Bank Connectedness: Qualitative and Quantitative Disclosure Similarity and Future Tail Comovement

Transcription

1 Bank Connectedness: Qualitative and Quantitative Disclosure Similarity and Future Tail Comovement Robert M. Bushman Kenan-Flagler Business School University of North Carolina-Chapel Hill Jason V. Chen University of Illinois at Chicago Christopher D. Williams Ross School of Business University of Michigan September 12, 2016 Preliminary and incomplete: comments welcome We appreciate the helpful comments of Donny Zhang and seminar participants at Chicago Booth. Bushman thanks Kenan-Flagler Business School, University of North Carolina at Chapel Hill, Chen thanks the University of Illinois for their support, and Williams thanks the Stephen M. Ross School of Business for financial support.

2 Bank Connectedness: Qualitative and Quantitative Disclosure Similarity and Future Tail Comovement Abstract Pillar 3 of the Basel framework stresses the importance of supplementing quantitative information with qualitative disclosures for understanding risk characteristics of banks. One important aspect of risk is a bank s contribution to systemic risk driven by its degree of connectivity with other banks. In this paper, we measure bank connectedness by constructing text-based measures of similarity between banks based on 10-K Business Description section and MD&A disclosures. We examine whether this qualitative connectedness measure contains information about future tail comovement incremental to information in accounting-based quantitative similarity measures, and in past equity returns. Focusing first on average similarity of a bank s qualitative disclosures with those of all other banks, we find that comovement between the lower tail of a given bank s future equity return distribution and the lower tail of the banking system s returns is increasing in the bank s average similarity, and that the relation between future banking system returns and a bank s current period returns is increasing in average qualitative similarity. We also construct groups of peer banks with the most (least) similarity in qualitative discussions, finding that banks co-move more (less) in the tails with peers that share a higher (lower) cosine similarity, and that tail returns of the portfolio comprised of high (low) similarity peer banks are more (less) affected by tail outcomes of an individual bank. Finally, we construct a measure of the banking system s vulnerability to systemic risk by using LDA analysis to construct clusters of banks whose 10-K discussions focus on similar topics in a given year. Measuring systemic risk vulnerability using the size of the largest cluster, we find that the number of future bank failures is increasing in this measure.

3 Introduction The Financial Crisis of has focused significant attention on methods for assessing and managing downside tail risk of banks. The focus is not just on standalone risks of individual banks, but encompasses the entire banking system and the complex web of interconnections between banks through which illiquidity, insolvency, and losses can spread during periods of financial distress. In this regard, the Pillar 3 disclosure requirements issued by the Basel Committee on Banking Supervision (2015) stress the importance of supplementing quantitative information with qualitative disclosures to facilitate a fuller understanding of a bank s risk profile. One important aspect of a bank s risk profile is the contribution to systemic risk driven by its degree of connectivity with other banks. Recent literature suggests that systemic risk measurement cannot be reduced to a simple formula, where many indicators of systemic risk exposure have been proposed (e.g., Hansen, 2014; Bisias, 2012). A common approach to measuring bank connectedness and systemic risk exposure relies exclusively on quantitative information using return series of publicly traded securities or some combination of returns and balance sheet data (e.g., Billio et al., 2012; Huang et al., 2011; Adrian and Brunnermeier, 2016; Acharya et al., 2016). However, these measures ignore potentially valuable information embedded in the rich qualitative disclosures contained in mandatory financial reports. In this paper, we complement and extend the banking and text analysis literatures in several important ways. First, we develop a new time-varying measure of bank interconnectedness that spans qualitative 10-K disclosures of publicly traded banks, and examine its incremental information content for assessing systemic risk. To measure qualitative connectedness across banks we analyze the text of discussions contained in both the business description and MD&A 1

4 sections of 10-K reports and construct pair-wise measures of similarity between all banks. 1 We then use the matrix of pair-wise similarities to design measures of connectedness and to isolate specific subsets of banks that are highly interconnected. Second, we introduce a new accountingbased, quantitative measure of connectedness by estimating pair-wise cosine similarity between all banks based on the entire, standardized vector of quantitative accounting data required to be reported in regulatory financial report filings. Our qualitative and quantitative connectedness measures together allow us to examine the relative information value of distinct dimensions of bank financial reporting with respect to assessing systemic risk. Specifically, we investigate whether qualitative connectedness contains information about future comovement in the tails of banks equity return distributions incremental to quantitative connectedness and past equity return tail co-movement. 2 Finally, we exploit the rich texture of banks qualitative disclosures to develop a novel measure of the banking system s vulnerability to systemic risk at a point in time. Here, we use LDA analysis to organize all banks in a given year into clusters based on the similarity of the topics discussed in their 10-K reports. We conjecture that, regardless of the nature of the specific topics discussed, as the number of banks that concentrate their discussions on the same set of topics increases, the more vulnerable the banking sector becomes to systemic risk. The idea that our qualitative connectedness measure could have incremental information about future tail comovement relative to quantitative accounting information builds on a growing literature using textual analysis to extract information from 10-K reports in a variety of contexts. 1 Throughout the paper, we will use the terms bank similarity and qualitative connectedness to refer to this textbased measure. 2 Tail co-movement between a pair of random variables describes how outcomes in the tails of their distributions comove. In our paper, the distributions under consideration are the equity return distributions of publicly traded banks, where we consider how outcomes in the lower tail of a bank s equity distributions comove with tail outcomes of portfolios comprised of selected groups of other banks. 2

