A Textual Analysis of Logograms in Chinese IPO Roadshows: How Agreement between Investors and Management Relates to Pricing and Performance

Abstract

We analyze the interaction between management and investors during Chinese IPO roadshows through Jaccard Similarity analysis of written Chinese logograms. We provide evidence that when agreement is high, investor optimism increases, leading to relatively large first-day underpricing. We further show that high agreement biases investors to systematically overestimate IPO prospects leading to poor long-run abnormal performance. Jaccard Similarity is different from current content analysis methodologies because it is language and culture agnostic, requiring no a priori construction of thematic dictionaries. Elimination of such dictionaries removes the danger that the researcher has imposed predispositions upon the study.

1. Introduction

The path-breaking work of Loughran and McDonald (2011, 2013, 2014) has legitimized the use of content analysis in financial economics research. The primary method to conduct content analysis is to use a pre-existing dictionary such as the Harvard Psychosociological Dictionary or DICTION, a popular textual analysis software program. (e.g., Das and Chen (2007) and Tetlock (2007) use the Harvard Dictionary; Ober et al. (1999) use DICTION). If a pre-existing dictionary is not used, then a customized dictionary is created as in Loughran and McDonald (2011), Brau et al. (2016, 2021) or Afanasyev et al. (2021). These customized dictionaries are argued to add more specificity to the content analysis methodology in financial economics. However, the construction of a researcher-customized dictionary potentially adds researcher bias to the analysis (Benson et al. 2015). In this paper we present an approach to remove such bias.

Prior to presenting our contribution to the literature, it is instructive to position our paper vis-à-vis prior IPO literature. The seminal articles, and many of those that have followed, explored the notion of three IPO phenomena. Namely, IPO underpricing (e.g., Brau et al. 2008); IPO long-run performance (e.g., Brau et al. 2007, 2012; Boubaker et al. 2017, 2020; Cao-Alvira and Rodríguez 2017); and the IPO hot markets cycle (e.g., Brau and Fawcett 2006a, 2006b).

Along with these three original anomalies of IPO study, the field has branched into: the costs of IPOs (Ang and Brau 2002); IPO lockups (Brau et al. 2004, 2005); SB IPOs (Brau and Carpenter 2012a, 2012b, 2013); real estate IPOs (Brau and Heywood 2008; Brau et al. 2013); healthcare IPOs (Brau and Holloway 2009; Brau and Carpenter 2017; Adcock et al. 2020); IPO surveys (Brau and Fawcett 2006a, 2006b; Brau et al. 2006); IPO harvesting (Brau et al. 2003, 2010); SCOR IPOs (Brau and Osteryoung 2001; Brau and Gee 2010); international IPOs (Brau and Rodríguez 2009; Brogi et al. 2020); why firms do IPOs (Ang and Brau 2003; Brau 2012); and earnings management in IPOs (Brau and Johnson 2009).

The great majority of this prior IPO research uses quantitative data, and neglects any potential information contained in the text of roadshows. As discussed above, those authors who do study soft information through textual or content analysis have typically explored the offering prospectus S-1 documents (e.g., Loughran and McDonald 2013). As explained below, we extend the research along two primary fronts.

First, we contribute to the prior IPO literature by performing content analysis on Chinese IPO roadshows. Second, our application is a method that does not require a dictionary or word list. We operationalize Jaccard Similarity, which in our application measures linguistic convergence between management and potential investors during the Q&A portion of the roadshow. The method of Jaccard similarity has been established in the English finance literature by Cohen et al. (2020). The efficacy of the Jaccard similarity approach in our current study is that it is language neutral.

We believe we are the first to apply it to native Chinese writing (logograms) and the first to study IPO roadshows with it. We apply fundamental tenets of communication accommodation theory to interpret increased language similarities as increased levels of agreement and positive interaction between managers and investors. When managers and investors are in agreement, they tend to adopt each other’s word choices and communication patterns, i.e., they form communicative unity. We argue that our Jaccard similarity measurement serves as an effective metric to measure this positive interaction between the parties. Throughout this paper, we will describe the positive interaction between managers and investors, and the resulting linguistic convergence, as agreement.

Our textual analysis work fits into several strands of prior readability literature. The Jaccard/Cosine similarity approach originates in the natural language processing literature (Manning and Schütze 1999), the IPO literature (Hanley and Hoberg 2010) and the conference call literature (Cicon 2017). Specifically, Hanley and Hoberg (2010) use cosine similarity to compare information content among pairs of matched IPOs. Cicon (2017) uses the Jaccard Cosine Similarity approach to analyze information production in corporate conference calls. Cicon (2017) shows a direct positive relationship between information production and firm performance. Other textual analysis methodologies include the readability approach used by Boubaker et al. (2019). Their approach measures the ‘difficulty level’ required to read particular documents by looking at length of words, number of syllables per words, and number of words per sentence.

In this paper we significantly extend the above literature by providing a methodology which permits researchers to study text written in languages in which they have little to no proficiency. So, whereas Cicon (2017) and Boubaker et al. (2019) study corporate documents written in their native language, in this paper, the English speaking authors develop a technique to permit them to study corporate documents written in Mandarin Chinese, of which we have little to no proficiency. This approach gains particular power when it is used by researchers, who use a phonological writing system. (e.g., English, which uses an alphabet of symbols to represent distinct spoken sounds (consonants and vowels) which are stringed together to represent words and phrases), to analyze written text based on logograms (e.g., Mandarin, which uses symbols representing entire words and phrases).

Our results find that agreement matters because it is priced. In our study, the top 20% of firms (by agreement) have mean underpricing of 105.08%. On the other hand, the bottom 20% of firms have mean underpricing of 34.85%. The difference of means between the two is statistically significant at the 1% level. Our empirical evidence suggests that agreement indicates strong demand at the time of the IPO (via underpricing) followed by disappointing long-run performance (via returns) indicating initial overestimation of firm value. Our study suggests that disagreement during the roadshow may help price IPOs more accurately, at least in China.

