# Tone of language, financial disclosure, and earnings management: a textual analysis of form 20-F – Financial Innovation

May 10, 2022

### Summary statistics

The industry subsamples are based on the Global Industry Classification Standard (GICS) developed by the S&P Dow Jones Indices, a leading provider of global equity indices. After excluding the financial industry (see Appendix 2), there were nine industries in the sample. Over the entire sample period, the information technology industry accounts for 42.98% of the total observations, while the GICS utility industry accounts for only 1.78%. The weights of the other industries vary from 2.67% (communication services) to 16.7% (consumer discretionary). In general, sample firms are widely dispersed across various industries.

Table 1 reports the summary statistics for the selected variables. For example, for the dependent variable, EM, the minimum is 0, and the maximum is 0.341, with a mean of 0.031 and a median of 0.017. These results suggest variations in the amount of earnings management in the sample firms. Table 1 also reports the descriptive statistics of the eight sentiment measures, the percentages of words in Form 20-F selected from each word list. The mean values are greater than the median values for most sentiment variables, implying that the data distribution is skewed and outliers. Specifically, negative, litigious, and uncertain have the highest mean percentages (0.007, 0.007, and 0.006, respectively). In addition, it reveals that firm size (LNSIZE) and change in sales (REV) have the highest standard deviations, 0.927 and 0.480, respectively, suggesting that the sample firms’ revenues vary remarkably by firm size.

This study also examines the effect of the 2008 financial crisis on the selected variables. Specifically, it divides the sample into two sub-samples: a sub-sample of 86 observations during 2002–2007 and a sub-sample of 363 observations during 2008–2014. This study uses the beginning of 2008 as the cut-off point for the financial crisis. It compares the means of selected variables before and after 2008. The mean EM is slightly higher in the post-crisis sub-sample than in the pre-crisis period (0.031 vs. 0.025). We also find that the values for all sentiment variables are lower in the post-crisis period. For example, the mean of litigious drops from 0.009 to 0.006. Finally, the results show that the debt ratio (0.403 vs. 0.322) increases, while the cash ratio (0.234 vs. 0.300), cash flow ratio (0.074 vs. 0.137), and change in sales (0.251 vs. 0.431) decrease after the financial crisis. The results suggest that most firms have experienced financial constraints during the economic problems.

Table 2 reports the correlations between key variables. Most sentiment variables are positively associated with EM, except for litigious. For example, Positive, Uncertainty, Strongml, Moderml, and Weakml are positively correlated with EM (the coefficients are 0.196, 0.169, 0.092, 0.132, and 0.170, respectively), although the correlations between EM and Negative (0.029) and Irrverb (0.028) are weak. These results are consistent with those of Loughran and McDonald (2013), although there is an overlap between some categories in the word list. For example, there are strong correlations between Weakml and Uncertainty (0.954), Uncertainty and Positive (0.895), and Moderml and Positive (0.887), suggesting that multicollinearity will be a problem if it includes all the sentiment variables in one regression. Therefore, this study follows Loughran and McDonald (2013) and runs the regressions individually for each category of sentiment variables. In addition, there is a high correlation coefficient between CF and ROA (0.755), suggesting that firms with higher cash flows are more likely to have better performance. A possible explanation is that these firms have lower fixed asset costs and more capital turnover, resulting in higher profits.

### Baseline regression results

To further examine the link between earnings management and sentiment, we first report the baseline regression results, including the firm, year, and industry fixed effects in Table 3. Column (1) shows that the correlation between firm nature and earnings management is positive (0.013) and significant at the 5% level. This demonstrates that the managers of state-owned enterprises engage in more earnings management, consistent with studies of the Chinese financial markets (Bao and Lewellyn 2017). The firm size and cash flow coefficients are negative and significant, suggesting that larger firms with higher cash flows have fewer earnings management activities. In other words, small firms and firms with inadequate cash flow are more likely to manipulate their earnings. The coefficient of firm leverage is positive and significant, suggesting that firms with high debt ratios engage in earnings management, consistent with the original assumption.