5 This research provides evidence consistent with the narrative text of 10-K reports providing incremental information about a firm s prospects beyond the quantitative information in the financial statements. This includes findings that there is incremental information content in the tone of the text in the 10-K (Feldman, Givindaraj, Livnat, and Segal, 2010; Loughran and McDonald, 2011), in its readability (Li, 2008; Loughran and McDonald 2014), and in year-onyear changes to the MD&A section (Brown and Tucker, 2011; Cohen, Malloy, and Nguyen, 2015). We extend this literature by examining qualitative similarities in10-k disclosures across banks in the context of bank connectedness and tail comovement across banks. There are also plausible reasons supporting the possibility that qualitative connectedness has incremental information relative to a contemporaneous measure of tail co-movement based on banks realized equity returns. First, taking the efficient markets perspective that stock returns impound all available information, there is still potentially scope for qualitative disclosures to play a role in predicting future tail comovement. While stock prices optimally aggregate all available information for valuation purposes, the weights on information signals that are optimal for valuing equity may not be optimal for assessing tail comovement. 3 Also, stock price impounds information arrival over time, aggregating older news with more recent news that may be more timely and pertinent. To the extent that 10-K narrative disclosures reflect more recent information, including forward looking assessments, our qualitative connectedness measure allows this information to be disentangled from aggregated information in prices and weighted separately in predicting future tail comovement. 4 Second, Billio et al. (2012) provide evidence that bank connections are not priced contemporaneously due to limits to arbitrage. It is 3 This argument is analogous to the arguments developed in Paul (1992) and Bushman and Indjejikian (1993) to explain why accounting earnings are used in executive compensation contracts in addition to stock returns. 4 In this spirit, Bushman, Hendricks and Williams (2016) provide evidence that a text-based measure of bank competition captures changes in the competitive environment in more timely fashion that standard measures of competition based on historical information. 3

6 also possible that investors do not fully extract all the information about connectedness from qualitative disclosures. Whether our measure of qualitative connectedness has incremental information is ultimately an empirical question. Our approach to measuring bank connectedness from qualitative 10-K disclosures is related to the approach used by Hoberg and Philips (2016) to organize all publicly traded firms into textbased network industries. Applying textual analysis to the business description section of the 10- K, Hoberg and Philips (2016) build measures of pairwise product similarity capturing how firms are related to each other in terms of product offerings. They then use product similarity to create new industry classifications that reflect assessed competition in the product market. Hoberg and Philips (2016) show that their product similarity-based industry groupings offer some significant improvements relative to existing industry classifications and also allows them to examine dynamic changes in competition over time. Our paper differs from Hoberg and Philips (2016) in several important respects. First, we focus on a single industry, banking, where Hoberg and Philips consider all industries, including banking. More importantly, our objective is distinct from theirs. Hoberg and Philips focus on building connections between firms based on product market competition, where we seek to estimate measures of bank connectedness that capture the extent to which a group of banks share similar vulnerabilities to downside tail risk. Such risk vulnerabilities are likely driven by sources of connectedness beyond product market competition, including key aspects of the bank s current and forecasted situation that spans performance, business models, credit risk, investment concentrations, funding sources, and liquidity exposures, among other issues. To capture similarity across a range of important dimensions, our text analysis combines the business 4

7 description section (similar to Hoberg and Philips) with the Management Discussion and Analysis disclosure (MD&A) incorporated in banks 10-K reports. We believe including the MD&A section in our measure is important because the SEC intends that MD&A provide investors with an opportunity to see the company through the eyes of the management (SEC 1987). The SEC has guided companies to present known trends, events, commitments, plans, and uncertainties that are likely to materially affect company liquidity, capital resources, or future operations (SEC 1989, 2003). Voluntary projections of anticipated trends are also encouraged. Recent research finds evidence that MD&A discussions contain useful information that is incremental to other available sources of information. 5 We construct time-varying measures of cosine similarity between a bank s business description and MD&A discussions and those of all other publicly traded banks. We posit that by calculating how similar a bank is to all other banks, we can construct a network of banks that share similar vulnerabilities to downside tail risk. Using pairwise similarity scores, we can group banks together with other banks with which they are most similar (peers) and those with which they are least similar (non-peers). A bank is similar (not similar) to another bank if their cosine similarity is above (below) a certain percentile of its cosine similarity distribution. 6 Using these groupings, we then investigate whether the lower tail of a given bank s future equity returns distribution commoves more strongly with the tail of its peer group return distribution than with the tails of non-peers. Our first set of empirical analyses investigates whether our qualitative connectedness measure contains incremental information about future co-movement in the tails of banks equity 5 See for example, Ball, Hoberg and Maksimovic (2016), Mayew, Sethuraman, and Venkatachalam (2015), Cohen, Malloy, and Nguyen (2015), Volkan Muslu, Radhakrishnan, Subramanyam, and Lim (2014), Brown and Tucker (2011), among others 6 The percentiles that we use are 5%, 10%, 25% and 50% of the distribution. We repeat this process for each calendar year to create timing varying groupings of similar and non-similar banks. 5

