 57, 1795-1830. Roll, Richard, 1984, A simple implicit measure of the effective bid-ask spread in an efficient market, Journal of Finance 39, 1127-1139. Shepherd, Michael, Carolyn, Watters, and Ajitha T. Marath, 2002, Adaptive user modeling for filtering electronic news, Working Paper, Dalhousie University. Shiller, Robert, 2000, Irrational Exuberance. Princeton University Press, Princeton, NJ. Tetlock, Paul, 2003, Giving content to investor sentiment: The role of media in the stock market, Working Paper, University of Texas. Wall Street Journal, July 12, 1999, E-Commerce (A Special Report): eToys Story: the tale of one site that has had a short, tumultuous – and so far, successful – life online; what’s its secret? Wysocki, Peter, D., 1999, Cheap talk on the web: The determinants of postings on stock message boards, Working Paper, University of Michigan. 34 Aggregate News Coverage (1996-2000) Good Good Bad Bad Total Total Net Net -10,000 10,000 30,000 50,000 70,000 90,000 110,000 130,000 NON-INTERNET INTERNET N u mb er o f N e w s I t em s Figure 1-a. Aggregate news coverage for the entire sample period (1996-2000) Aggregate News Coverage: Prior to Firm's Maximum Market Cap Good Good Bad Bad Total Total Net Net -10,000 10,000 30,000 50,000 70,000 90,000 NON-INTERNET INTERNET N u mb er o f N e w s I t em s Figure 1-b. Aggregate news coverage prior to firm’s maximum market cap Aggregate News Coverage: Post Firm's Maximum Market Cap Good Good Bad Bad Total Total Net Net -10,000 10,000 30,000 50,000 70,000 90,000 NON-INTERNET INTERNET N u mb er o f N e w s I t em s Figure 1-c. Aggregate news coverage post firm’s maximum market cap 35 Aggregate News Coverage: Prior to Nasdaq's Peak (3/24/2000) Good Good Bad Bad Total Total Net Net -10,000 10,000 30,000 50,000 70,000 90,000 NON-INTERNET INTERNET N u m b er o f N e w s I t em s Figure 1-d. Aggregate news coverage prior to Nasdaq’s Peak (March 24, 2000) Aggregate News Coverage: Post Nasdaq's Peak (3/24/2000) Good Good Bad Bad Total Total Net Net -10,000 10,000 30,000 50,000 70,000 90,000 NON-INTERNET INTERNET N u mb er o f N e w s I t em s Figure 1-e. Aggregate news coverage post Nasdaq’s Peak (March 24, 2000) 36 Daily News Per Firm (1996-2000) Good Good Bad Bad Total Total Net Net -0.20 0.00 0.20 0.40 0.60 0.80 1.00 NON-INTERNET INTERNET N u mb er o f N e w s I t ems P er F i r m Figure 2-a. Daily average news coverage per firm for the entire sample period (1996-2000) Daily News Per Firm: Prior to Firm's Maximum Market Cap Good Good Bad Bad Total Total Net Net -0.20 0.00 0.20 0.40 0.60 0.80 1.00 1.20 NON-INTERNET INTERNET N u m b er o f N e w s I t em s P er F i r m Figure 2-b. Daily average news coverage per firm prior to firm’s maximum market cap Daily News Per Firm: Post Firm's Maximum Market Cap Good Good Bad Bad Total Total Net Net -0.20 0.00 0.20 0.40 0.60 0.80 1.00 1.20 NON-INTERNET INTERNET N u mb er o f N e w s I t em s P er F i r m Figure 2-c. Daily average news coverage per firm post firm’s maximum market cap 37 Daily News Per Firm: Prior to Nasdaq's Peak (3/24/2000) Good Good Bad Bad Total Total Net Net -0.20 0.00 0.20 0.40 0.60 0.80 1.00 1.20 NON-INTERNET INTERNET N u mb er o f N e w s I t em s P er F i r m Figure 2-d. Daily average news coverage per firm prior to Nasdaq’s Peak (March 24, 2000) Daily News Per Firm: Post Nasdaq's Peak (3/24/2000) Good Good Bad Bad Total Total Net Net-0.20 0.00 0.20 0.40 0.60 0.80 1.00 1.20 NON-INTERNET INTERNET N u mb er o f