The remainder of the paper proceeds as follows: in the next section we analyze the extant economic literature regarding IPO pricing in the short and long-run. We then review the literature on content analysis and specifically how it has been applied in the IPO financial economics literature. After the literature review, we develop our testable hypotheses in Section 3. Section 4 provides a detailed explanation of the method, as it is a relatively new procedure for finance. Section 5 describes the data and provides summary statistics. The empirical results and robustness tests are discussed and presented in Section 6 and Section 7. The final section summarizes and concludes, to include implications, limitations, and future research avenues.

2. Literature Review

This review consists of three main sections: asset pricing, theory of communication accommodation, and empirical methodology. The pricing of IPOs in both the short term and long term has been studied extensively (e.g., the seminal work of Stoll and Curley (1970), Logue (1973) and Ibbotson (1975) on IPO underpricing and Ritter (1991) and Loughran and Ritter (1995) on long-run IPO returns). In general, IPOs have been shown to experience positive first-day returns, known as underpricing, and negative abnormal long-run returns (Bhabra and Pettway 2003). Communication accommodation has been widely studied in the fields of sociology and communication (Fischer 1958; Giles et al. 1973, 1987, 1991; Beebe and Giles 1984; Thakerar et al. 1982; Triandis 1960), and is a leading theory for how social dynamics affect communication. Textual or content analysis (to include Jaccard similarity) and its implications has been established in the finance literature fairly recently, so a summary in this section is appropriate. Extending Akerlof (1970), one can conceptualize a reduction in information asymmetry as the relationship between managers and investors improves and agreement increases. However, when selling securities on the primary market, homogeneous probabilities and information sets between investors and management motivates higher underpricing. The higher underpricing when agreement is high is tied to the fact that managers have an incentive to provide their primary investors with better initial returns. When agreement is high, management teams tend to provide a higher “discount” to their primary investors where those investors have more accurate expectations of the fundamental value of the company due to less information asymmetry through higher agreement.

A voluminous academic literature documents and attempts to explain why most IPOs end the first day of trading with a stock price that is significantly greater than the offer price that was sold to the primary market that same day. Underpricing is measured by most researchers as:

Underpricing = (1st Day Closing Price/IPO Offer Price) − 1

(1)

Over the period of 1960–2016, average underpricing in the US has been 16.76%, and for our sample period of 2004–2012, average underpricing has been 12.7%1. This underpricing represents “money left on the table” by the issuing firms (Loughran and Ritter 2002). Through the years many hypotheses have been proposed to explain why underpricing is observed and why it persists.

Theoretical models, such as Baron (1982) and Rock (1986), began explaining IPO underpricing in the context of asymmetric information (see also Li et al. (2005) and Boulton et al. (2021) for more recent treatments of information asymmetry in IPOs). Early empirical explanations of this anomaly have focused on quantifiable market and financial statement data (Beatty and Ritter 1986; Hanley 1993; Brau et al. 2014). Fedorova et al. (2022) demonstrate the influence of behavioral characteristics of management on IPO pricing. Finally, Brau et al. (2016) focused on the pricing of soft information on IPO pricing using a strategic word dictionary. They show that soft information exhibits a positive correlation to underpricing and a negative correlation to long-run stock performance. Specifically, IPOs with greater (less) usage of positive (negative) strategic words are correlated with greater (less) underpricing and better (worse) long-run performance. Communication Accommodation Theory (CAT) is a generalized extension of Speech Accommodation Theory, which encompasses all forms of communication. CAT has been largely supported by theoretical and empirical studies involving interpersonal (Giles et al. 1973; Giles 2016; Muir et al. 2016) and multicultural communication behaviors (Hajek et al. 2008; Simard et al. 1976). We are not the first to use similarity methods in measuring communication accommodation (Jones et al. 2014; Sagi and Diermeier 2017).

According to CAT, when two parties engage in positive social interactions (i.e., agreement) they tend to converge their communication style (word choice, intonation, phrasing, volume, etc.) in order to unify themselves with the other party (Giles et al. 1987; Kulesza et al. 2014). Conversely, insufficient accommodation or intentional non-accommodation can result in greater levels of disunity and conflict between two parties (Gasiorek and Giles 2012; Giles and Gasiorek 2013).

Empirical data supports the application of CAT and its subfields (Speech Accommodation Theory, Linguistic Style Accommodation, etc.) to negotiations (Sagi and Diermeier 2017; Swaab et al. 2011; Taylor and Thomas 2008) and online-multiparty interactions (Huffaker et al. 2011; Ludwig et al. 2014; Muir et al. 2017).

As seen in empirical studies, convergence in word choice reflects positive interactions and results in greater levels of unity and agreement (Kulesza et al. 2014; Linnemann and Jucks 2016; Muir et al. 2016, 2017; Sagi and Diermeier 2017; Van Baaren et al. 2003). We extend the communicative predictions, assumptions, and explanations of CAT to the financial negotiations during IPO roadshows. This approach allows us to quantify the quality of agreement between investors and firms by measuring the similarity in their word choices (using Jaccard (cosine) similarity methods) rather than using a dictionary of positive/negative terms which may unintentionally include cultural and language bias. Jaccard (cosine) similarity originates in the natural language processing literature (Manning and Schütze 1999) the information processing literature (Kwon and Lee 2003) and the finance and IPO literature (Hanley and Hoberg 2010). Specifically, Hanley and Hoberg (2010) use cosine similarity to compare content among pairs of matched IPOs. We use this approach in our paper because it enables us to study documents written in languages for which we have no reading or writing proficiency. This approach gains particular power when it is used by researchers, who use a phonological writing system (e.g., English which uses an alphabet of symbols to represent distinct spoken sounds, i.e., consonants and vowels, which are stringed together to represent words and phrases), to analyze written text based on logograms (e.g., Mandarin which uses symbols representing entire words and phrases).