8 returns. We measure tail risk comovement by constructing two measures inspired by the Marginal Expected Shortfall measure developed in Acharya et al. (2016). Our first measure is constructed as the average of the abnormal returns of a given bank over the 20 days where the portfolio of banks in a chosen index group has it lowest abnormal performance for the year. Our second measure is the number of days in which a bank experiences one its 20 lowest abnormal return days at the same time that the portfolio of banks in a chosen index group is also experiencing one its 20 lowest return days. The portfolio of banks used to construct our index groups is either the entire banking sector, or portfolios consisting of banks designated as a bank s peer or non-peer banks based on our qualitative measure of connectedness. We first examine the relation between a bank s average cosine similarity with all other banks and its tail co-movement with a portfolio consisting of all other banks. We find that comovement between the lower tail of a given bank s future equity return distribution and the lower tail of the banking system s return distribution is increasing in the average cosine similarity of the bank. Our results hold after controlling for the current level of tail comovement and quantitative bank similarity. We next implement a Granger causality test of the relation between a bank s buy and hold returns and future banking sector returns (e.g., Billio et al 2012). We find that the relation between future banking system returns and a bank s current period returns is increasing in the bank s average qualitative similarity. Next, rather than considering all banks together as a single group, we instead focus on groups of banks designated as a bank s peer or non-peer banks based on our qualitative measure of connectedness. Here we investigate whether co-dependence between the lower tail of a given bank s future equity returns distribution and the lower tail of its peer group return distribution is significantly higher than is its co-dependence with the tail returns of non-peers. In essence, this 6

9 analysis examines the extent to which bank similarity allows us to effectively cluster banks into subsets of the entire banking sector in which future tail movement is expected to be the highest or lowest. We find that in the future, banks co-move significantly more (less) in the tails with peers that share a higher (lower) cosine similarity. Our analyses of qualitative connectedness so far examine the extent to which the returns of a bank are low when conditioning on whether similar banks are experiencing low returns. We now turn this around and examine whether the returns of the group of banks with which an individual bank is similar are low when conditioning on the individual bank s returns being low. As noted in Adrian and Brunnermeier (2016), the direction of conditioning matters and the incidence of a bank s tail outcomes conditioned on peer banks having tail outcomes need not be the same as the incidence of peer banks having tail outcomes conditioned on an individual bank having a tail outcome. Building on the approach used in Boyson, Stahel and Stulz (2010) to examine hedge fund contagion using daily stock returns, we regress the proportion of banks in an individual bank s peer (non-peer) group experiencing low performance on a given day on whether or not the individual bank is also having a low performance day. We find that proportion of banks in an individual bank s peer (non-peer) group experiencing low performance is higher (lower) when the individual bank is having a low performance day. Finally, we explore the topics discussed in the business description and MD&A sections of banks 10-Ks using Latent Dirichlet Allocation (hereafter LDA). LDA is a textual analysis technique which finds topics in unstructured text. Building on the LDA analysis, we apply cluster analysis techniques to form groups of banks discussing similar topics in a given year. We construct a novel measure of the banking system s vulnerability to systemic risk based on the size distribution of these clusters across the entire banking sector. Measuring systemic risk 7

10 vulnerability using either the size of the largest cluster or the skewness of the size distribution, we find that the number of future bank failures is increasing in this measure. This paper makes several contributions to prior literature. First, this is the only paper to our knowledge that uses qualitative disclosures from banks 10-Ks to aid in understanding systemic risk. This complements Rönnqvist and Sarlin (2016), who estimate interconnections between large European banks based on co-occurrences of bank names in news articles, and papers measuring bank connections using quantitative data ((e.g., Billio et al., 2012; Huang et al., 2011; Adrian and Brunnermeier, 2016; Acharya et al., 2016; Cai et al., 2016). Given the regulatory attention to identifying, understanding and measuring systemic risk, our measure provides a unique way of measuring connectedness within the banking system. Second, we provide a unique measure of quantitative accounting similarity. Given the regulatory nature of the banking industry, banks are required to file uniform reports that insure consistency in the reporting of accounting information. This uniformity in reporting allows us to quantify accounting similarity and then test the information content of qualitative vs. quantitative disclosures. Finally our paper provides interesting insights into the role of clustering. Our final analyses suggest that it is not similarity per se that is a driver of system-wide risk, but a lack of diversity in similarities. As the concentration of banks increase around a particular set of topics, this reduces the diversity in the system and magnifies the impact of any shocks. Therefore, it is not a problem if there is clustering as long as there is diversity in concentrations across clusters. The papers proceeds in as follows, section 2 discusses the sample selection and measurement of connectedness. Section 3 discusses the empirical approaches we take and the results. Section 4 concludes. 8