N e w s I t em s P er F i r m Figure 2-e. Daily average news coverage per firm post Nasdaq’s Peak (March 24, 2000) 38 Total News Items Per Month (Calendar Time) 0 1,000 2,000 3,000 4,000 5,000 6,000 7,000 2-1996 6-1996 10-1996 2-1997 6-1997 10-1997 2-1998 6-1998 10-1998 2-1999 6-1999 10-1999 2-2000 6-2000 10-2000 Ne w s It e m s P e r M o n t h Non-Internet: Total News Internet: Total News Non-Internet: Market Cap. Internet: Market Cap. M a rk et C a p . Pe r Fi rm ( $ m i l) Figure 3-a. Total number of news articles per month for the entire sample period (1996-2000) Net News Items (Good - Bad) Per Month (Calendar Time) -1,800 -800 200 1,200 2,200 3,200 2 - 19 96 6 - 19 96 10 - 1 9 9 6 2 - 19 97 6 - 19 97 10 - 1 9 9 7 2 - 19 98 6 - 19 98 10 - 1 9 9 8 2 - 19 99 6 - 19 99 10 - 1 9 9 9 2 - 20 00 6 - 20 00 10 - 2 0 0 0 Ne t N e ws I t e m s P e r M o nt h Non-Internet: Net News Internet: Net News Non-Internet: Market Cap. Internet: Market Cap Ma r k e t C a p . P er Fi rm ( $ m i l ) Figure 3-b. Net number of news articles (good – bad) per month for the entire sample period (1996-2000) 39 Total News Items Per Firm Per Month (Calendar Time) 0 0.5 1 1.5 2 2.5 3 2-1996 6-1996 10-1996 2-1997 6-1997 10-1997 2-1998 6-1998 10-1998 2-1999 6-1999 10-1999 2-2000 6-2000 10-2000 N e w s I t e m s Pe r Fi rm Pe r Mo n t h Non-Internet: Total News Internet: Total News Non-Internet: Market Cap. Internet: Market Cap. M a rk et C a p . Pe r Fi rm ( $ b i l) Figure 3-c. Total news articles per firm per month for the entire sample period (1996-2000) Net News (Good-Bad) Per Firm Per Month (Calendar Time) -0.3 0.2 0.7 1.2 1.7 2.2 2.7 3.2 2-1996 6-1996 10-1996 2-1997 6-1997 10-1997 2-1998 6-1998 10-1998 2-1999 6-1999 10-1999 2-2000 6-2000 10-2000 N e t N e w s I t e m s Pe r Fi rm Pe r M o n t h Non-Internet: Net News Internet: Net News Non-Internet: Market Cap. Internet: Market Cap. Ma rk et C a p . Pe r Fi rm ( $ b i l ) Figure 3-d. Net number of news articles (good – bad) per firm per month for the entire sample period (1996-2000) 40 Total News Items Per Month (Event Time) 0 1,000 2,000 3,000 4,000 5,000 6,000 7,000 8,000 9,000 -30 -27 -24 -21 -18 -15 -12 -9 -6 -3 0 3 6 9 12 15 18 21 24 27 30 N e w s I t em s Pe r Mo n t h Non-Internet: Total News Internet: Total News Non-Internet: Market Cap. Internet: Market Cap. M a rk et C a p . P e r F i r m ( $ m i l) Figure 3-e. Total number of news articles per month based on firm’s maximum market cap Net News Items (Good-Bad) Per Month (Event Time) -1,200 -700 -200 300 800 1,300 1,800 2,300 2,800 3,300 3,800 -30 -27 -24 -21 -18 -15 -12 -9 -6 -3 0 3 6 9 12 15 18 21 24 27 30 Ne t Ne ws I t e m s P e r M o nt h Non-Internet: Net News Internet: Net News Non-Internet: Market Cap. Internet: Market Cap. M a rk et C a p . Pe r Fi rm ( $ m i l) Figure 3-f. Net number of news articles (good – bad) per month based on firm’s maximum market cap 41 Total News Items Per Firm Per Month (Event Time) 0 0.5 1 1.5 2 2.5 3 3.5 -30 -27 -24 -21 -18 -15 -12 -9 -6 -3 0 3 6 9 12 15 18 21 24 27 30 N e w s I t e m s Pe r Fi rm Pe r Mo n t h Non-Internet: Total News Internet: Total News Non-Internet: Market Cap. Internet: Narket Cap. M a rk et C a p . Pe r Fi rm ( $ b i l) Figure 3-g. Total number of news articles per firm per month based on firm’s maximum market cap Net News Items (Good-Bad) Per Firm Per Month (Event Time) -0.4 0.1 0.6 1.1 1.6 2.1 2.6 3.1 3.6 -3 0 -2 7 -2 4 -2 1 -1 8 -1 5 -1 2 -9 -6 -3 0 3 6 9 12 15 18 21 