We generalize linguistic convergence (similarity of word choice) as an indicator that communication between the two parties is positive and unifying. For simplicity, we define positive interactions during negotiations as agreement, i.e., the greater the linguistic convergence the greater the agreement between managers and investors.

3. Hypothesis Development

We use Jaccard similarity to measure agreement between management and investors during the IPO roadshow. Similarity in this context is consistent with a wide array of literature which relates linguistic decisions to interaction quality. We believe a plausible explanation is that managers do their best to match the style and diction of investors’ questions in their responses in order to improve interpersonal capital and perception (Jenkins and Owen 2021a, 2021b). Swaab et al. (2011) find that linguistic mimicry improved negotiation outcomes by improving interpersonal capital. Huffaker et al. (2011) find that linguistic convergence increased agreement between potential coalition partners in multiparty negotiations through improvements to interpersonal rapport. Kulesza et al. (2014) find that the repetition of a counterpart’s words increases their ability to elicit prosocial behavior from them. Sagi and Diermeier (2017) use latent semantic analysis similarity to measure the change in language similarity among negotiators and find that an increase corresponds to greater agreement among the parties. We should therefore expect similarity to proxy for agreement and perception in as much as it captures these linguistic strategies. Furthermore, we do not believe that positive perception will go unnoticed since Blankespoor et al. (2017) find that investor perception of CEOs is positively associated with price. We hypothesize that investors quiz management until they have sufficient information to make their investment decision. When the interaction between managers and investors is positive, we hypothesize that investors reach a satisfactory state more quickly and begin reaffirming previously stated information and providing feedback.

Langberg and Sivaramakrishnan (2010) suggest that this feedback may take the form of verbal praise in which the investor affirms that the information, provided by management, is valuable to them. In such a situation, the analyst is likely to repeat words and phrases that management used previously. The more quickly management and investors reach agreement the more time they will spend accommodating and converging their speech.

Likewise, if responses from management are dissatisfactory, investors will pose additional questions until satisfied. Interactions that struggle to become satisfactory yield less reaffirmation, feedback, and accommodation.

Our algorithms identify these situations by comparing the Chinese logograms used by managers to those used by investors. When both participants are using similar logograms, we say that there is agreement between manager and investor. We hypothesize two consequences to manager and investor agreement during a roadshow:

Hypothesis 1.

When investor and manager agreement is high during a roadshow, first day underperformance will be high.

Hypothesis 2.

When investor and manager agreement is high during a roadshow, subsequent long-term performance will be poor (investor over-optimism).

The first hypothesis is rooted in the notion that greater agreement indicates greater demand, which then manifests as greater underpricing (i.e., greater first-day return). The second hypothesis is founded on studies that show other types of over-optimism in IPOs (e.g., earnings management as in Teoh et al. 1998) lead to poor long-run performance. Empirically, Brau et al. (2016) show that greater first-day returns contribute to poor long-run underperformance as the stock price adjusts back to pricing equilibrium.

4. Textual Modeling Methodology

Measuring Agreement

Following Cohen et al. (2020), the Jaccard similarity measure is defined as the size of the intersection divided by the size of the union of two sets:

$$Jaccardsimilarity=\frac{\left|Q^{\prime }\cap A^{\prime }\right|}{\left|Q^{\prime }\cup A^{\prime }\right|}$$

(2)

We extend this methodology to Chinese logograms, first parsing each roadshow transcript call into two separate documents. The first document contains only those comments made by management during the roadshow. The second document contains only those comments made by the investors during the roadshow.

We serialize the two documents into computer memory, Chinese character by Chinese character, to create logogram vector representations of each document. We remove non-informative logograms (Chinese equivalents of “the,” “a,” “is,” etc.). Each vector contains an element for each unique logogram in the section, where the element is a scalar count of the number of times that logogram appears in the session. We compute the Jaccard similarity between these vectors to determine the degree to which the two vectors are alike. A more rigorous analysis and application of this methodology can be found in Cicon (2017).

5. Data Description

The IPO companies are from the SME and ChiNext in ShenZhen Stock Exchange. Companies listed on the SME or ChiNext hold net roadshows at http://rsc.p5w.net/, (accessed on 7 February 2021) which allows us to use Python to web scrape the releases.

Our initial sample range is from 2004 and 2012. We end the sample of IPOs in 2012 because there are no IPOs in 2013, and China’s Securities Regulatory Commission stipulated an IPO’s first-day return cannot exceed 44% after January 1, 2014. See Appendix A.1, Appendix A.2, Appendix A.3 and Appendix A.4 for examples of interviews from which our textual analysis data is gleaned.

5.1. Dependent Variables

Our first hypothesis uses the dependent variable of IPO underpricing. The common definition uses First-day Return (Underpricing) as the IPO initial return, which is the percentage change in price from the IPO offer price to the closing price on the first day of trading. Our second hypothesis studies the impact of Agreement on Cumulative Abnormal Returns (CARs) relative to the ShenZhen index over a 1-, 2-, and 3-year horizon. The definition of the dependent variable here is CAR with the respective year.

We define CAR as:

AR_i,t = R_i,t − E[R_i,t|SZ_t]

(3)

CAR_i,t+k = Σ_kAR_i,t+k

(4)

where AR is the abnormal return for firm i on day t, R is the observed return of the firm on that day, and E[] is the expected return of the firm that day (R) adjusted for the return on the ShenZhen market that day (SZ), CAR is then the sum of the ARs for k equaling one-, two-, or three-year horizons for firm i.

5.2. Independent Variables

We use variables that have been shown in the extant literature to impact underpricing, along with our main variable of interest, Agreement. Following Borochin et al. (2018) we define each variable as:

Agreement = Jaccard similarity of the investor questions and management answers on the IPO Net Roadshow.