11 2. Sample Selection, Measures of Connectedness and Descriptive Statistics Our sample consists of all banks with two digit historic SIC codes between 60 and 62 which are available in Compustat Annual or Compustat Annual Bank. We download each financial institutions 10-K and 10-K405 filings from the SEC EDGAR online filling system. Our sample of 10-K filings begins in 1995 and ends in The sample begins in 1995 because this is the first year in which the SEC required all publically traded companies to make their filings publically available electronically through the EDGAR filing system. The Management Discussions and Analysis (hereafter MD&A) and Business (here after BUS) sections of each 10- K are extracted using PERL. For our accounting-based quantitative similarity measures we access Call Reports and FR Y-9C filings from various regulatory databases. Financial information was obtained from Compustat Annual and Compustat Annual Bank. Market returns and pricing information is obtained from Eventus and CRSP where needed. In the remainder of this section, we discuss in detail how we measure qualitative similarity (section 2.1), accounting-based quantitative similarity (section 2.2), and tail risk comovement (section 2.3). We also provide descriptive statistics for these measures (section 2.4). 2.1 Qualitative Measure of Connectivity To construct our qualitative measure of connectivity between banks we use the 10-K business description section (BUS) and the MD&A disclosures to compute pairwise word similarity scores for each pair of banks in a given year (we also compute measures using BUS and MD&A separately). The business description section typically appears as Item 1 or Item 1A in 10-K reports. To estimate cosine similarities, we first extract BUS and MD&A from each financial institution s 10-K filing for each year and combine the two disclosures. As is common in the literature, stop words are eliminated from the text. Using the text from the combined BUS 9

12 and MD&A disclosures, we then construct a vector summarizing each bank s usage of words. The number of elements in these vectors is equal to the number of unique words used in the set of all combined BUS and MD&A. Each element of a vector represents the number of times that a unique word is mentioned in a bank s BUS and MD&A text in a given year. For each year, we then estimate pairwise cosine similarity between a given institution s word vector and the vectors of all other banks in the sample. Cosine similarity is a technique from the field of textual analysis which calculates the similarity between two sets of texts (Kogan et al 1998). The technique has had wide spread use in the areas of computer science and web development (Joydeep et al 2000). Recently studies in accounting have begun using this technique to examine changes in firm s fundamentals and similarity in product market offerings (Brown and Tucker 2011, Hoberg and Phillips 2015). The cosine similarity between two banks is the cosine of the angle between the vectors of words that comprise the combined BUS and MD&A. Specifically, the cosine similarity between two vectors of words B 1 and B 2 is calculated as follows: Cosine Similarity B B 1 2 =. B B 1 2 where indicates vector dot product, and B is the length of vector B. B 1 and B 2 are the vectors of words for two distinct bank s being compared. The axes of each vector are the unique words in the text and the magnitude of the axis is the number of times that the given word is mentioned in the given text. The cosine similarity is higher when banks i and j use more of the same words with similar intensity, where a cosine similarity of 1 means that the two word vectors are identical. As discussed in more detail below, we use the matrix of pairwise cosine similarities in several different ways. For some analyses, we compute the average cosine similarity between a 10

13 given bank and all other banks in the sample in a given year (AvgCos_MDABUS). In Table 1 we report that AvgCos_MDABUS has a mean value of 0.70 with a standard deviation of We also compute cosine similarity separately using either MD&A alone or BUS alone. Table 1 reports that the mean value of AvgCos_MDA (AvgCos_BUS) is 0.68 (0.57) with a standard deviation of 0.09 (0.11). Note that AvgCos_BUS is similar in spirit to the product similarity measures constructed in Hoberg and Philips (2015), as both our measure and theirs is based on text analyzing the business description section of 10K alone. We also use pairwise cosine similarities to form the group of banks with which a given bank has high (low) cosine similarity in a given year. Our procedure for forming groups is similar in many respects to that used in Hoberg and Philips (2015) to place firms into product similarity industries. Instead of text analyzing the business description section to create product similarity industries, we create bank connectedness groups based on all dimensions of activities described in banks business description section and MD&A disclosures. 2.2 Accounting-Based Quantitative Measure of Connectivity We construct a novel measure of banks quantitative similarity by estimating pairwise cosine similarity between banks based on the entire, standardized vector of quantitative accounting data required to be reported in banks mandated regulatory filings. All banks prepare their regulatory filings using the reporting template required by bank regulators. This template structure implies that we can construct vectors of accounting data for each bank based on the same standardized set of required reporting fields. Specifically, we calculate cosine similarities using financial institutions mandatory call report filings or FR Y-9C filings as appropriate, which have identical reporting fields (but different call letters). A complete set of call reports is 11