24 27 30 Ne t Ne ws I t e m s P e r F i r m P e r M o n t h Non-Internet: Net News Internet: Net News Non-Internet: Market Cap. Internet: Market Cap. M a rk et C a p . P e r F i r m ( $ b i l ) Figure 3-h. Net number of news articles (good – bad) per firm per month based on firm’s maximum market cap 42 Actual Cumulative Abnormal Returns vs. Cumulative Abnormal Returns Predicted by News 0 50 100 150 200 250 300 350 400 1-19 97 2-19 97 4-19 97 6-19 97 8-19 97 10 -1 99 7 12 -1 99 7 2-19 98 4-19 98 6-19 98 8-19 98 9-19 98 11 -1 99 8 1-19 99 3-19 99 5-19 99 7-19 99 9-19 99 11 -1 99 9 1-20 00 3-20 00 5-20 00 6-20 00 8-20 00 10 -2 00 0 12 -2 00 0 Ac t u a l Cu m u l a t i v e Ab n o r m a l Re t u r n s 98 99 100 101 102 103 104 105 106 107 Cu m u l a t i v e Ab n o r m a l Re t u r n s P r e d i c t e d b y Ne ws Cumulative Abnormal Returns for Non-Internet Firms -- Actual Cumulative Abnormal Returns for Internet Firms -- Actual Cumulative Abnormal Returns for Non-Internet Firms -- Predicted by News Cumulative Abnormal Returns for Internet Firms -- Predicted by News Figure 4. The overall news effect on stock returns 43 Table 1. Descriptive statistics of sample firms The sample period is 1996-2000. Internet IPOs are identified as in Loughran and Ritter (2004). Gross proceeds do not include the over-allotment option. The expected offer price is calculated as the midpoint of the indicative filing range. Price revisions are the percentage change between the expected and final offer prices. Initial return is the first-day close price over the offer price, minus one. Gross spread is the total manager’s fee expressed as the percentage of offer price. Information regarding venture capital backing is from Securities Data Corporation (SDC). Length of the book building period is the pre-IPO stage between the filing day (when a company files a preliminary prospectus with the SEC) and the pricing day (when the final offer price is set). Age is IPO year minus founding year. We manually collect missing founding date for 193 issues within the non-internet sample and 222 issues within the internet sample from SEC prospectuses, subsequent 10-Ks, or news sources. “High-tech” industries are classified by the first three-digit SIC codes 283, 357, 366, 367, 381, 382, 383, 384, 737, 873, and 874 covering industries such as pharmaceuticals, computing, computer equipment, electronics, medical and measurement equipment, biotech, and software industries. This definition follows Benveniste, Ljungqvist, Wilhelm and Yu (2003) and “Hi Tech Industry Group” defined by SDC. ***, **, * represents difference of internet sample from non-internet sample at 1%, 5% and 10% level (two-sided, Satterthwaite test for means and Wilcoxon signed rank test for medians), respectively. Non-Internet Sample Internet Sample Statistical Significance Gross proceeds (in $MM) Mean 87.96 88.22 Std. Dev. 122.72 124.54 Median 60.50 61.05 No. of obs. 458 458 Filing price range Mean 1.98 1.96 Std. Dev. 0.46 0.57 Median 2 2.00 No. of obs. 458 456 Expected offer price Mean 13.24 12.14 *** Std. Dev. 3.45 4.34 Median 13 11.5 *** No. of obs. 458 456 Final offer price Mean 13.67 14.76 *** Std. Dev. 4.67 5.62 Median 13 14 ** No. of obs. 458 458 Price revisions Mean 4.15% 23.00% *** Std. Dev. 27.66% 37.29% Median 0.00% 18.18% *** No. of obs. 458 456 Initial returns Mean 41.09% 83.72% *** Std. Dev. 68.07% 100.57% Median 17.68% 49.17% *** No. of obs. 458 458 Gross Spread Mean 6.97% 