Columns (2)–(9) of Table 3 examine each of the eight sentiment categories as independent variables to explore whether the language used in financial disclosures can explain earnings management. The results show that most sentiment variables (Positive, Uncertainty, Strongml, Moderml, and Weakml) have positive and statistically significant coefficients for E.M. In Column (2), Positive has a positive and significant coefficient (6.482), suggesting that firms using a higher proportion of positive words in their 20-Fs experience more earnings management. This finding implies that companies use positive words more frequently in financial statements to conceal their true earnings status. Similarly, uncertainty has a positive coefficient (3.793), suggesting that firms using higher proportions of uncertain words in their 20-Fs engage in more earnings management. A one standard deviation increase in the percentage of uncertainty leads to a 0.009 increase in EM (3.793 multiplied by the standard deviation of 0.00234). This suggests that companies that frequently use words indicating uncertainty about the future may engage in more earnings management activities than others. Similarly, Columns (6)–(8) show that Strongml, Moderml, and Weakml have positive and significant coefficients (6.372, 8.741, and 5.357, respectively), suggesting that firms that use a high proportion of modal words engage in more earnings management. Specifically, a one standard deviation change in the percentage of Strongml, Moderml, and Weakml is linked to 0.005 (6.372 times 0.0008), 0.005 (8.741 times 0.00058), and 0.008 (5.357 times 0.00153) increase in earnings management, respectively.Footnote 5 We also find that the coefficients of REV and ROA are positive and significant. A possible explanation is that companies with better growth prospects or higher profitability ratios may have more motives for earnings management and manipulating numbers (Espahbodi et al. 2021). Overall, the results demonstrate that the tone of 20-Fs relates to earnings management. Specifically, a positive correlation exists between a positive, uncertain, or modal tone in 20-Fs and earnings management behaviors, suggesting that the sentiment words “good, may, could, depend, and approximately” may reflect managers’ immoral behaviors. In summary, the results support both hypotheses and indicate that qualitative textual tones can provide valuable information to investors.

These results are consistent with those obtained for the U.S. companies in previous studies (Jegadeesh and Wu 2013; Kang et al. 2018; Jaspersen et al. 2021). For example, Kang et al. (2018) claim that it is risky to interpret business managers’ positive expressions because the data can include their subjective opinions and viewpoints from the companies’ perspectives. Furthermore, they employ the text-mining method to identify the tones of the 10-K narratives to determine whether the changes are consistent with the current earning levels. Jaspersen et al. (2021) suggest that qualitative disclosures are additional sources of information about a company’s financial situation. Still, executives likely hide their earnings management activities in these disclosures. Nevertheless, their results demonstrate that qualitative disclosures can predict earnings management and are useful for learning about companies’ accounting choices. In addition, Jegadeesh and Wu (2013) find a significant relationship between document tone and market reaction to 10-K filings for negative and positive words. Furthermore, they indicate a need to partition words into positive and negative word lists subjectively. However, our study focuses on 20-Fs, different from previous studies on 10-Ks.

### Multicollinearity test

A possible problem in the baseline regression models is the strong correlation between the variables and relatively small sample size, leading to multicollinearity. This study conducts an extra multicollinearity test commonly used in multiple regression models and reports the variance inflation factors (VIFs) in Table 4 to address this issue. It shows that the average VIF values are less than 10 for all multiple regressions, suggesting that there are no significant multicollinearity issues between the sentiment variables. These results confirm the robustness of the baseline regression results.

### An alternative earnings management variable

This study further investigates whether the impact of textual analysis on earnings management persists when an alternative measure of earnings management is used to check the robustness of the empirical results. The performance-adjusted discretionary accruals variable developed by Kothari et al. (2005) is the alternative measure. It estimates the following regression in Eq. (3):

$$TAccr_{it} = alpha_{0} + alpha_{1} left( {frac{1}{{Asset_{i,t – 1} }}} right) + alpha_{2} Delta Rev_{it} + alpha_{3} PPE_{it} + alpha_{4} ROA_{it} + varepsilon_{it}$$

