A. Data Sample

Our data sample includes all public firms with information on options trading available during the period from 1996 to 2017 as a result of combining different databases. We obtain options data from OptionMetrics, accessed via WRDS. For these firms, we download all 10-K filings during the same period from the EDGAR website maintained by the Securities and Exchange Commission (SEC) and construct textual similarity measures. Finally, we construct firm-level financial variables, such as market capitalization, book-to-market ratio, stock return, return on assets (ROA), cash holdings, capital expenditure (CAPEX), and tangibility, based on items from Compustat. All variable definitions are provided in the Appendix. We keep firm-year observations with information on options trading available both before and after the 10-K release. Our final data sample consists of 4,604 unique firms and 28,704 firm-year observations.

B. Main Variables

Our primary outcome variable is the change in volatility smirk ( $ \Delta Smirk $ ) before and after the 10-K release date. Motivated by prior research, we use the smirk change to capture options trading in response to new information disclosed in the most recent 10-K. Specifically, we first calculate daily volatility smirk as the difference in implied volatilities between OTM put options and ATM call options. Following Xing et al. (Reference Xing, Zhang and Zhao2010), we classify moneyness between 0.95 and 1.05 as ATM call options and moneyness between 0.80 and 0.95 as OTM put options. For stocks with multiple ATM (OTM) options trading on a given day, we use the average of implied volatilities to compute the difference.

Next, we compute changes in volatility smirk as follows: We keep firm-year observations with volatility smirk information available within 3 months both before and after the 10-K release date. We use the first available option volatility smirk, which is at least 8 weeks (i.e., 56 days) before the 10-K release date, as the “Baseline Smirk.” The baseline smirk is defined sufficiently before the 10-K release to alleviate the concern that trading driven by proprietary information and/or other information events (e.g., earnings announcement) may cause the smirk change ahead of the release date. We use the first available option volatility smirk immediately after the 10-K release as the “Post-release Smirk.” The difference between the two smirks reflects changes in options trading surrounding the 10-K release, that is, $ \Delta Smirk=``Post{\textstyle \hbox{-}}release\ Smirk" $ - $ ``Baseline\ Smirk" $ .Footnote ² Alternatively, we use the “Post-release Smirk” ( $ Smirk $ ) as the outcome variable and obtain similar results.Footnote ³

Our main explanatory variable is a textual similarity measure of 10-Ks between two consecutive fiscal years. Following Brown and Tucker (Reference Brown and Tucker2011), Peterson et al. (Reference Peterson, Schmardebeck and Wilks2015), and Cohen et al. (Reference Cohen, Malloy and Nguyen2020), we measure textual changes based on the cosine similarity between the current year’s 10-K and the prior year’s 10-K. A lower such cosine similarity implies more year-to-year textual changes. Suppose that the two 10-Ks have $ n $ unique words,Footnote ⁴ each 10-K is then represented by an $ n $ -dimension vector— $ {v}_1 $ for the current year and $ {v}_2 $ for the previous year:

$$ {v}_1=\left({w}_1,{w}_2,\dots, {w}_n\right),\mathrm{and}\;{v}_2=\left({z}_1,{z}_2,\dots, {z}_n\right), $$

where $ {w}_i $ and $ {z}_i $ are counts of each word $ i\in \left[1,n\right] $ . The cosine similarity score is computed as follows:

$$ Cosine\left(\theta \right)=\left({v}_1/\left\Vert {v}_1\right\Vert \right)\left({v}_2/\left\Vert {v}_2\right\Vert \right)=\left({v}_1\cdot {v}_2\right)/\left(\left\Vert {v}_1\right\Vert \left\Vert {v}_2\right\Vert \right), $$

where $ \theta $ is the angle between $ {v}_1 $ and $ {v}_2 $ , $ \left(\cdot \right) $ is the dot product operator, $ \left\Vert {v}_1\right\Vert $ is the vector length of $ {v}_1 $ , and $ \left\Vert {v}_2\right\Vert $ is the vector length of $ {v}_2 $ . This score is bounded between 0 and 1 with a lower score indicating lower similarity.