Tone of Question = Based on the NTUSD word list. The NTUSD word list is from Taipei university and is a general word list appropriate for native Chinese logograms.

Tone of Answer = Based on the NTUSD word list. The NTUSD word list is from Taipei University and is a general word list appropriate for native Chinese logograms.

Age = The age of the firm in years at the time it goes public.

Size = Total assets logged in the year before IPO.

ROE = Return on equity in the year before going public.

SOE = Dummy equaling one if state-owned and zero otherwise.

Lottery rate = The number of shares issued/the number of subscripted shares.

Offerprice = IPO offer price.

Offer_PE = IPO offer price/Earnings.

IPO Proceeds = The total number of shares offered in the IPO multiplied by price.

IPO Fees/IPO Proceeds = IPO Fees divided by IPO Offering Proceeds.

Primary Shares Offered = The number of shares offered where proceeds go to the firm.

Underwriter Prestige = Dummy equaling one if the underwriter is top 10.

Prior 3-month IPO volume = Sum of gross IPOs in the prior three months.

Average Prior 3-month Underpricing = Leading 3-month average underpricing IPOs.

Three-week Prior Market Return = ShenZhen stock market return for the 15 trading days before going public.

Fixed Effects = Controls for the year and industry of each offer.

5.3. Data Descriptive Statistics

Table 1. Distribution of Sample by Industry and by Year. This Table reports the distribution of the sample by industry and by year. The full sample consists of 980 firm observations drawn from the SME and ChiNext on the ShenZhen Stock Exchange. The sample spans 2004 to 2012. The industry distributions are reported in Panel A and annual distributions are reported in Panel B.

Panel A: Industry Distribution
Industry Classification	Number of Firms	Percent of Sample
Agriculture, Forestry, and Fishing	16	1.63%
Business Service	15	1.53%
Construction	26	2.65%
Culture Communication	11	1.12%
Environment	11	1.12%
Finance	5	0.51%
Health	4	0.41%
Information	105	10.71%
Lodging, Catering	2	0.20%
Manufacturing	715	72.96%
Mining	10	1.02%
Real Estate	9	0.92%
Research Service	9	0.92%
Transportation, Storage	10	1.02%
Water, Electric, Gas	8	0.82%
Wholesale Trade, Retail Trade	24	2.45%
Total	980	100%
Panel B: Annual Distribution
IPO Year	Number of Firms	Percent of Sample
2004	33	3.37%
2006	47	4.80%
2007	97	9.90%
2008	65	6.63%
2009	89	9.08%
2010	316	32.24%
2011	230	23.47%
2012	103	10.51%
Total	980	100%

reports the sample distribution by industry and by year. The complete sample is composed of 980 firm observations drawn from the SME and ChiNext on the ShenZhen Stock Exchange, from 2004–2012.

As expected for the Chinese economy, manufacturing dominates the industry count with 715 IPOs (72.96% of the sample). The second most populous industry is information (n = 105, 10.71%). Every other industry has less than 3% representation and is fairly evenly distributed. The annual distribution of listings shows a low of 33 in 2004, following the global debt crisis, and a high of 316 in 2010.

The summary statistics for the sample are reported in

Table 2. Summary Statistics of Sample Variables. This Table reports the summary statistics for our sample of 980 firm observations drawn from the SME and ChiNext on the ShenZhen Stock Exchange. The sample spans 2004 to 2012. Reported are the means, medians, standard deviation, minimum and maximum values for each of the variables used within our analysis. IPO Underpricing is the common definition that uses First-day Return (Underpricing) as the IPO initial return, which is the percentage change in price from the IPO offer price to the closing price on the first day of trading. Agreement is the Jaccard similarity of investor questions and management answers on the IPO Net Roadshow. Tone of Question = Based on the NTUSD word list. The NTUSD word list is from Taipei university and is a general word list appropriate for native Chinese logograms. Tone of Answer is based on the NTUSD word list. The NTUSD word list is from Taipei University and is a general word list appropriate for native Chinese logograms. Age is the age of the firm in years at the time it goes public. Size is the total assets logged in the year before IPO. ROE is defined as the return on equity in the year before going public and SOE is a dummy equal to one if state-owned and zero otherwise. Lottery rate is the number of shares issued/the number of subscripted shares. Offerprice and Offer PE are the IPO offer price and IPO offer price/earnings, respectively. IPO Proceeds represents the total number of shares offered in the IPO multiplied by price. IPO Fees/IPO Proceeds are the IPO Fees divided by IPO Offering Proceeds. Primary Shares Offered are the number of shares offered where proceeds go to the firm. Underwriter Prestige is a dummy equaling one if the underwriter is top 10. Prior three-month IPO volume is the sum of gross IPOs in the prior three months. Average Prior three-month Underpricing is the leading three-month average underpricing IPOs. Lastly, Three-week Prior Market Return is the ShenZhen stock market return for the 15 trading days before going public.

Variable	Mean	Median	Standard Deviation	Minimum	Maximum
IPO Underpricing	0.636	0.375	0.797	−0.263	5.381
Agreement	0.213	0.212	0.034	0.109	0.309
Tone of question	0.582	0.578	0.111	0.162	1.000
Tone of answer	0.800	0.807	0.057	0.558	0.952
Firm Characteristics
Age	2.018	2.079	0.596	0.000	3.296
Size	19.946	19.863	0.782	18.043	24.758
ROE	0.277	0.256	0.126	0.018	1.669
SOE	0.110	0.000	0.313	0.000	1.000
Offer Characteristics
Lottery rate	0.009	0.006	0.025	0.000	0.655
Offer Price	23.641	20.000	15.028	2.880	148.000
Offer PE	45.279	40.930	21.196	6.670	150.820
Offer Proceeds	6.898	5.420	5.550	0.905	59.348
IPO Fees/IPO Proceeds (%)	0.070	0.067	0.024	0.018	0.174
Underwriter Prestige	0.448	0.000	0.498	0.000	1.000
	60.210	69.000	25.041	0.000	101.000
Industry Characteristics
Prior 3 Month IPO Volume	60.210	69.000	25.041	0.000	101.000
Prior 3 Months Underpricing	0.634	0.469	0.582	0.000	2.988
3 Week Prior Market Return	−0.000	−0.001	0.070	−0.212	0.212

. The average (median) underpricing for the sample is 63.6% (37.5%). Benchmarked to the US average underpricing of 17.6% mentioned earlier, this level of underpricing is relatively high. The agreement number has an average (median) of 21.3% (21.2%).