7.09% ** Std. Dev. 0.58% 0.99% Median 7.00 7.00 44 No. of obs. 458 458 Fraction Venture Capital Backed Mean 54.59% 69.65% *** Length of book-building period (in days) Mean 104.57 91.46 ** Std. Dev. 94.93 47.65 Median 77 77 No. of obs. 457 455 Firm age (in Years) Mean 9.56 4.84 *** Std. Dev. 12.49 4.64 Median 5 3 *** No. of obs. 434 444 Fraction of high tech issues Mean 62.45% 67.47% Fraction of issues traded at NYSE 8.95% 1.09% Nasdaq 87.12% 95.63% American 1.31% 0.87% OTC or Small Cap Market 2.63% 2.41% 45 Table 2. Difference in media coverage The news item data is hand-collected from Dow Jones Interactive and Factiva for both the internet and the non-internet sample IPOs. We read and classify them as good, bad or neutral news. Net news is the difference between the number of good news and the number of bad news. For each sample, we report the average daily and average monthly news item per firm before and after the peak, where the peak is measured in both calendar time, centered on March 24th, 2000 when Nasdaq’s composite index QQQ reached its highest value, and event time, when a firm reaches its own maximum market cap during the sample period. Large IPOs are issues with offer sizes greater than the combined sample median. The remaining issues are classified as small IPOs. p-values testing the difference in the degree of media coverage between the pre-peak period and the post-peak period are based on Satterthwaite standard errors and are reported in column (3) and (6) respectively. p-values testing the difference in the degree of media coverage between the internet sample and the non-internet sample are based on Satterthwaite standard errors and reported in column (7) and (8) respectively. (1) (2) (3) (4) (5) (6) (7) (8) Internet IPOs Non-Internet IPOs p-value for difference between Before After p-value Before After p-value (1) and (4) (2) and (5) Panel A: Calendar Time Daily News Items Total News 1.035 0.649 0.000 0.303 0.231 0.000 0.000 0.000 Net News 0.098 -0.115 0.000 0.068 -0.011 0.000 0.000 0.000 Good News 0.389 0.171 0.000 0.148 0.085 0.000 0.000 0.000 Bad News 0.292 0.286 0.258 0.080 0.096 0.000 0.000 0.000 Large IPOs Total News 1.154 0.731 0.000 0.365 0.261 0.000 0.000 0.000 Net News 0.131 -0.132 0.000 0.099 -0.004 0.000 0.001 0.000 Small IPOs Total News 0.941 0.528 0.000 0.263 0.191 0.000 0.000 0.000 Net News 0.072 -0.089 0.000 0.047 -0.020 0.000 0.001 0.000 Tech IPOs Total News 1.055 0.683 0.000 0.331 0.253 0.000 0.000 0.000 Net News 0.100 -0.114 0.000 0.058 -0.025 0.000 0.001 0.001 Non-Tech IPOs Total News 1.038 0.593 0.000 0.278 0.202 0.000 0.000 0.000 Net News 0.097 -0.118 0.000 0.075 0.007 0.000 0.044 0.000 VC Backed IPOs Total News 1.282 0.746 0.000 0.352 0.263 0.000 0.000 0.000 Net News 0.086 -0.151 0.000 0.071 -0.026 0.000 0.092 0.000 Non-VC Backed IPOs Total News 0.675 0.427 0.000 0.270 0.193 0.000 0.000 0.000 Net News 0.120 -0.030 0.000 0.065 0.007 0.000 0.000 0.000 Monthly News Items Total News 21.039 14.008 0.300 6.002 4.928 0.000 0.000 0.000 Net News 2.173 -2.276 0.000 1.420 -0.192 0.000 0.000 0.000 Good News 7.980 3.792 0.000 2.966 1.830 0.000 0.000 0.000 Bad News 5.808 6.068 0.000 1.546 2.021 0.000 0.000 0.000 Large IPOs Total News 23.531 16.357 0.000 7.470 5.592 0.000 0.000 0.000 Net News 3.015 -2.618 0.000 2.206 0.103 0.000 0.022 0.000 Small IPOs Total News 20.079 11.658 0.000 5.308 4.254 0.000 0.000 0.000 Net News 1.742 -1.929 0.000 1.048 -0.492 0.000 0.002 0.000 Tech IPOs 46 Total News 