14 obtained from the Federal Financial Institutions Examination Council for the year 2001 to Call reports for the years 1994 to 2000 are obtained from the Federal Reserve Bank of Chicago. 8 FR Y-9C reports were obtained from the datasets provided by the Chicago Federal Reserve Bank. 9 We prepare the data in the call reports and FR Y-9C filings by first aggregating quantitative information based upon their main series and sub series classifications. 10 This allows us to create variables which are comparable across reports. A firm s quantitative information is then represented as a vector where the axis is the specific quantitative variable and the magnitude of the axis is its reported value. We then calculate the cosine similarity between each banks quantitative information vector and those of all others banks in the same calendar year. This measure allows us to assess the similarity across the entire set of quantitative measures reported by banks in their regulatory reports (AvgCos_Report). In Table 1 we report that across the sample, the average quantitative cosine similarity between a given bank and all other banks in the sample in a given year, AvgCos_Report, has a mean value of.74 with a standard deviation of.18. The sample size for AvgCos_Report, 5,499, is smaller than those used for our qualitative connectedness measures because we require financial institutions to have a regulatory report. 2.3 Tail Risk Comovement Following the recent financial crisis there has been considerable interest in modeling and measuring systemic risk, the risk that many banks will simultaneously experience financial distress and impose externalities on the overall economy. There is no agreed upon approach to Main series variables are the sum of certain sub series variables. Some banks gave main series variables while others provide the sub series variables underlying the main series variables. To ensure consistency across all banks regulatory reports, we create main series variables by summing the respective sub series variables when necessary. For a description of all the main and sub series variables see 12

15 this measurement (e.g., Bisias et al., 2012, Hansen, 2014). One important stream of literature exploits the high frequency observability of bank s equity prices to extract measures of systemic risk, focusing on comovement in the tails of equity returns across banks (Acharya et al., 2016, Adrian and Brunnermeier, 2016). We measure tail risk comovement by constructing measures inspired by the marginal expected shortfall measure (MES) developed in Acharya et al. (2016). The MES measure of Acharya et al. (2016) captures the connection between a bank s equity returns and market equity returns on days where the market return is in the bottom 5% for the year. That is, it measures the extent to which an individual bank s returns are low when the overall (banking) market returns are low. Building on this idea, we create two measures of tail comovement. Our first measure, LFM Days, reflects the number of days in year t where bank i and a portfolio of banks in an index group simultaneously experience low returns performance. An extreme low performance day occurs if it is in the set of the lowest 20 return days for year t based upon daily abnormal returns. A bank s daily abnormal return is calculated using Eventus, and is the difference between the bank s return and a value weighted market return. Eventus calculates the value weighted market return using NYSE, AMEX, and Nasdaq stocks. We calculate a daily market return by summing the abnormal returns each day for all the banks in a specific portfolio of banks selected to represent the comparison index group, and then find the lowest 20 market performance days in a calendar. Next, for each bank we calculate daily abnormal returns and then find their bottom 20 performance days in a given year. LFM Days is the number days in a given calendar year in which the bank and the selected bank index group both have low performance. This measure can vary from 0 (no overlap of low days for bank i and the index) and 20 (the low return days of bank i and the index perfectly overlap). Depending 13

16 on the specific analysis, the portfolio of banks in the index group will be comprised of either all banks in the sample (excluding bank i), or of a group of banks formed on the basis of high cosine similarity with bank i (highly connected peers) or low similarity (weakly connected non-peers). Table 1 shows that the mean value for LFM Days when the index is defined as all banks in the sample (excluding bank i) is 4.12 with a standard deviation of This measure varies from 1 at the 5 th percentile to 9 days at the 95 th percentile. Our second measure, LM AbnRet, is measured as the average of the abnormal returns of bank i over the 20 days where the portfolio of banks in the index group has it lowest performance for the year. Table 1 shows that the mean value for LF AbnRet when the index is defined as all banks in the sample (excluding bank i) is -0.01, with a standard deviation of Univariate Correlations In Table 2 we report univariate correlations between our main variables of interest. While our main qualitative connectedness measure of interest is AvgCos_MDABUS, we see that this measure has Pearson correlation with AvgCos_MDA of.79 and with AvgCos_BUS of.72. AvgCos_MDA and AvgCos_BUS have a Pearson correlation of.56, implying that each measure contains orthogonal information. Also, AvgCos_MDABUS has Pearson correlation with AvgCos_Report of only.29, suggesting that there is substantial scope for our qualitative connectedness measure to contain incremental information about tail comovement relative to quantitative similarity. In terms of tail comovement, Table 2 reports that AvgCos_MDABUS has a Pearson correlation with LFM Days of.11. This implies that banks with higher qualitative connectedness are more likely than less connected banks to have low returns days at the same time that all other banks as a group are also experiencing low returns. Similarly, AvgCos_MDABUS has a Pearson 14