6. Empirical Results

In this section we discuss the empirical findings for our two testable hypotheses.

Table 3. Correlations.

	Variables	(1)	(2)	(3)	(4)	(5)	(6)	(7)	(8)	(9)	(10)	(11)	(12)	(13)	(14)	(15)	(16)	(17)
(1)	IPO Underpricing	1
(2)	Agreement	0.313 ***	1
(3)	Tone of question	−0.001	−0.216 ***	1
(4)	Tone of answer	0.059 *	0.238 ***	0.266 ***	1
(5)	Age	−0.144 ***	−0.064 **	−0.012	−0.0230	1
(6)	Size	−0.105 ***	0.038	−0.015	−0.061 *	0.064 **	1
(7)	ROE	−0.126 ***	0.029	−0.075 **	0.020	−0.070 **	−0.072 **	1
(8)	SOE	0.178 ***	0.101 ***	0.014	0.0450	0.007	0.186 ***	−0.153 ***	1
(9)	Lottery rate	−0.190 ***	−0.130 ***	−0.029	−0.029	0.028	0.118 ***	0.204 ***	−0.073 **	1
(10)	Offer Price	−0.372 ***	−0.254 ***	0.109 ***	−0.023	0.062 *	−0.021	0.435 ***	−0.163 ***	0.243 ***	1
(11)	Offer PE	−0.271 ***	−0.348 ***	0.217 ***	−0.040	0.009	−0.202 ***	−0.069 **	−0.125 ***	0.0290	0.597 ***	1
(12)	Offer Proceeds	−0.302 ***	−0.135 ***	0.072 **	−0.018	0.016	0.527 ***	0.366 ***	−0.014	0.247 ***	0.639 ***	0.330 ***	1
(13)	IPO Fees / IPO Proceeds (%)	0.051	−0.015	−0.048	0.001	0.130 ***	−0.435 ***	−0.175 ***	−0.075 **	−0.039	−0.245 ***	−0.120 ***	−0.478 ***	1
(14)	Underwriter Prestige	−0.086 ***	−0.105 ***	0.068 **	0.056 *	−0.002	0.045	0.029	−0.002	−0.013	0.063 **	0.066 **	0.063 **	0.0290	1
(15)	Prior 3 Month IPO Volume	−0.503 ***	−0.387 ***	0.109 ***	−0.131 ***	0.095 ***	0.010	−0.007	−0.147 ***	0.153 ***	0.485 ***	0.575 ***	0.355 ***	−0.0290	0.075 **	1
(16)	Prior 3 Months Underpricing	0.720 ***	0.297 ***	0.070 **	0.096 ***	−0.214 ***	−0.005	−0.078 **	0.143 ***	−0.197 ***	−0.312 ***	−0.258 ***	−0.199 ***	−0.123 ***	−0.030	−0.568 ***	1
(17)	3 Week Prior Market Return	0.285 ***	0.050	0.021	0.038	0.038	−0.025	−0.030	−0.001	−0.036	−0.064 **	−0.016	−0.057 *	−0.018	−0.032	−0.086 ***	0.137 ***	1

This Table reports correlations between the study variables, where ***, ** and * denote significance at 1, 5 and 10 percent levels, respectively. Table acronyms are defined as follows: ROE stands for return on equity over the past year, SOE is a dummy equal to one if state-owned, and Offer PE is the IPO offer price/earnings. Most significant for our research are the correlations between agreement and IPO underpricing, Tone of question, Tone of answer, SOE, lottery rate, offer price, offer PE, offer proceeds, underwriter prestige, prior three-month IPO volume, and prior three-month underpricing that are significant at the 1% level. In fact, only three variables in the study showed correlation with agreement at a significance level less than 10%, namely size, ROE, and three-week prior market return.

reports Spearman correlation coefficients. Most significant for our research are the correlations between agreement and IPO underpricing, Tone of question, Tone of answer, SOE, lottery rate, offer price, offer PE, offer proceeds, underwriter prestige, prior three-month IPO volume, and prior three-month underpricing that are significant at the 1% level. In fact, only three variables in the study showed correlation with agreement at a significance level less than 10%, namely size, ROE, and three-week prior market return.

6.1. Univariate Analysis

We begin with a univariate analysis of the agreement-underpricing relationship. We divide sample IPO firms into five groups based on agreement. Group 1 has the lowest agreement while Group 5 has the highest. We then calculate the average underpricing for each of the five groups and plot the results in Figure 1. Figure 1 shows that the average underpricing in Group 1 is only 34.85%, whereas the average underpricing in Group 5 is 105.08%. The difference between the average underpricing in the two groups is statistically significant (p value < 0.01). The figure also shows that average underpricing increases monotonically when moving from Group 1 to Group 5.

Figure 1 plots the average underpricing for quintiles by Agreement. We divide the sample into quintiles of interaction quality and plot the mean first-day return for each quintile. The x-axis is the respective quintile, and the y-axis is the average underpricing reported as a decimal. The difference between the average underpricing of the lowest and highest quintiles is statistically significant at the 1% level.