We retrieve from WRDS SEC Analytic Suite the cosine measures constructed based on the entire 10-K documents.Footnote ⁵ We use the same method to construct cosine measures for individual items in 10-Ks. Other qualitative measures, such as log file size and tone optimism, are also constructed based on 10-Ks following the method of Loughran and McDonald (Reference Loughran and McDonald2011), (Reference Loughran and McDonald2014). Figure 1 shows the time pattern of average annual report cosine similarity for our sample firms. Year-to-year similarity in 10-K filings exhibits a clear upward trend, consistent with the findings in Brown and Tucker (Reference Brown and Tucker2011) based on a smaller sample.Footnote ⁶

FIGURE 1The Time Pattern of Average Annual Report Similarity

Figure 1 shows the time-series pattern of annual cosine similarity measures averaged across all sample firms. The cosine similarity measure for a firm quantifies textual changes between the current year’s 10-K and the prior year’s 10-K. A lower value implies more year-to-year textual changes.

C. Summary Statistics

Table 1 presents the descriptive statistics for the main variables. Consistent with prior studies (Xing et al. (Reference Xing, Zhang and Zhao2010)), the post-release option volatility smirks are on average positive at around 5%. The year-to-year changes in option volatility smirks exhibit a slightly negative mean (–0.047%) and median (–0.149%), indicating a modest decrease on average after the release of 10-K filings. However, there are considerable variations in smirk changes across firm-years with a high standard deviation of 7.134%. Our primary explanatory variable is the annual report cosine similarity with a mean of 0.971 and a median of 0.988.Footnote ⁷

TABLE 1Descriptive Statistics

We also report two other variables related to 10-K filings. The first is tone optimism, which represents the number of Loughran–McDonald positive words minus negative words scaled by the total number of words in a filing. The second variable is log file size, which denotes the logarithm of the total number of words in a 10-K filing. Overall, the file tone measure is slightly negative, indicating that managers tend to adopt a neutral or slightly pessimistic tone in their 10-K reports.

A. Baseline Results

We use both sorting and Fama–MacBeth regressions to examine the response of option volatility smirks to 10-K textual changes. Sorting relies less on parametric assumptions, and Fama–MacBeth regressions allow us to control for various firm-level characteristics that may impact options trading.

We first sort all firms by the annual report cosine similarity into quintile groups and compare average changes in volatility smirks between the top and bottom quintiles. Note that Q1 (Q5) consists of firms with the most (least) textual changes to 10-Ks. Table 2 presents the sorting results. The average changes for Q1 are positive, suggesting an increase in volatility smirk for firms with the most textual changes. By contrast, we find the opposite for Q5. The equal- and value-weighted spreads in volatility smirk changes are –0.299 and –0.361, respectively. These are around 2–2.5 times the median volatility smirk change and are statistically significant for both equal- and value-weighted portfolios. Hence, these results are consistent with Hypothesis 1 that options traders buy more OTM put options on firms with more textual changes.

TABLE 2The Effect of Textual Changes on Option Volatility Smirk Changes: Cosine-Sorted Portfolios

We next examine the effect of 10-K textual changes on volatility smirks using the following Fama–MacBeth regression model:

(1) $$ \Delta {Smirk}_{it}={\beta}_0+{\beta}_1{Cosine}_{it}+\phi {X}_{it}+{\varepsilon}_{it}, $$

where $ i $ indexes a company and $ t $ indexes a certain fiscal year. $ \Delta Smirk $ is the option volatility smirk right after the 10-K release minus the option volatility smirk at least 8 weeks before the release date.Footnote ⁸ $ Cosine $ , our main variable of interest, is the cosine similarity between the current and previous fiscal year’s 10-Ks. As a lower such cosine similarity implies more textual changes to 10-Ks, which on average contain more negative information regarding future stock returns and firm fundamentals, we expect a negative sign of the estimated coefficient $ {\beta}_1 $ .