(3)

where (TAccr_{it}) is total accruals, measured as the change in non-cash current assets minus the change in current non-interest-bearing liabilities, minus depreciation and amortization expenses for firm i in year t, scaled by lagged total assets ((Asset_{i,t – 1}));(Delta Rev_{it}) is the annual change in revenue scaled by lagged total assets; (PPE_{it}) is property, plant, and equipment costs for firm i in year t, scaled by lagged total assets; and (ROA_{it}) is the return on investments for firm i in year t. The residuals from the regression model are discretionary accruals. This study uses the absolute value of discretionary accruals (AVDA) as an alternative proxy for earnings management. Table 5 shows the regression results. As shown in Columns (2), (4), (7), and (8) of Table 5, the significant relationship between most sentiment words and earnings management continues to hold as an alternative measure of earnings management. This result confirms the main results presented in Table 3. This shows that companies that use more positive or uncertain words in their financial reports are more likely to engage in earnings management.

### Controlling for the length of financial statements

The percentage of sentiment words used in financial statements may be affected by the text length. In this study, we use file size, measured by the natural logarithm of the file size in megabytes (Lnfsize) of the “complete submission text file” for Form 20-F filing, as a proxy for the length of the text. As shown in Table 6, the significance levels and signs of the coefficients of the sentiment word variables do not change significantly, indicating that the baseline regression results are robust after controlling for the length of the financial statements.

### Impact of the 2008 financial crisis

This study also analyzes the impact of the financial crisis on earnings management because firms may face different regulations during crisis periods, influencing how they report earnings. Additionally, firms may change the tone of their financial disclosures during a crisis period because of increased uncertainty. Therefore, it divides the sample period into two subsamples: 2002–2007 (pre-crisis) and 2008–2014 (post-crisis). This analysis identifies whether there is a difference in the tone of financial statements before and after the crisis.

The results for two subsamples are reported in Panels A and B in Table 7. The first column of each panel includes only control variables. Columns (2)–(9) add the variables for each sentiment category as the main independent variables. Positive has positive coefficients on earnings management for both periods (10.10 and 6.991, respectively), suggesting no significant difference in the relationship between positive words and earnings management in these periods. Column (9) of Panel A shows that Irrverb has a marginally significant negative coefficient of − 5.993. However, the results may be biased due to the small sample size for this category (only 86 observations). Therefore, Irrverb has less power to explain earnings management in the pre-crisis subsample. Columns (3)–(9) of Panel B in Table 7 show that the sentiment variables Uncertainty, Strongml, Moderml, and Weakml have significant positive coefficients on EM after the crisis, consistent with the results for the full sample in Table 3. The results confirm that these modal variables have similar and significant relationships with EM in the pre -and post-crisis periods.

Overall, there is no significant change in the relationship between Positive, Uncertainty, Strongml, Moderml, and Weakml and EM in the pre -and post-crisis periods. Therefore, the main results are robust after controlling for the effects of the 2008 financial crisis.

### Information technology versus non-information technology industries

The sample reveals that 42.98% of observations are from the information technology industry. Thus, we divide the sample into two subsamples: information technology and non-information technology industries. Panels A and B of Table 8 present the regression results for the information technology industry and non-information technology industry subsamples, respectively. Columns (4)–(6) of Panel A show that the coefficients of Litigious, Strongml, and Moderml are significant, indicating that firms in information technology are more likely to use a positive tone in their financial disclosures to conceal earnings management. However, there are no significant changes in the signs of the estimated coefficients, indicating that the robustness of the empirical results holds for all industries.

### Discussion of the empirical results

The main results are summarized as follows. First, positive, uncertain, and modal words are positively and significantly correlated with earnings management. Table 3 shows that firms with a higher proportion of positive, uncertain, and modal words in financial reports are more likely to engage in earnings management. The results imply that companies use more positive words when concealing earnings management, probably to attract potential investors or other businesses. Second, although the VIF values indicated possible multicollinearity among the selected variables, the mean values of the VIFs are less than 10 in each regression. This suggested no significant multicollinearity issues in the main regression models. Third, it uses the absolute values of discretionary accruals as an alternative proxy for earnings management and re-estimates the main regression. These results provide further support for the main results in Table 3, even after controlling for the effect of the length of financial statements. In addition, this study considers the impact of the financial crisis and finds that the main results hold in the pre- and post-crisis periods. Finally, it divides the sample into information technology and non-information technology industries and finds that the results do not exhibit significant differences between these two sub-samples.