Different ways of managing risk as reported in 10-Ks: A supervised learning approach

We use supervised learning on annual reports of publicly listed US firms (10-Ks) to build textual measures of risk management via derivatives, insurance, diversification, long-run contracts, and credit lines. Validation exercises favor these supervised learning-based measures over those based on word lists. Panel regressions (1996–2015) indicate that firms using one form of risk management are more likely to also use other forms. In contrast, intensive use of one risk management technique associates with less intensive use of other methods. Findings are consistent with a model featur-ing fixed costs of organizational capacity for managing risks and increasing marginal costs of hedging.

A challenge that researchers face is that many aspects of risk management are difficult to observe; for instance, use of insurance is not available in any standard data sets. To circumvent this problem, we develop measures of risk management based on textual analysis of annual reports (10-Ks) that we combine with financial data from COMPUS-TAT. Several researchers have previously applied textual analysis to 10-Ks (see, e.g., Loughran & McDonald, 2016 for a survey) but in terms of measurement this article contributes to the existing literature in two ways: (i) it broadens the scope of textual analysis of risk management to include, for instance, use of insurance, and (ii) rather than relying on prespecified word lists, a supervised learning algorithm is applied to identify sentences that discuss ways of managing risk. First, we manually read and annotate a randomly selected set of 10-Ks and train the algorithm on these data.
We then use our algorithm to read and classify all text in Items 1 and 1A of US annual reports filed with the Securities and Exchange Commission (SEC) between fiscal years 1996 and 2015. We validate our text-based measures using outside measures of risk management that are available for subsets of our sample (e.g., an indicator of derivatives use based on COMPUSTAT data and credit line use from Sufi, 2009). When we examine externally available measures of derivatives use (from COMPUSTAT and from Almeida et al., 2014), we find that the supervised learning algorithm is more accurate than a word list in measuring risk management, a contribution that we believe will be valuable for future work.
Our textual analysis delivers measures on five risk management practices that vary by firm and fiscal year: derivatives, insurance, credit lines, long-run contracts, and diversification. A large body of literature examines different ways of managing risk, but it typically focuses on only one way of managing risk in a given paper. 1 In contrast, our key focus is on how different ways of managing risk interact. Are firms that use one way of managing risk also using other ways of managing risk? In the empirical literature on risk management, this question has largely been framed in terms of whether different ways of managing risks are "complements" or "substitutes." 2 As we discuss later, some risks can almost exclusively be handled by one risk management tool, while several other risks can at least partly be handled with multiple tools. For instance, an exporter aiming to limit exchange rate exposure can use foreign currency derivative contracts to hedge, but can also hedge operationally by producing locally in foreign markets. Such logic would suggest that financial hedging and operational hedging are alternatives or substitutes.
However, one can also envision circumstances when financial hedging and operational hedging are observed simultaneously and are complements-for instance if different tools are suited for different horizons. The literature that examines the combined use of different ways of managing risk has been especially focused on the relation between the use of currency derivatives and international operations, where the latter is taken as a measure of operational hedging. Painting with a broad brush, the results in Allayannis et al. (2001), Kim et al. (2006), and Bartram et al. (2010) all suggest complementary effects of financial hedging and operational hedging. An important strand of the literature thus interprets the results as supporting the complements interpretation.
On the other hand, substantial research supports the substitutes interpretation. In particular, there is robust evidence that cash holding is negatively related to diversification, thus indicating that operational hedging (i.e., diversified operations) lowers the need for precautionary cash holding (Bakke & Gu, 2017;Duchin, 2010). Other dimensions of the interaction between different ways of managing risk are less explored, but there are several additional investigations that also support the substitutes interpretation. 3 1 The literature is especially rich when it comes to precautionary cash holding (Almeida et al., 2014;Bates et al., 2009) and financial hedging (Bodnar et al., 1998;Geczy et al., 1997).
At first glance, the previous literature thus seems divided on whether different ways of managing risks are substitutes or complements. To understand the varying results, we note that previous literature has frequently been hampered by having only indicator variables of risk management (e.g., a dummy variable to capture if the firm uses derivatives). We show that the results differ substantially depending on whether indicator variables or continuous measures of risk management are used. All the risk management tools we investigate are positively correlated when we use indicator variables to capture different ways of managing risk. 4 For instance, a firm that uses insurance to deal with risks is also more likely to use derivatives. If we instead capture the extent of risk management by continuous measures, a negative relation dominates. For instance, more diversified firms use less derivatives, insurance, credit lines and long-run contracts and hold less cash. An exception to this negative pattern is that the use of credit lines, insurance, and derivatives tends to be positively correlated.