The control variables, $ X $ , include log file size, tone optimism, log market cap, book-to-market, 1-year stock return, ROA, cash, CAPEX, and tangibility. All variables are measured at the previous fiscal year-end. Prior research shows that these firm characteristics could also induce options market reactions and are likely correlated with 10-K textual changes (Ertugrul, Lei, Qiu, and Wan (Reference Ertugrul, Lei, Qiu and Wan2017), Cohen et al. (Reference Cohen, Malloy and Nguyen2020)). Large and mature firms tend to have relatively stable operations, lower information asymmetry, and therefore lower downside risk. Firms with high book-to-market and tangible-to-total assets ratios are likely more transparent and easier to value. These firms are, therefore, less likely to experience widening of volatility smirks driven by investors with information advantages. Moreover, cash holdings and investment expenditures may also affect firm risk, though the direction likely depends on investment efficiency (Cheng, Gawande, and Qi (Reference Cheng, Gawande and Qi2021)).

Table 3 presents the Fama–MacBeth regression results. Column 1 includes only the cosine similarity measure, and column 2 adds tone optimism and log file size to control for the effects of other textual information. Column 3 is our main specification and includes firm-level financial variables as additional controls. The estimated effect of textual changes is very similar across different specifications. The coefficient estimates of the cosine measure are all negative and statistically significant at the 1% level. This confirms that more textual changes (proxied by a lower cosine measure) are associated with larger increases in option volatility smirks. Regarding the economic magnitude, if the cosine measure decreases from the 90th percentile (0.998) to the 10th percentile (0.924), the coefficient in column 3 would imply a twofold increase in the change of option volatility smirks relative to the sample median. This magnitude is quite similar to the one obtained in the sorting analysis.

TABLE 3The Effect of Textual Changes on Option Volatility Smirk Changes

As to control variables, most of the firm-level financial variables do not bear a significant relation with changes in option volatility smirks. Consistent with our conjecture, we find that firms with higher book-to-market ratios and more tangible assets tend to have significantly smaller smirk changes. Because the estimated effects of these control variables are very similar across regressions, we omit reporting them in the following tables for brevity.

We conduct several additional tests to ensure that our results are robust to alternative measures of textual changes and econometric specifications. Despite being widely used in measuring textual similarity, the cosine measure is subject to the criticism that it does not have the triangle inequality property and violates the coincidence axiom. To address such concerns, we use three alternative measures of annual report similarity: Jaccard similarity, modified Jaccard similarity, and minimum edit distance. More textual changes to 10-Ks are associated with lower Jaccard or modified Jaccard similarity but higher minimum edit distance. Table A3 in the Supplementary Material presents the Fama–MacBeth regression results, which are consistent with the findings in Table 3. Option volatility smirk changes are negatively related to Jaccard or modified Jaccard similarity and positively related to minimum edit distance. All of the estimated effects are statistically significant at the 1% level. Therefore, our baseline results are robust to alternative measures of textual similarity.

To ensure that our findings are not affected by unobserved industry and time factors, we estimate pooled regressions with industry (4-digit SIC code) and year fixed effects. We cluster standard errors at the firm level. Table A4 in the Supplementary Material reports the estimation results. Consistent with the baseline results in Table 3, more textual changes to 10-Ks are associated with larger option volatility smirk changes. The coefficients of the cosine measure are statistically significant at the 5% level. Moreover, the magnitudes of estimated coefficients are similar to those obtained from Fama–MacBeth regressions.