21.446 14.653 0.000 6.578 5.426 0.000 0.000 0.000 Net News 2.174 -2.221 0.000 1.258 -0.484 0.000 0.000 0.000 Non-Tech IPOs Total News 21.178 13.004 0.000 5.560 4.351 0.000 0.000 0.000 Net News 2.263 -2.394 0.000 1.556 0.187 0.000 0.000 0.000 VC Backed IPOs Total News 26.037 16.062 0.000 6.952 5.651 0.000 0.000 0.000 Net News 1.949 -3.007 0.000 1.532 -0.497 0.000 0.000 0.000 Non-VC Backed IPOs Total News 13.911 9.302 0.000 5.429 4.139 0.000 0.000 0.000 Net News 2.611 -0.550 0.000 1.348 0.171 0.000 0.000 0.000 Panel B: Event Time Daily News Items Total News 1.102 0.720 0.000 0.354 0.208 0.000 0.000 0.000 Net News 0.135 -0.076 0.000 0.090 -0.014 0.000 0.000 0.000 Good News 0.418 0.216 0.000 0.178 0.075 0.000 0.000 0.000 Bad News 0.283 0.292 0.100 0.084 0.088 0.007 0.000 0.000 Large IPOs Total News 1.206 0.808 0.000 0.406 0.242 0.000 0.000 0.000 Net News 0.180 -0.091 0.000 0.121 -0.011 0.000 0.000 0.000 Small IPOs Total News 1.024 0.608 0.000 0.312 0.173 0.000 0.000 0.000 Net News 0.101 -0.056 0.000 0.066 -0.016 0.000 0.000 0.000 Tech IPOs Total News 1.168 0.731 0.000 0.387 0.227 0.000 0.000 0.000 Net News 0.130 -0.072 0.000 0.086 -0.035 0.000 0.000 0.000 Non-Tech IPOs Total News 1.045 0.705 0.000 0.323 0.187 0.000 0.000 0.000 Net News 0.156 -0.081 0.000 0.096 0.009 0.000 0.000 0.000 VC Backed IPOs Total News 1.348 0.829 0.000 0.393 0.243 0.000 0.000 0.000 Net News 0.125 -0.120 0.000 0.098 -0.037 0.000 0.001 0.001 Non-VC Backed IPOs Total News 0.681 0.520 0.000 0.323 0.176 0.000 0.000 0.000 Net News 0.164 0.007 0.000 0.085 0.008 0.000 0.000 0.916 Monthly News Items Total News 21.857 15.346 0.000 6.922 4.494 0.000 0.000 0.000 Net News 2.596 -1.382 0.000 1.713 -0.139 0.000 0.000 0.000 Good News 8.251 4.697 0.000 3.446 1.690 0.000 0.000 0.000 Bad News 5.655 6.079 0.114 1.733 1.829 0.224 0.000 0.000 Large IPOs Total News 23.009 17.974 0.000 5.380 7.839 0.000 0.000 0.000 Net News 3.640 -1.715 0.000 0.143 2.152 0.000 0.000 0.000 Small IPOs Total News 22.007 12.976 0.000 6.328 3.827 0.004 0.000 0.000 Net News 2.128 -1.075 0.000 1.428 -0.351 0.000 0.000 0.018 Tech IPOs Total News 22.915 15.509 0.000 7.392 4.984 0.000 0.000 0.000 Net News 2.381 -1.274 0.000 1.588 -0.571 0.000 0.160 0.000 Non-Tech IPOs 47 Total News 21.200 15.181 0.000 6.488 4.029 0.000 0.000 0.000 Net News 3.267 -1.568 0.000 1.865 0.302 0.000 0.000 0.000 VC Backed IPOs Total News 26.931 17.561 0.000 7.557 5.337 0.000 0.000 0.000 Net News 2.406 -2.269 0.000 1.821 -0.585 0.000 0.000 0.080 Non-VC Backed IPOs Total News 13.282 11.241 0.000 6.445 3.770 0.000 0.000 0.000 Net News 3.159 0.316 0.000 1.647 0.259 0.000 0.000 0.767 48 Table 3. The double standard of media coverage This table reports the difference in media coverage between the internet and the non-internet sample firms, following a price increase or decrease from previous date. The news item data is hand-collected from Dow Jones Interactive and Factiva for both the internet and the non-internet sample IPOs. We read and classify them as good, bad or neutral news. Net news is the difference between the number of good news and the number of bad news. RET t-1 is the stock return on day t-1 for daily analysis, and month t-1 for monthly analysis. k% = 1% for daily news analysis, and 10% for monthly news analysis. Using abnormal returns instead of raw returns yields virtually identical results and is hence omitted. TN t and NN t are the average number of total news and net news (good news – bad news) collected on