In unreported analyses we have conducted t-tests (means) and Wilcoxon nonparametric tests (medians) along with truncation and winsorization. Our results are robust to these tests. We feel our reported multivariate models are superior to pairwise means and medians testing as we can control for confounding effects and mitigate omitted variable bias. Figure 1 expands the statistics reported in Table 2 for underpricing. Next, we explore whether this effect persists in a multivariate framework.

6.2. Agreement and Underpricing

Table 4. Agreement and Underpricing.

Variable	Model 1	Model 2	Model 3	Model 4
Agreement	1.778 ***	2.205 ***	1.941 ***	2.018 ***
	(3.03)	(3.22)	(2.77)	(3.24)
Tone of question		−0.033	0.089	−0.001
		(−0.17)	(0.56)	(−0.01)
Tone of answer		−0.444	−0.336	−0.437 *
		(−1.60)	(−1.36)	(−1.79)
Age		0.022	0.024	0.006
		(0.60)	(0.64)	(0.24)
Size		−0.162 ***	−0.130 ***	−0.119 ***
		(−5.54)	(−3.42)	(−3.38)
ROE		−0.443 **	−0.092	−0.204
		(−2.63)	(−0.67)	(−1.33)
SOE		0.134 **	0.126 *	0.185 ***
		(2.19)	(1.88)	(3.22)
Lottery rate			−0.274	0.199
			(−0.87)	(0.59)
Offer Price			−0.005 ***	−0.004 ***
			(−3.29)	(−3.26)
Offer PE			−0.001	−0.002 *
			(−1.19)	(−1.97)
Offer Proceeds			0.002	0.002
			(0.52)	(0.60)
IPO Fees/IPO Proceeds (%)			2.523 **	2.887 **
			(2.17)	(2.13)
Underwriter Prestige			−0.073 *	−0.059 *
			(−1.86)	(−1.74)
Prior Three Month IPO Volume				−0.005
				(−1.40)
Prior Three Months Underpricing				0.622 ***
				(5.61)
Three Week Prior Market Return				1.998 ***
				(3.24)
Year Fixed Effects	yes	yes	yes	yes
Industry Fixed Effects	yes	yes	yes	yes
Constant	0.318 *	3.773 ***	2.885 ***	2.692 ***
	(1.91)	(7.26)	(3.74)	(3.09)
Observations	980	980	980	980
R-squared	0.516	0.542	0.556	0.629

This Table reports the relationship between ‘Agreement’ and IPO Underpricing. We perform OLS regressions clustering standard errors by year and industry. We report four different models to demonstrate the robustness of our agreement measure and to eliminate concerns of multicollinearity. Model 1 consists of only underpricing and agreement (as well as fixed effects). Model 2 includes tone and firm characteristics. Model 3 adds offer characteristics. Model 4 adds in industry characteristics. Table acronyms are defined as follows: ROE stands for return on equity over the past year, SOE is a dummy equal to one if state-owned, and Offer PE is the IPO offer price/earnings. Statistical significance of differences in means is represented by ‘*’ at the 10% level, ‘**’ at the 5% level, and ‘***’ at the 1% level or better.

is central to our paper because it clearly demonstrates a correlation between agreement and underpricing after adjusting for other factors. We report three different models to demonstrate the robustness of our agreement measure and to eliminate concerns of multicollinearity.

Model 1 shows a significant relationship between agreement and underpricing with no control variables other than year and industry fixed effects. Model 2 performs well with an increase in the goodness of fit, which is expected when additional controls are added. In Model 2 three new variables (size, roe, and SOE) are found to be significant with p value less than 5%. Model 3 also performs well. We see an incremental loss of economic significance when we add offer characteristics, but the loss is not remarkably large. Within our full model, Model 4, the agreement coefficient retains significant economic and statistical significance.

6.3. Agreement and Long-Term Performance

Like Table 4,

Table 5. Agreement and Subsequent Cumulative Abnormal Performance.

Variable	CARs 1 Year	CARs 2 Year	CARs 3 Year
Agreement	−0.768 **	−1.059 ***	−1.135 **
	(−2.20)	(−2.69)	(−2.00)
Tone of question	−0.219 *	−0.015	−0.001
	(−1.83)	(−0.13)	(−0.01)
Tone of answer	0.423 **	0.538 **	1.006 ***
	(2.04)	(2.03)	(3.91)
Age	−0.014	−0.015	0.016
	(−0.73)	(−0.45)	(0.50)
Size	−0.026	−0.054 *	−0.137 ***
	(−0.97)	(−1.77)	(−3.59)
ROE	−0.043	−0.252	−0.287
	(−0.28)	(−1.41)	(−1.41)
SOE	0.049	0.045	0.055
	(1.30)	(0.88)	(0.86)
Lottery rate	0.383	−0.251	−0.380
	(1.44)	(−0.86)	(−1.34)
Offer Price	0.002 *	0.004 **	0.004
	(1.88)	(2.13)	(1.16)
Offer PE	−0.000	−0.002	−0.002
	(−0.19)	(−1.45)	(−1.33)
Offer Proceeds	−0.006	−0.006	−0.004
	(−1.58)	(−1.10)	(−0.59)
IPO Fees/IPO Proceeds (%)	−0.603	0.293	1.335
	(−0.89)	(0.26)	(1.20)
Underwriter Prestige	0.025	0.012	−0.031
	(0.92)	(0.33)	(−0.99)
Prior Three Month IPO Volume	0.002 *	0.000	0.001
	(1.67)	(0.04)	(0.57)
Prior Three Months Underpricing	0.091 **	0.068	0.081
	(2.03)	(1.03)	(0.78)
Three Week Prior Market Return	−0.060	−0.217	0.087
	(−0.25)	(−0.93)	(0.25)
Year Fixed Effects	yes	yes	yes
Industry Fixed Effects	yes	yes	yes
Constant	0.303	0.993	1.881 **
	(0.49)	(1.41)	(2.58)
Observations	980	980	980
R-squared	0.205	0.285	0.320