The mostly positive relation between indicator variables of different ways of managing risk at the same time as the overall relations are mostly negative may initially seem puzzling. An important contribution of this article is not only to consider a wider set of ways managing risk, and how they are combined, but also to provide a model-based interpretation of results that appear contradictory at first. We show how fixed costs of building capacity to use a specific tool to manage risk, combined with increasing marginal costs to use a certain way to cover a particular risk, can generate the observed pattern. Fixed costs of building capacity may, for instance, reflect the costs of creating management processes and hiring the critical level of staff needed to use derivatives or to commit to long-term contracts with suppliers.
Increasing marginal costs of using a particular form of risk management capture that-for example, some exposures are likely to be relatively cheap to cover with derivatives but the more exposures that a firm covers with derivatives, the more complex instruments and analysis are needed.
Consider a car manufacturer, for instance. Covering some exposures (such as short-term exposure to the euro) with derivatives is likely to be cheap, at the same time as covering other exposures with long-run supplier contracts is also likely to be cheap. Conditional on having taken the fixed costs, it is cheaper for such a firm to use both derivatives and long-run contracts to manage exposures rather than to try to cover all risks with only derivatives or only with longrun contracts. Fixed costs of capacity and increasing marginal costs for using a particular tool to manage risk can thus explain why larger firms are more likely to engage in risk management and why several ways of risk management tend to be used in tandem. On the other hand, if one form of managing risks becomes more expensive-for instance, if longterm contracts become costlier because of changing legal rules-then such a firm would naturally tend to rely less on long-term contracts and more on derivatives. That is, the intensity of use of different ways of managing risk would tend to be negatively related.
One simple intuition for our findings notes that it is not puzzling if a consumer who enjoys fruit buys both apples and pears (positive indicator variables for both) but buys more apples and fewer pears if the price of pears increases (goods are substitutes). This is also the reason why we diverge from much of the previous literature and avoid the commonly used terms substitutes and complements to interpret our results. In economic theory, these terms have a precise definition tied to the sign of cross-price effects, whereas the use of these terms in the risk management literature varies to some extent between articles and is not always clearly defined.
The next section presents the data and the text-based measures of risk management. 5 Section 3 provides validation exercises for the text-based measures of risk management. Section 4 presents the main empirical analysis and interprets the findings. Section 5 concludes. 4 The exception is cash holding, which is not well captured by an indicator variable. 5 An Appendix in the Supporting Information with details on how the textual data were assembled as well as many robustness exercises is available in the supporting materials section online.

Building measures of risk management from annual reports
We create measures of risk management by textual analysis of Items 1 and 1A in the annual reports (form 10-K) of US firms. 6 The regulations stipulate that Item 1 describes the general development of the business of the filing firm: for instance, principal products, dependence on a limited set of customers, and competitive conditions are topics that must be discussed. Item 1A requires a concise discussion of the most significant factors that make the offering risky.
Many firms discuss risk factors from the start of the sample period, but it is required only from 2005 onward. For brevity, we simply refer to these sections as Item 1 such that if not stated otherwise, Item 1 is considered to include Item 1A.
We download all 10-Ks from the SEC EDGAR servers for fiscal years 1995-2015 and identify Item 1. After excluding financial firms, this gives us a final data set that includes 75,627 annual reports that each, on average, has 343 sentences in Item 1. To analyze the extent to which these texts discuss risk management, we rely on two methods. The first is a supervised learning algorithm that relies on manual reading and classification, and the second relies on a word count that uses prespecified dictionaries.

A supervised learning algorithm
We first develop a sentence classification algorithm that is trained on a randomly selected, manually annotated set of Item 1's. The training data set comprised a random draw of 10% of Item 1's for calendar years 2006 and 1998. Through the lens of three broad classes of risk management practices (financial hedging, operational hedging, and access to liquidity), we manually read and tag sentences dealing with risk management in the training data. We now provide a brief summary of the classification scheme as well as the training of the classifier. The Appendix in the Supporting Information presents detailed accounts of all aspects of the supervised learning algorithm.