17 correlation with LF AbnRet of -0.08, implying that banks with higher qualitative connectedness have lower average abnormal returns than less connected banks on days when the banking sector as a whole is experiencing low returns. Finally, Table 2 reports that AvgCos_Report has a Pearson correlation with LFM Days of.02 and with LF AbnRet of Empirical Results In this section, we discuss the main empirical results of our analyses of relations between our bank connectedness measures and tail comovement across banks. The section is organized as follows. Section 3.1 discusses our analyses using a bank s average cosine similarity with all other banks to measure connectedness; Section 3.2 examines Granger causality tests of how average cosine similarity conditions the relation between a bank s buy and hold returns and future banking sector returns; Section 3.3 focuses om tail comovement within groups of banks designated as a bank s peer or non-peer banks based on qualitative connectedness; Section 3.4 uses a contagion framework (Boyson et al., 2010) to examine a bank s poor performance is associated with a higher proportion of its peers also experiencing poor performance; and finally, Section 3.5 runs LDA topic analyses and then applies cluster analysis techniques to form groups of banks discussing similar topics in a given year 3.1 Average Cosine Analyses In this section we consider bank connectedness using a bank s average cosine similarity with all other banks in the market. We examine relations between our qualitative and quantitative similarity measures and tail comovement in a multivariate setting using the following regression:!"#$%&'#()&!" =!! +!!!"#$%&_!"#$%&!" +!!!"#$%&_!"#$%&!" +!!!""#$"!" +!!!"#$!" +!!!"#$%&'#()&!"!! +!"#$%& +!!", (1) 15

18 where the RiskMeasure variable will either LFM Days or LM AbnRet as proxies for tail comovement. The coefficient! 1 (! 2 ) captures the extent to which a bank s qualitative (quantitative) similarity to all other banks is associated with a bank s susceptibility to tail comovement. We control for bank size (Assets), the bank s general correlation with the banking sector (Beta), and the lagged RiskMeasure (LFM Days and LM ABnRet). Detailed descriptions of all varibles are contained in the Appendix. The results from the estimation of (1) are reported in Table 3 panels A and B. In Panel A, column one, we only consider our qualitative measure of bank similarity, AvgCos_MDABUS. The coefficient on AvgCos_MDABUS is and is significant at the 0.01 level. This result suggests that the higher the average qualitative similarity of a bank to the rest of the banks in a given year, the more susceptible the bank is to systemic risk. This result is economically significant, where a one standard deviation increase in AvgCos_MDABUS results in a 9% increase in the number of days that bank i and the banking market overlap in their lowest return days. In column 2 we investigate the standalone association between quantitative accounting connectedness, AvgCos_Report, and future tail comovement. We find that the coefficient AvgCos_Report is positive (0.7827) and statistically significant at the 0.01 level. We also find that while the effect of quantitative variable is statistically significant, its economic significance is about one-third that of qualitative connectedness (i.e., 3.4% vs. 9%). However, the difference in economic significance decreases when both the qualitative and quantitative measures are included together in column 3. When both are included, the economic significance of the qualitative measure drops from about 9% to 4.4% and the economic significance of the quantitative measure goes from 3.4% to 3%. 16

19 In Panel B in Table 3, we re-estimate equation (1) using LM AbnRet as the risk measure. Because LM AbnRet is a returns-based measure of poor performance, negative coefficients on our measures of qualitative and quantitative connectedness are consistent with greater future tail comovement. Similar to Panel A, we find that both our qualitative (AvgCos_MDA) and our quantitative (AvgCos_Report) measures of similarity are statistically significantly negative. The results in column 3 suggest that for a one standard deviation increase in our qualitative (quantitative) measure AvgCos_MDABUS (AvgCos_Report), there is a 9.4% (6.1%) reduction in the bank s average abnormal return over the banking market s lowest return days. The results in Table 3 show that both our qualitative and quantitative connectedness measures are significantly associated with a bank s future tail comovement with other banks. Moreover, while both measures have significant associations, the economic significance of our qualitative measure is generally higher than that of our qualitative connectedness measure. This result may seem somewhat surprising, but note that model (1) includes the lagged risk measure, and so some of the explanatory power of the quantitative accounting may be captured in price since they are measured over the same time period. This evidence is consistent with 10-K narrative disclosures reflecting information, including forward looking assessments, that is more timely and pertinent for predicting future tail comovement than is historical accounting information. As a result, our qualitative connectedness measure allows this information to be disentangled from aggregated information in prices and weighted separately in predicting future tail comovement. This can occur because qualitative cosine similarity is measured in the cross section at a point in time, and therefore may not suffer as much from stale data issue necessary for return based calculations. 17