This Table reports the impact of Agreement on Cumulative Abnormal Returns relative to the ShenZhen index. We perform OLS regressions clustering standard errors by year and industry. Table acronyms are defined as follows: ROE stands for return on equity over the past year, SOE is a dummy equal to one if state-owned, and Offer PE is the IPO offer price/earnings. Statistical significance of differences in means is represented by ‘*’ at the 10% level, ‘**’ at the 5% level, and ‘***’ at the 1% level or better. We winsorize at the 0.1% level.

is central to our thesis. For this data, we used the IPO performance model. The results are displayed for several different periods ranging from one to three years. The post-IPO stock returns are significantly and negatively associated with agreement up to three years after the IPO. Such long-term underperformance by IPOs is common and aligns with the current theory. The economic impact of agreement on the post-IPO returns is also significant. A one standard deviation increase of agreement (0.034) is associated with one-year, two-year, and three- year post IPO abnormal returns of 2.61%, 3.60%, and 3.86%, respectively. This data shows that agreement retains its ability to predict CARs for at least three years. Agreement, therefore, is a key long-term variable currently missing from existing CAR models. The Agreement variable remains significant for Underpricing, BHAR-1, and BHAR-2. The Tone of Answer variable is robust to Underpricing, BHAR-1, and BHAR-3. Control variables also remain robust in each specification.

7. Robustness Checks

7.1. Different Dependent Variables

Following Chen et al. (2015), we calculate market-adjusted underpricing:

Underpricin_g = (P₁ − P₀)/P₀ − (M₁ − M₀)/M₀

(5)

where, P₁ is the closing price of the first trading day, P₀ is the IPO offering price, and M₀ and M₁ are the A-share market index closing prices on the IPO issuing day and the first trading day, respectively. For robustness, we provide

Table 6. Robustness Check: Market Adjusted Underpricing and BHAR.

Variable	Underpricingadj	BHARs 1 Year	BHARs 2 Year	BHARs 3 Year
Agreement	2.078 ***	−0.870 **	−1.087 **	−2.356
	(3.29)	(−2.40)	(−2.03)	(−1.58)
Tone of question	−0.006	−0.283	0.169	−0.171
	(−0.03)	(−1.50)	(0.93)	(−0.39)
Tone of answer	−0.447 *	0.397 *	0.309	2.211 **
	(−1.84)	(1.96)	(0.74)	(2.00)
Age	0.007	−0.010	−0.008	0.036
	(0.26)	(−0.64)	(−0.28)	(0.82)
Size	−0.118 ***	−0.022	−0.034	−0.088
	(−3.33)	(−0.85)	(−0.80)	(−0.75)
ROE	−0.205	−0.076	−0.127	0.044
	(−1.35)	(−0.54)	(−0.62)	(0.07)
SOE	0.194 ***	0.035	0.016	−0.011
	(3.30)	(0.92)	(0.24)	(−0.09)
Lottery rate	0.139	0.268	0.094	−0.667
	(0.39)	(0.98)	(0.24)	(−1.26)
Offer Price	−0.004 ***	0.003 *	0.004 *	0.005
	(−2.99)	(1.68)	(1.79)	(1.13)
Offer PE	−0.002 *	−0.001	−0.002	−0.002
	(−1.80)	(−0.50)	(−1.18)	(−0.98)
Offer Proceeds	0.002	−0.004	−0.004	−0.007
	(0.59)	(−1.15)	(−0.53)	(−0.33)
IPO Fees/IPO Proceeds (%)	2.804 **	−0.967	0.291	3.015
	(2.07)	(−1.53)	(0.20)	(1.26)
Underwriter Prestige	−0.059 *	0.045	0.002	−0.122
	(−1.75)	(1.15)	(0.05)	(−1.41)
Prior 3 Month IPO Volume	−0.005	0.003 **	0.001	0.007
	(−1.35)	(2.02)	(0.32)	(1.27)
Prior 3 Months Underpricing	0.647 ***	0.128 ***	−0.057	−0.099
	(5.60)	(3.40)	(−0.71)	(−0.85)
3 Week Prior Market Return	1.410 **	0.054	−0.070	0.350
	(2.23)	(0.24)	(−0.28)	(0.57)
Year Fixed Effects	yes	yes	yes	yes
Industry Fixed Effects	yes	yes	yes	yes
Constant	2.661 ***	0.238	0.719	0.010
	(3.05)	(0.40)	(0.83)	(0.00)
Observations	980	980	980	980
R-squared	0.615	0.285	0.207	0.167

This Table reports the impact of Agreement on Market Adjusted Underpricing and BHAR. We perform OLS regressions clustering standard errors by year and industry. Table acronyms are defined as follows: ROE stands for return on equity over the past year, SOE is a dummy equal to one if state-owned, and Offer PE is the IPO offer price/earnings. Statistical significance of differences in means is represented by ‘*’ at the 10% level, ‘**’ at the 5% level, and ‘***’ at the 1% level or better. We winsorize at the 0.1% level.

, which is economically equivalent to Table 4.

Following Brau et al. (2016), we use the abnormal buy-and-hold return (BHAR) to measure long-term performance. Again for robustness, we provide Table 6 which is qualitatively similar to Table 5.

We define BHAR as:

BHAR_i,t+k = Π_k(1 + R_i,t+k) − Π_k(1 + SZ_i,t+k)

(6)

where R is the observed return for firm i on day t, and SZ is return on the ShenZhen market index that day. BHAR is computed as the product of one plus R over periods of k of one, two, and three years for firm i minus the product of one plus SZ over the corresponding periods.

7.2. Industry-Adjusted Agreement

In this section we test for industry impacts that could possibly alter our findings as per Ibbotson et al. (1988, 1994) and Alti (2005). To operationalize this robustness test, we adjust the agreement variable by the median agreement of the IPO’s industry at the time of going public.