B. Timing of Changes in Volatility Smirks

To provide further evidence on the timing of changes in volatility smirks, we conduct the Fama–MacBeth regression in equation (1) for each week during the 8-week period before and after the 10-K release. To reduce missing observations, we use the option volatility smirk on the day closest to the 10-K release date in a given week as the weekly measure ( $ {Smirk}_{wt} $ ). Then, we compute the weekly change in option volatility smirk ( $ \Delta {Smirk}_{wt} $ ) as the difference between the weekly smirk ( $ {Smirk}_{wt} $ ) and the baseline smirk measured at least 8 weeks before the 10-K release date ( $ `` Baseline\ Smirk" $ ): $ \Delta {Smirk}_{wt}={Smirk}_{wt}- Baseline\ Smirk $ .Footnote ⁹

Table 4 presents the weekly Fama–MacBeth regression results. Panel A reports the responses in option volatility smirks to textual changes during weeks before and after the 10-K release, respectively. In general, the coefficients of cosine measure have smaller magnitude and lack statistical significance during prerelease weeks. By contrast, the coefficients are much larger in magnitude and statistically significant during each of the first 7 weeks after the 10-K release. This suggests that the significant relation between option volatility smirks and 10-K textual changes is mainly driven by postrelease trading.

TABLE 4The Effect of 10-K Textual Changes on Option Volatility Smirk Changes Before and After Release

Note that the response in option volatility smirks to 10-K textual changes in week [–2] is statistically significant and has a magnitude comparable to the postrelease responses. One interpretation is that some options traders have access to proprietary information before the public release. Although we cannot rule out the possibility of informed trading based on private information, it is difficult to reconcile with why such trading occurs in week [–2] but not in week [–1] or other weeks. A more plausible explanation is that the spike in trading activity stems from a separate information event that coincides with the content later disclosed in the 10-K filing. This is likely because many firms disclose their fourth-quarter earnings approximately 2 weeks before releasing their 10-Ks. As extensively documented in the accounting and finance literature (e.g., Amin and Lee (Reference Amin and Lee1997), Campbell, Ramadorai, and Schwartz (Reference Campbell, Ramadorai and Schwartz2009), Truong and Corrado (Reference Truong and Corrado2014), and Barron, Schneible, and Stevens (Reference Barron, Schneible and Stevens2018)), trading around earnings announcements is often heavy in both the stock and options markets, driven by the anticipation and arrival of new information about a firm’s profitability. The subsequent 10-K filing contains more detailed financial results and price-relevant information that can help investors better understand the earnings result. This drives additional trading during the post 10-K release weeks.

To test this explanation, we exclude firms reporting their fourth quarter earnings within 2 weeks before 10-K releases and re-estimate the weekly Fama–MacBeth regressions. As shown in Panel B of Table 4, the cosine coefficient for week [–2] is now much smaller in magnitude and no longer statistically significant. By contrast, the postrelease cosine coefficients remain large and statistically significant for most weeks. This finding helps address concerns that option volatility smirk changes might be driven by pre-10-K release trading activity. The overall evidence thus supports our hypothesis that options investors gain a competitive advantage through their superior skill in processing public information.

C. Mechanism

Our hypothesis about the relation between textual changes in 10-K filings and changes in option volatility smirks relies on several implicit assumptions. These assumptions are grounded in existing literature on the informational content of 10-K textual changes and the determinants of option volatility smirks. In this section, we present additional evidence to support the proposed mechanism.

The key assumption is that 10-K textual changes typically signal negative information about a firm’s profitability, leading sophisticated investors to increase their demand for OTM put options.Footnote ¹⁰ Cohen et al. (Reference Cohen, Malloy and Nguyen2020) provide evidence that greater textual changes predict negative news, including lower future earnings and a higher likelihood of bankruptcy. It is plausible, however, that not all textual changes convey negative news. If options traders truly possess superior information processing skills, then volatility smirk increases should primarily stem from textual changes containing new negative information.