20 3.2 Granger Causality Billio et al. (2012) apply pairwise Granger-causality tests across financial institutions to identify the network of statistically significant Granger-causal relations among these institutions. In this section, we instead implement a Granger causality test of the relation between the stock returns of an individual bank i s returns today and future banking market returns, conditioned on bank i s qualitative similarity. Rather than the focusing on future tail comovement as we did in the previous section, we here examine whether the extent to which a bank Granger causes future banking sector returns is increasing in the bank s qualitative connectedness with all other banks. Specifically, we estimate the following regression:!"#$%&!"#$%&! =!! +!!!"#$!"#$%&!"!! +!!!"#$%&_!"#$%&!" +!!!"#$!"#$%&!"!!!"#$%!!"#$%&!" +!!!"#$%&_!"#$%&!" +!!!"#$%&_!"#$%!!"!"#$!"#$%&!"!! +!, (4) where Market Return is the sum of the buy and hold returns for calendar year t across all banks (less bank i) in our sample. Bank Return is bank i s buy and hold return over the calendar year t- 1. AvgCos_MDABUS and AvgCos_Report are as defined previously. Under this framework a positive!! would indicate that the bank i s return Granger causes future market returns. If our qualitative connectedness measure is powerful in identifying groups that are likely to commove in the future, then we predict that the coefficient on the interaction term,!!, will be positive and significant. We also examine if our quantitative connectedness measure is associated with a higher Granger causality coefficient as reflected in β 5 > 0. We report the results of estimating equation (4) in Table 4. The results in Table 4 show that the coefficient β 3 is positive and statistically different from zero, while the coefficient on β 5 not statistically significant. This evidence suggests that shocks to a bank characterized by high 18

21 qualitative connectedness with other banks in the market have a strong association with future market movements. 3.3 Group Cosine Analysis To further investigate the power of our qualitative similarity measure for predicting future systemic risk we investigate subgroups based on qualitative similarity. A benefit of our methodology in computing similarity is that it allows us to refine our definition from the entire banking market to focus on subgroups within the market that share significant similarities. We use this subgroup analysis to explore the possibility that qualitative similarity can identify groups of banks that are particularly susceptible to tail comovement. We construct subgroups for each bank by selecting other banks with which the given bank is most and least similar based on cosine similarity measures. A bank is similar (not similar) to another bank if their qualitative similarity is above (below) certain cutoff percentiles of their qualitative similarity distribution. The percentiles that we use are 5%, 10%, 25%, and 50% of the qualitative similarity measures distribution. A 5 percent cutoff choses that 5% of all other banks with which a given bank is most or least similar. We repeat this process each calendar year and so allow these groups to evolve dynamically over time. An interesting property of such classification is that for each individual bank, the group of banks that are in close proximity in similarity need not be the same. For example, suppose that for Bank A the banks in the market that are most similar to Bank A are Bank B and Bank C. However for Bank B the two most similar banks in the market are Bank X and Bank Z. It is also possible that the banks in close proximity to a given bank change over time as strategies and circumstances evolve. We develop some descriptive statistics to examine the 19

22 dynamic evolution of peer groups through time. For each bank, in each year we construct a vector that reflects the banks in its peer group that year. The number of elements in these vectors is equal to the number of banks minus 1 (to exclude the bank around which the peer group is built). Then for each bank, we compute the cosine similarity between the vectors for year t and t+1. This cosine similarity provides information on how similar the peer groups are across years. We compute this for every bank in a year and compute the average cosine similarity across banks for the year. We plot the results in Figure 1. While there is evidence of some persistence, there is also evidence of significant change over time in the banks comprising peer groups. If we use 5 percent (50 percent) as cutoffs to determine peer groups, there is a change approximately 50% (20%) in the banks comprising peer groups. Our first test examines the difference in the systemic risk between a bank and the group of banks with which it has highest and lowest similarity. We posit that a bank will exhibits more tail comovement with those banks with which it has high cosine similarity. We run the following regression separately for the high and low cosine similarity groups:!"#$!"#$%&"!" =!! +!!!""#$"!" +!!!"#$!" +!!!"#$!"#$%&"!"!! +!"#!"# +!!", (2) where all variables are defined above, except that the portfolio of banks now used to compute the risk measures LFM Days and LM AbnRet are a given banks specific qualitative similarity subgroups rather thanthe entire market as in our earlier analyses In equation (2) our interest is in the sign and significance of!!. When we use the LFM Days (LM AbnRet) we expect a positive (negative) intercept for both the high and low groups but the magnitude should be significantly more positive (negative) for the high group compared to the low group. The results for the 20

23 estimation of (2) are reported in Table 5 panels A and B. As reported, the results are consistent with our predictions that for each of the similarity cutoffs, high cosine similarity groups have a stronger association with future systemic risk than do low cosine similarity groups. With the exception of the 25% cutoff in Panel A, it should also be noted that the difference between the high and low monotonically increases as we move from the 50% to 5%. These results provide unique evidence that our qualitative cosine similarity measure is able to identify subgroups within a bank that are more likely in the future to commove in the tails. 3.4 Proportion Low Performance Our analyses of qualitative connectedness so far examine the extent to which the returns of a bank are low when conditioning on whether similar banks are experiencing low returns. We now turn this around and examine whether the returns of the group of banks with which an individual bank is similar are low when conditioning on the individual bank s returns being low. As noted in Adrian and Brunnermeier (2016), the incidence of a bank s tail outcomes conditioned on peer banks having tail outcomes need not be the same as the incidence of peer banks having tail outcomes conditioned on an individual bank having a tail outcome. We build on Boyson et al. (2010) and regress the proportion of banks in an individual bank s peer (nonpeer) group experiencing low performance on a given day on whether or not the individual bank is also having a low performance day. Specifically we estimate the following models by high and low cosine group for each bank:!"#$#"%&#'!"#_! =!! +!!!"#!"#!,! +!"#$!! +!"#$%& +! (3) 21