Table 7. Robustness Check: Industry-adjusted Agreement.

Variable	Underpricing	CARs 1 Year	CARs 2 Year	CARs 3 Year
Agreement	1.793 ***	−0.761 **	−1.018 ***	−1.120 **
	(2.97)	(−2.11)	(−2.82)	(−2.28)
Tone of question	−0.016	−0.220 *	−0.013	−0.001
	(−0.09)	(−1.82)	(−0.12)	(−0.01)
Tone of answer	−0.394 *	0.419 **	0.527 **	0.998 ***
	(−1.67)	(2.03)	(1.99)	(3.97)
Age	0.007	−0.014	−0.015	0.015
	(0.26)	(−0.73)	(−0.46)	(0.49)
Size	−0.119 ***	−0.026	−0.054 *	−0.136 ***
	(−3.35)	(−0.96)	(−1.77)	(−3.60)
ROE	−0.202	−0.043	−0.252	−0.286
	(−1.32)	(−0.28)	(−1.41)	(−1.41)
SOE	0.182 ***	0.050	0.047	0.056
	(3.17)	(1.33)	(0.91)	(0.89)
Lottery rate	0.149	0.395	−0.232	−0.363
	(0.43)	(1.45)	(−0.80)	(−1.30)
Offer Price	−0.004 ***	0.002 *	0.004 **	0.004
	(−3.29)	(1.90)	(2.15)	(1.17)
Offer PE	−0.002 *	−0.000	−0.002	−0.002
	(−1.92)	(−0.20)	(−1.45)	(−1.34)
Offer Proceeds	0.003	−0.006	−0.006	−0.004
	(0.67)	(−1.64)	(−1.15)	(−0.62)
IPO Fees/IPO Proceeds (%)	2.857 **	−0.587	0.313	1.359
	(2.10)	(−0.87)	(0.28)	(1.23)
Underwriter Prestige	−0.061 *	0.025	0.012	−0.031
	(−1.79)	(0.94)	(0.35)	(−0.97)
Prior 3 Month IPO Volume	−0.005	0.002 *	0.000	0.001
	(−1.40)	(1.68)	(0.06)	(0.59)
Prior 3 Months Underpricing	0.622 ***	0.091 **	0.068	0.082
	(5.60)	(2.03)	(1.03)	(0.79)
3 Week Prior Market Return	1.991 ***	−0.058	−0.213	0.091
	(3.23)	(−0.24)	(−0.92)	(0.26)
Year Fixed Effects	yes	yes	yes	yes
Industry Fixed Effects	yes	yes	yes	yes
Constant	3.182 ***	0.105	0.725	1.590 **
	(3.24)	(0.17)	(1.02)	(2.10)
Observations	980	980	980	980
R-squared	0.628	0.204	0.284	0.320

This Table reports the impact of Industry-adjusted Agreement on Underpricing and CAR. We perform OLS regressions clustering standard errors by year and industry. Table acronyms are defined as follows: ROE stands for return on equity over the past year, SOE is a dummy equal to one if state-owned, and Offer PE is the IPO offer price/earnings. Statistical significance of differences in means is represented by ‘*’ at the 10% level, ‘**’ at the 5% level, and ‘***’ at the 1% level or better. We winsorize at the 0.1% level.

provides evidence that the Agreement variable is robust with all four dependent variables—Underpricing, 1-Year CAR, 2-Year Car, and 3-Year CAR. These results are robust to the entire portfolio of control variables as well as year and industry fixed effects. The Agreement variable remains significant for Underpricing, BHAR-1, and BHAR-2. The Tone of Answer variable is robust to Underpricing, BHAR-1, and BHAR-3. Control variables also remain robust in each specification. Thus, the Agreement variable and Tone of Answer variables remain robust. In addition, the relevant control variables continue to hold explanatory power.

8. Conclusions

We find statistically significant correlations between agreement at the Chinese roadshows and key metrics of the resulting IPOs, including underpricing and cumulative abnormal returns. Our research suggests that a company is better able to avoid underpricing its IPO when there is serious discussion and disagreement during the Q&A session of its roadshow. Were this pattern found to be consistent across all roadshows and all languages, a company could strategically encourage disagreement at its roadshow question and answer sessions to price its IPO to minimize underpricing and maximize equity issuance revenue.

The agreement variable also provides an additional metric that both investors and managers can use to predict short- and long-term stock value behavior from as early as the issuer’s IPO roadshow. While agreement cannot be used alone as a crystal ball investing tool, our research suggests that agreement fits into, and improves existing IPO pricing models leading to more precise calculations and therefore reducing money lost due to model error. Agreement is a particularly good indicator of cumulative abnormal returns for at least three years after an IPO is made.

Our findings have practical and policy implications. Regulators could improve original IPO pricing by encouraging or creating policies that promote vigorous debate and discussion pertaining to the firm in question. This increased rigor could lead to less underpricing and more efficient pricing. In addition to initial IPO pricing, improved policy to increase careful and thoughtful discussion during roadshows can lead to better long-term price predictions.

Like all studies, our paper contains limitations to consider. We study only one country, China, with one method, Jaccard Similarity. We cannot say whether our results will extend to other countries, cultures, or languages. As such, further research avenues could include using Jaccard Similarity methods on other languages and in other countries to study agreement in the roadshow. In addition, other methods of agreement could be developed, as well as studying alternative corporate events surrounding IPOs, such as analyst reports and filing prospectuses.

In conclusion, this study has presented empirical evidence that agreement among investors and insiders in IPOs during the roadshow has statistically significant pricing impacts both in the short- and long-term. We apply Jaccard Similarity to the IPO literature as a language-neutral textual analysis tool and demonstrate that soft information that is gleaned from roadshows is not trivial.