period t (day or month) per firm conditional on the previous period (t-1) price movement. Nasdaq Market Peak is March 24, 2000, the day when Nasdaq 100 index reaches its highest level during the sample period. Firm Market Cap. Peak is the firm-specific day when a firm achieves the highest market capitalization during the sample period. The last two columns present the p-values based on Satterthwaite standard errors testing the difference in conditional mean number of news between the two samples. (1) (2) (3) (4) (5) (6) Internet IPOs Non-Internet IPOs p value of test for TN t NN t TN t NN t (1)=(3) (2)=(4) Panel A: Entire Period (1996-2000) Daily News RET t-1 >0 0.928 0.072 0.284 0.045 0.000 0.000 RET t-1 <0 0.836 -0.058 0.272 0.013 0.000 0.000 RET t-1 > k% 0.934 0.078 0.288 0.049 0.000 0.000 RET t-1 < -k% 0.838 -0.063 0.273 0.010 0.000 0.000 Monthly News RET t-1 >0 20.237 0.627 5.618 0.569 0.000 0.783 RET t-1 <0 15.904 -0.464 5.334 0.601 0.000 0.000 RET t-1 > k% 21.345 0.714 6.171 0.587 0.000 0.636 RET t-1 < -k% 16.045 -0.665 5.592 0.643 0.000 0.000 Panel B: Sub-Periods Prior to Nasdaq Market Peak (January, 1st, 1996 to March 24, 2000) Daily News RET t-1 >0 1.126 0.186 0.325 0.089 0.000 0.000 RET t-1 <0 1.006 0.026 0.310 0.050 0.000 0.002 RET t-1 > k% 1.141 0.200 0.334 0.094 0.000 0.000 RET t-1 < -k% 1.015 0.026 0.310 0.047 0.000 0.014 Monthly News RET t-1 >0 23.364 2.028 6.030 1.307 0.000 0.014 RET t-1 <0 19.017 2.298 5.978 1.513 0.000 0.002 RET t-1 > k% 24.364 2.049 6.740 1.365 0.000 0.047 RET t-1 < -k% 20.029 2.424 6.563 1.676 0.000 0.019 Post Nasdaq Market Peak (March 24, 2000 to December 31st, 2000) Daily News RET t-1 >0 0.681 -0.071 0.240 -0.001 0.000 0.000 RET t-1 <0 0.659 -0.146 0.237 -0.022 0.000 0.000 RET t-1 > k% 0.683 -0.072 0.240 0.003 0.000 0.000 RET t-1 < -k% 0.663 -0.151 0.240 -0.022 0.000 0.000 Monthly News RET t-1 >0 15.034 -1.703 5.130 -0.304 0.000 0.000 RET t-1 <0 13.614 -2.496 4.818 -0.130 0.000 0.000 RET t-1 > k% 16.107 -1.602 5.550 -0.262 0.000 0.000 RET t-1 < -k% 13.470 -2.662 4.929 -0.063 0.000 0.000 49 Panel C: Sub-Periods Prior to Firm Market Cap. Peak Daily News RET t-1 >0 1.183 0.217 0.364 0.108 0.000 0.000 RET t-1 <0 1.077 0.057 0.364 0.074 0.000 0.093 RET t-1 > k% 1.201 0.234 0.378 0.116 0.000 0.000 RET t-1 < -k% 1.093 0.061 0.375 0.075 0.000 0.198 Monthly News RET t-1 >0 25.159 2.229 6.640 1.342 0.000 0.014 RET t-1 <0 18.668 2.949 7.191 2.066 0.000 0.003 RET t-1 > k% 26.162 2.322 7.230 1.223 0.000 0.008 RET t-1 < -k% 19.437 3.073 7.781 2.298 0.000 0.035 Post Firm Market Cap. Peak Daily News RET t-1 >0 0.765 -0.021 0.219 -0.006 0.000 0.029 RET t-1 <0 0.725 -0.111 0.216 -0.024 0.000 0.000 RET t-1 > k% 0.766 -0.021 0.218 -0.004 0.000 0.019 RET t-1 < -k% 0.726 -0.117 0.215 -0.026 0.000 0.000 Monthly News RET t-1 >0 16.364 -0.633 4.692 -0.132 0.000 0.040 RET t-1 <0 14.886 -1.721 4.390 -0.143 0.000 0.000 RET t-1 > k% 17.138 -0.690 5.156 -0.023 0.000 0.029 RET t-1 < -k% 14.915 -1.910 4.561 -0.138 0.000 0.000 50 Table 4. The explanatory power of the media in explaining abnormal returns The news item data is hand-collected from Dow Jones Interactive and Factiva for both the internet and the non-internet IPO samples. We read and classify them as good, bad or neutral news. Net news is the difference between the number of good news and the number of bad news. Regressions are estimated for the internet IPO sample and the non-internet IPO sample during the bubble period and post bubble period