24 where Proportion Low is the proportion of banks in a given bank s subgroup on day t that are have a low performance day. Low Day is an indicator variable set to 1 for bank i on day t if its daily abnormal return is in the bottom 5% of the entire set of its daily returns, and zero otherwise. We predict that when bank i is experiencing a low return day, a higher (lower) percentage of banks in its high (low) cosine similarity subgroup will also experience a low return. We estimate equation (3) and report the results in Table 6. Similar to Table 5, we again present the results for different group similarity cutoffs (i.e., 5%, 10%, 25%, and 50%). Consistent with our predictions, we find that there is a positive coefficient on Low Day for the high cosine group. We also find that the effect is monotonically increasing as we move from the 50% cutoff to the 5% cutoff. For the low cosine groups we find that coefficient on Low Day is negative and significant for all cutoffs, suggesting that an individual bank has less influences on other banks with which it is least connected. 3.5 Topic Analysis In this final set of analyses, we more fully exploit the texture of the text in banks 10-Ks by examining the topics discussed in the business description section and MD&A disclosures using Latent Dirichlet Allocation (hereafter LDA). LDA is a textual analysis technique which finds topics in unstructured text. We use the Gensim implementation of LDA in Python as the basis for our analysis. We keep only nouns and combined words (discussed below) for use in our analysis. Following prior studies, we set alpha = 50 / number of topics, eta = 0.025, and decay = 0.5 to improve the separation of the topics. The total number of topics is set to 50, similar to prior studies, and the topics are estimated based upon the entire sample of MD&A s. We perform 22

25 5 passes to stabilize the topics. After the 50 topics are estimated we calculate the relation between each topic and each MD&A in our sample. We prepared the business description section and MD&A disclosures for topic analysis by first combining words which are meant to be interpreted together. For example, the phrase balance sheet would be interpreted as two independent words in a typical LDA analysis, thereby potentially leading to a misinterpretation of the meaning conveyed by the phrase. To address this issue we perform a part of speech analysis (hereafter POS) of each sentence in the MD&A s. Our POS analysis combines adjacent adjectives and nouns and treats the combined words as a single word in the LDA analysis. This technique is similar to the common gram technique used in prior studies. However, unlike prior research, this method explicitly uses the structure of each sentence to determine what words should be combined Clustering Banks by Topic Similarity Our objective in this analysis is examine the extent to which fragility of the banking system is reflected in how intensively banks concentrate their 10-K discussions on similar topics. The idea is that when a large concentration of banks cluster around similar topic discussions, such concentrated connectedness makes the system more vulnerable to downside comovement. We measure the degree to which financial institutions discuss similar topics in their MD&A by performing cluster analysis of the topics from the LDA analysis. Specifically, we use agglomerative clustering (i.e. each financial institutional begins as its own cluster) with the Ward minimum variance method as our objective function. The number of clusters is set to 50. We then measure the clustering of financial institutions around topics as either the percentage of 11 The following is an example of the output from and LDA analysis of the sentence The numbers in the balance sheet follow GAAP : The (DT) numbers (NN) in (PRP) the (DT) balance (NN) sheet (NN) follow (VB) GAAP (NN); where DT = Determiner; NN = Noun; PRP = Preposition;VB = Verb. 23

26 financial institutions in the largest cluster, or the skewness of the number of financial institutions across the 50 clusters in each year. In Figure 2, we provide descriptive statistics on the percentage of all banks that are in the largest topic cluster in a given year. There is significant variation across years where for example, the largest cluster was comprised of 8% of all banks in 1999, while in 2005 the largest cluster comprised 22% of all banks. We include two vertical lines to designate the timing of the recession in and the beginning of the financial crisis after We see that following a negative shock to the economy the percentage of all banks in the largest cluster decreases substantially, suggesting divergence in the responses across banks to the shock. We also find that the topics discussed by banks in the largest cluster changes significantly after an economic shock. To see this, each year we compute the average of the topic loadings on each one of the 50 possible topics across all the banks in the largest cluster by year. We then form a vector consisting of the average topic loadings on each of the 50 topics in each year. The difference in the topics discussed in the largest cluster is then measured by calculating the cosine similarity between the vector of average topic loadings in year t and year t +1. A high cosine similarity indicates that the mix of topics discussed by banks in the largest is similar across years, and vice versa for low cosine similarity between topic loading vectors. We report the results of this exercise in Figure 3, where again we include vertical lines to designate economic shocks. We see that prior to the two financial shocks there is high cosine similarity in the topics discussed by the largest clusters in contiguous years. However, following the shocks the topics of discussion change significantly as indicated by a significant drop in cosine similarity Topic Clustering Concentration and Bank Failures 24