Our first test examines whether it is the negative information that drives the relation between cosine measures and volatility smirk changes in Table 3. Following Cohen et al. (Reference Cohen, Malloy and Nguyen2020), we measure the negative sentiment of changes by calculating the ratio of negative words to total words in the changed text.Footnote ¹¹ This measure serves as our proxy for negative news content in textual changes. After dividing our sample at the median value of negative sentiment measure, we run the regression in Table 3 separately for each subsample. As reported in Table 5, the coefficient of cosine measure is much more negative and statistically significant only for the subsample with above-median changes in negative sentiment. In other words, greater textual changes predict increased volatility smirks only when 10-K filing reveals substantially negative information.

TABLE 5Analyses of Option Volatility Smirk Changes Based on Sentiment Subsamples

We next test whether more textual changes lead to higher demand for OTM put options. Following Lakonishok, Lee, Pearson, and Poteshman (Reference Lakonishok, Lee, Pearson and Poteshman2007), we compute delta-adjusted open interest by multiplying an option’s open interest by its delta and dividing the result by the number of outstanding shares of the underlying stock. This adjustment converts open interest into equivalent underlying shares and normalizes it by the total shares outstanding. We then average the delta-adjusted open interest across all OTM put options on the first day after the 10-K release. Table 6 reports the Fama–MacBeth regression results on the relation between the delta-adjusted open interest and the cosine measure. As shown in column 1 for the full sample, the coefficient of cosine is negative and statistically significant, confirming that greater textual changes are associated with higher open interest for OTM put options. Columns 2 and 3 report regression results for above- and below-median negative sentiment subsamples, respectively. The results indicate that the positive relation is primarily driven by textual changes containing substantially negative information. These findings align with our volatility smirk change results reported in Table 5.

TABLE 6Analyses of Option Open Interest Based on Sentiment Subsamples

The results in Tables 5 and 6 suggest that negative information disclosed through 10-K textual changes increases the demand for OTM puts, subsequently resulting in higher option prices (as reflected by implied volatility). However, if competitive market makers can perfectly hedge inventory risk, the option supply curves are essentially flat and option prices are determined by no-arbitrage conditions, irrespective of demand pressure. Upward-sloping supply curves will arise if market makers are risk averse and unable to hedge perfectly. Bollen and Whaley (Reference Bollen and Whaley2004) is the first paper to empirically link changes in demand to changes in option prices via limits to arbitrage. Increased demand for certain options results in higher prices to compensate market makers for the elevated hedging risk they bear. As market makers gradually rebalance their portfolios and/or demand pressure subsides, option prices revert at least partially to their previous levels.

Motivated by the findings in Bollen and Whaley (Reference Bollen and Whaley2004), Gârleanu, Pedersen, and Poteshman (Reference Gârleanu, Pedersen and Poteshman2009) formally develop a demand-based option pricing model in which competitive, risk-averse dealers cannot hedge perfectly. The market incompleteness could be driven by various factors, including discrete trading, stochastic volatility, jump risk, and transaction costs. The model predicts that equilibrium option prices increase with demand pressure from end-users. Essentially, higher prices reflect the premium required by market makers to supply the volume of options contracts demanded by end-users. The model remains agnostic regarding the sources of end-users’ demand pressure—whether motivated by hedging or driven by information. The magnitude and duration of these demand-based price effects depend on the nature of market incompleteness, which in turn determines the speed at which market makers rebalance their inventory positions.Footnote ¹²

The evidence of our article is broadly consistent with Bollen and Whaley (Reference Bollen and Whaley2004) and Gârleanu et al. (Reference Gârleanu, Pedersen and Poteshman2009). In Table 4, we examine changes in option volatility smirks during the 8-week period following the 10-K release. To examine the longer term behavior of volatility smirks, we extend the window to 13 weeks in Table A7 in the Supplementary Material, covering approximately one quarter following the 10-K release. We limit the postrelease horizon to one quarter to avoid any confounding information events, such as new earnings announcement, that may affect option prices. As shown in Table 4 and Table A7 in the Supplementary Material, the coefficients for cosine measure are negative and statistically significant for the first 7 weeks and become statistically insignificant after week 8.Footnote ¹³ This suggests that the price impact on OTM puts, driven by informed options traders’ demand, persists for several weeks before reverting to pre-10-K release levels. The evidence thus supports the limits-to-arbitrage interpretation, as it takes time for options market makers to rebalance their inventory and/or to attract more market makers to provide liquidity in response to increased demand pressure.