respectively. Table 4 reports R 2 s obtained from estimating regression models with and without news variable. The dependent variable in all the regressions is the daily abnormal return from fitting a Fama-French 3-factor model for each firm for the combined sample firms. To obtain R 2 s without news effect, only ABRET t-1 , Bid-Ask t-1 , log(1+Vol t-1 ) are included as independent variables. ABRET t-1 is the abnormal return at day t-1. Bid- Ask t-1 is the bid-ask spread and Vol t-1 is the number of shares traded at day t-1. To calculate R 2 s with news effect, news variables NN t-1 , GN t-1 , and BN t-1 are also included respectively. NN t-1 , GN t-1 , and BN t-1 are the number of net news, good news, and bad news per firm at day t-1. The news variables (NN t-1 , GN t-1 and BN t-1 ) and control variables Bid-Ask t-1 and log(1+Vol t- 1 ) are differenced to remove firm fixed effects. Internet IPOs Non-Internet IPOs pre-peak post-peak pre-peak post-peak R 2 with news variables excluded 0.0034 0.0051 0.0019 0.0018 R 2 with news variable NN t-1 included 0.0038 0.0062 0.0030 0.0026 R 2 with news variables GN t-1 and BN t-1 included 0.0038 0.0062 0.0035 0.0029 51 Table 5. The marginal impact of the media on abnormal returns The news item data is hand-collected from Dow Jones Interactive and Factiva for both the internet and the non-internet IPO samples. We read and classify them as good, bad or neutral news. Net news is the difference between the number of good news and the number of bad news. The dependent variable is the daily abnormal return from fitting a Fama-French 3-factor model for each firm for the combined sample firms. NN t-1 , GN t-1 , BN t-1 and TN t-1 are the number of net news, good news, bad news, and total news per firm at day t-1. PostPeak is a dummy variable equal to 1 if it is after March 24, 2000, and 0 otherwise. ABRET t-1 is the abnormal return at day t-1. Bid-Ask t-1 is the bid-ask spread and Vol t-1 is the number of shares traded at day t-1. The news variables (NN t-1 , GN t-1 and BN t-1 ) and control variables Bid-Ask t-1 and log(1+Vol t-1 ) are differenced to remove firm fixed effect. p-values based on Newey-West standard errors are reported in parentheses. ***, **, * = significant at 1%, 5% and 10% (two-sided), respectively. Daily Abnormal Returns Internet IPOs Non-Internet IPOs Internet IPOs Non-Internet IPOs Difference in coefficients between Internet and Non- Internet IPOs (1) (2) (3) (4) pre-peak post-peak Intercept 0.253*** 0.087*** 0.267*** 0.086*** (0.00) (0.00) (0.00) (0.00) Variables of Interests NN t-1 0.090*** 0.228*** -0.137*** -0.078 (0.00) (0.00) (0.00) (0.18) NN t-1 ×PostPeak 0.125*** 0.066 (0.00) (0.27) GN t-1 0.119*** 0.219*** -0.100* 0.070 (0.00) (0.00) (0.08) (0.35) GN t-1 ×PostPeak 0.122*** -0.048 (0.01) (0.55) BN t-1 -0.037 -0.242*** 0.205*** 0.226*** (0.27) (0.00) (0.00) (0.01) BN t-1 ×PostPeak -0.160*** -0.181** (0.01) (0.04) Control Variables log(1+TN t-1 ) -0.316*** -0.142* -0.352*** -0.131 (0.00) (0.07) (0.00) (0.1) log(1+TN t-1 )×PostPeak 0.179** -0.165 0.195** -0.049 (0.03) (0.18) (0.02) (0.71) PostPeak -0.364*** -0.120*** -0.373*** -0.136*** (0.00) (0.00) (0.00) (0.00) ABRET t-1 0.036*** -0.013 0.036*** -0.013 (0.00) (0.14) (0.00) (0.14) ABRET t-1 ×PostPeak -0.105*** -0.030*** -0.105*** -0.029*** (0.00) (0.01) (0.00) (0.01) Bid-Ask t-1 0.055*** 0.047*** 0.056*** 0.047*** (0.00) (0.00) (0.00) (0.00) Bid-Ask t-1 ×PostPeak -0.011 -0.045** -0.011 -0.045** (0.71) (0.02) (0.70) (0.02) log(1+Vol t-1 ) 0.211*** 0.066*** 0.204*** 0.066*** (0.00) (0.00) (0.00) (0.00) log(1+Vol t-1 )×PostPeak -0.066 0.004 -0.061 0.007 52 (0.16) (0.89) (0.19) (0.81) R-squared 0.006 0.003 0.006 0.003 Number of observations 157,094 139,651 157,094 139,651 53 Table 6. Portfolio analysis This table reports portfolio analyses based on the internet sample and the non-internet sample. The dependent variable is the daily internet (non-internet) portfolio return constructed both equally-weighted (EW) and value-weighted (VW). NN t-1 , GN t-1 , BN t-1 and TN t-1 are the average number of net news, good news, bad news, and total news per firm in the portfolio at day t-1. PostPeak is a dummy variable equal to 1 if it is after March 24, 2000, and 0 otherwise. ABRET t-1 is the abnormal portfolio return at day t-1. p-values based on Newey-West standard errors are reported in parentheses. ***, **, * = significant at 1%, 5% and 10% (two-sided), respectively. Internet IPOs Non-Internet IPOs Difference in coefficients between Internet and Non-Internet IPOs Pre-Peak Post-Peak Daily Abnormal Portfolio Returns EW VW EW VW EW VW EW VW EW VW EW VW Intercept 0.001 0.000 0.002 0.003 0.001 0.001 0.000 0.001 (0.45) (0.68) (0.19) (0.18) (0.68) (0.70) (0.97) (0.68) Variables of Interests NN t-1 0.005 0.002 0.011*** 0.006 -0.006 -0.004 0.035* 0.039 (0.12) (0.70) (0.01) (0.23) (0.22) (0.50) (0.07) (0.11) NN t-1 ×PostPeak 0.039*** 0.038** -0.002 -0.005 (0.00) (0.02) (0.91) (0.79) GN t-1 0.004 -0.000 0.012*** 0.005 -0.008 -0.006 0.083*** 0.072** (0.24) (0.95) (0.01) (0.34) (0.14) (0.43) (0.00) (0.02) GN t-1 ×PostPeak 0.058*** 0.055*** -0.033* -0.023 (0.00) (0.01) (0.09) (0.32) BN t-1 -0.008* -0.008 -0.008 -0.007 0.000 -0.001 0.033 0.006 (0.07) (0.23) (0.14) (0.23) (0.99) (0.94) (0.27) (0.87) BN t-1 ×PostPeak -0.019 -0.019 -0.053** -0.025 (0.23) (0.36) (0.05) (0.41) Control Variables log(1+TN t-1 ) 0.001** 0.002** 0.002** 0.003*** 0.001 0.000 0.000 0.001 54 (0.04) (0.04) (0.02) (0.01) (0.41) (0.55) (0.55) (0.54) log(1+ TN t-1 )×PostPeak -0.007 -0.008 -0.012** -0.013* -0.003 -0.003 0.001 -0.001 (0.14) (0.19) (0.04) (0.07) (0.20) (0.24) (0.80) (0.84) PostPeak -0.002 0.001 -0.014* -0.009 -0.005 -0.003 0.009 0.005 (0.38) (0.69) (0.07) (0.24) (0.19) (0.43) (0.21) (0.53) ABRET t-1 0.127*** 0.056*** 0.126*** 0.054 -0.043 0.022 -0.043 0.023 (0.01) (0.22) (0.01) (0.23) (0.39) (0.57) (0.38) (0.56) ABRET t-1 ×PostPeak -0.108 -0.123 -0.111 -0.120 0.361*** 0.168** 0.354*** 0.164** (0.28) (0.22) (0.26) (0.23) (0.00) (0.02) (0.00) (0.03) R-squared 0.046 0.012 0.049 0.015 0.034 0.013 0.041 0.015 Number of observations 1,203 1,203 1,203 1,203 1,182 1,182 1,182 1,182 55 Table 7. Agreement in news classification Seven undergraduates read and classified one hundred news items into good, bad, or neutral news. Four undergraduates (U1-U4) read one hundred pieces of news from Sapient, Inc., a control firm; three undergraduates (U5-U7) read one hundred pieces of news from Yahoo!, an internet firm. This table shows the pairwise incidence of agreement between individuals; that is, the percent of one hundred news items to which two individuals both assign a value of good, bad, or neutral. Panel A: Sapient Panel B: Yahoo! Author 2 U1 U2 U3 U4 Author 2 U5 U6 U7 Author 1 71.00% 84.00% 71.00% 53.00% 70.00% 65.00% 44.00% 56.00% 71.00% Author 2 71.00% 60.00% 48.00% 77.00% 36.00% 38.00% 52.00% U1 67.00% 55.00% 72.00% U2 48.00% 70.00% U3 52.0% U4 U5 55.00% 50.00% U6 55.00%