Our final test examines how options trading responds to textual similarities in 10-K filings. If higher textual similarity (fewer textual changes) signals nonnegative or positive news, we expect stronger demand for bullish call options relative to protective puts. This, in turn, should lead to higher implied volatility for calls compared to puts. Similar to the baseline model, we measure the change in reverse volatility smirk from the first day after the 10-K release and the baseline day, which is set at least 8 weeks before the release. Table A6 in the Supplementary Material reports the Fama–MacBeth regression results on the relation between reverse volatility smirks and cosine measures. The coefficients of the cosine measure are positive and statistically significant at the 10% level.Footnote ¹⁴ This suggests that higher textual similarities in 10-K filings predict larger increases in reverse volatility smirks. The result contrasts with the changes in volatility smirks reported in Table 3. The evidence is thus consistent with informed investors buying OTM call options to trade on positive signals indicated by minimal or no textual changes in 10-K filings.

In summary, we find that negative information in 10-K textual changes leads to an increase in option volatility smirks. This reflects options traders buying more OTM put options in response to negative news extracted from 10-K textual changes. These results support our conjecture that sophisticated options traders have superior skills in processing public yet complex information disclosed in EDGAR filings.

A. Textual Changes by Reporting Items

We first explore whether textual changes in the product market description segment, usually displayed in Item 1 (Business) of 10-Ks, are related to options trading. Hoberg et al. (Reference Hoberg, Phillips and Prabhala2014) show that product market descriptions contain useful information on the firm’s competitive position in the industry, in addition to the standard industry classification codes. We utilize two measures of product market fluidity developed by Hoberg et al. (Reference Hoberg, Phillips and Prabhala2014) to capture changes in own and competitors’ product descriptions, respectively. To ensure consistent interpretation of textual change coefficient, we use 1 minus the product market fluidity measure in regressions so that lower values reflect greater textual changes in product descriptions. Accordingly, a negative coefficient suggests that more such changes are associated with steeper volatility smirks. Table 7 reports the Fama–MacBeth regression results with the product description similarity measures. We find that greater textual changes to a firm’s own product descriptions are associated with larger increases in option volatility smirks. In contrast, changes in competitors’ product descriptions have a limited impact on volatility smirk movements.

TABLE 7The Effects of Product Market Description Changes on Option Volatility Smirk Changes

Next, we examine textual changes across different items in 10-K filings, acknowledging that managers have varying levels of discretion over specific sections. For example, Cohen et al. (Reference Cohen, Malloy and Nguyen2020) document that firms’ reporting changes are mainly concentrated in Item 7 (Management’s Discussion and Analysis), Item 3 (Legal Proceedings), Item 7a (Quantitative and Qualitative Disclosures about Market Risk), and Item 1a (Risk factors). We expect changes in these discretionary items to be more informative and, therefore, to have a greater impact on options trading. To test this, we construct cosine similarity measures for each item based solely on the textual content within the respective section.

Table 8 shows the coefficients of cosine measures estimated from equation (1) for four 10-K items: Item 7, Item 3, Item 7a, and Item 1a.Footnote ¹⁵ Consistent with the strong return predictability documented by Cohen et al. (Reference Cohen, Malloy and Nguyen2020), we find that option volatility smirks respond more strongly to textual changes in these discretionary 10-K items. Moreover, the average cosine coefficient for these four items is significantly larger in magnitude compared to other sections of the 10-K. This difference is statistically significant at the 1% level.Footnote ¹⁶ These findings support our Hypothesis 2a, suggesting that options traders actively respond to textual changes in parts of the 10-K that are both more informative about firms’ fundamentals and more subject to managerial discretion.

TABLE 8The Effects of Textual Changes to Information Intensive and Discretionary Items

B. Information Environments

Previous studies find that the predictive power of option expensiveness for future stock returns is particularly pronounced among firms with higher levels of information asymmetry (Hu (Reference Hu2014)). This is because option traders are more likely to profit from information advantages in firms with more diverse investor opinions. Therefore, we expect that option volatility smirks will react more strongly to 10-K textual changes for these firms.

Drawing from the existing literature, we use the following two measures to proxy for information asymmetry: abnormal idiosyncratic volatility (AIV) and analyst coverage. Morck, Yeung, and Yu (Reference Morck, Yeung and Yu2000) and Durnev, Morck, and Yeung (Reference Durnev, Morck and Yeung2004) relate the idiosyncratic volatility (IV) of stock returns to firm-specific information flows. To isolate abnormal price variation stemming from informed trading, Yang et al. (Reference Yang, Zhang and Zhang2020) propose the AIV measure, defined as the IV prior to information events in excess of its normal levels. Specifically, we compute AIV as the difference between IV measured during preearnings-announcement windows and IV measured during nonearnings-announcement periods using daily returns for each fiscal year. Following Yang et al. (Reference Yang, Zhang and Zhang2020), we define the preearnings-announcement periods as the 5 business days preceding each of the four earnings announcement dates. The nonearnings-announcement periods comprise all days excluding the 11-day window surrounding each earnings announcement. Our next measure is analyst coverage (Hu (Reference Hu2014)), defined as the number of analysts following a firm, with data sourced from IBES. We compute these variables at the end of the previous fiscal year and partition the sample into high and low information asymmetry groups based on their respective median values.

Table 9 reports the regression results based on equation (1), estimated separately for the high versus low information asymmetry subsamples. The coefficients for the cosine measure are consistently negative but statistically significant only for firms characterized by high AIV and low analyst coverage. Hence, an increase in textual changes leads to significantly higher volatility smirk only among firms with greater information asymmetry. This finding is consistent with the conjecture in Hypothesis 2b.

TABLE 9The Effect of Textual Changes on Option Volatility Smirk Changes by Information Environments

C. Short-Sale Constraints

Options trading is not the only way sophisticated investors can exploit negative information in 10-K textual changes. Short selling offers informed traders another method to express bearish views. However, unlike options trading on centralized exchanges, the securities lending market is fragmented, and even large institutions often face high borrowing costs or may be unable to secure shares to borrow. Hence, informed investors are more likely to rely on options when short-sale constraints are binding. Consistent with this, Hu (Reference Hu2014) finds that options trading conveys more information about future stock prices for firms facing greater short-sale constraints. We conjecture in Hypothesis 2b that the effect of 10-K textual changes on option volatility smirks is stronger when short-sale constraints are more severe.

Following D’Avolio (Reference D’Avolio2002), Asquith, Pathak, and Ritter (Reference Asquith, Pathak and Ritter2005), and Hu (Reference Hu2014), we use institutional ownership as a proxy for short-selling costs. This is because lower institutional ownership is often associated with substantially higher costs for short sellers. We divide the sample into high-cost and low-cost groups based on the median of fiscal-year-end institutional ownership and estimate equation (1) separately for each group. Table 10 presents the regression results. The coefficient estimates of the cosine similarity measure are negative and statistically significant for both groups. However, the effect of 10-K textual changes is approximately four times larger for the high-cost group compared to the low-cost group. This finding supports that informed investors rely more on OTM puts to trade on negative information for firms facing greater short-sale constraints.

TABLE 10The Effect of Textual Changes on Option Volatility Smirk Changes by Short-Sale Constraints