First, we use the term financial hedging to denote contracts with financial counterparties that create an offsetting stream of revenue in the case that certain states of the world materialize. We consider two such forms of risk management: (i) derivatives (we make a broad interpretation and include options, futures, forwards, swaps, and other statecontingent contracts with financial partners) and (ii) insurance. 7 An example of a derivatives sentence is, "The Company has a fuel hedging program in which it enters into jet fuel, heating oil and crude oil hedging contracts to dampen the impact of the volatility of jet fuel prices." An example of a sentence capturing the use of insurance is, "We carry insurance for public liability, passenger liability, property damage and all-risk coverage for damage to our aircraft." Both these examples are from the 10-K of AMR (2006), the parent of American Airlines.
Second, the firm may act to modify the relation between operating profits and the state of the world. We refer to such measures as operational hedging, and in the analysis we use sentences that encompass five different forms of operational hedging. 8 We combine three of these (broad customer base, broad supplier base, and other diversification) under the heading "diversification." 9 To obtain the flavor of the type of sentences tagged, we exemplify a broad supplier base with Chiquita (2006): "The geographic diversity of our suppliers helps us to lessen the effects of any event that 6 See https://www.sec.gov/about/forms/form10-k.pdf. Information that must be included in these sections of 10-K is stipulated in Items 101 and 503(c) of Regulation S-K under the US Securities Act of 1933. 7 The use of the former are relatively well documented in the previous research (see, e.g., Geczy et al., 1997;Guay & Kothari, 2003), whereas the empirical literature on insurance is limited and mainly comprises case studies (e.g., reinsurance of catastrophe risk in Froot, 2001, or property insurance in Aunon-Nerin & Ehling, 2008. could cause production declines in a single region." A broad customer base can be exemplified with Ikon (2006), "Our customer base is broad and diverse, numbering approximately 500,000 organizations, including global and national companies, representing over 100 of the Fortune 500 companies, major regional companies, mid-size businesses, professional services firms, and state, local and federal government agencies." Other diversification is illustrated with a sentence from Alico (2006): "The assets of Plant World were purchased for the purpose of diversifying Alico's agricultural operations." Two other classes capture long-run contracts with suppliers and customers that contain a risk management dimension: one tags sentences capturing long-run contracts per se, and the other tags renegotiation clauses in long-run contracts. An example of a contract relating to long-run contracts is from Chiquita (2006): "For lettuce, we often enter into contracts with these growers to help mitigate supply risk and mitigate exposure to cost risks." Renegotiation clauses are illustrated by Coca-Cola (2007): "This baseline price may be adjusted periodically by the Company, up to a maximum indexed ceiling price, and is adjusted quarterly based on changes in certain sugar or sweetener prices, as applicable." Both the above ways of managing risk rely on ex ante measures that modify outcomes in different states of the world. In a third way to manage risk, the firm may undertake measures that strive to provide it with access to liquidity should adverse states materialize (or investment opportunities). A strong balance sheet can broadly provide access to liquidity. We follow previous literature and focus on two ways of providing liquidity: precautionary cash holding (Almeida et al., 2014;Bates et al., 2009) and lines of credit (Campello et al., 2011;Sufi, 2009). To capture this form of risk management we build a text-based measure of credit lines and rely on COMPUSTAT data to capture cash holdings.
The manually identified text serves as a training ground for the Stanford Classifier 10 from the Stanford Natural Language Processing Group in identifying sentences related to risk management in Item 1. It is a maximum entropy classifier of a type commonly used in machine learning.
Having trained the algorithm, we use it to classify all the text in Item 1 of all annual reports between fiscal years 1996 and 2015 and assign sentences to risk management categories. This procedure generates a "sentence-level" data set, which contains all the identified risk management sentences. We also store the length of the identified Item 1.
From these sentence counts, we create the following text-based measures of risk management: Derivatives it is defined as the number of derivative sentences for firm i in fiscal year t divided by the total number of sentences in Item 1 for firm i in fiscal year t. Insurance it , Credit lines it , Long-run contracts it , and Diversification it are calculated in the same way. All five text-based variables are thus continuous measures and we multiply the sentence shares by 100 such that they are expressed in percentage. We also create analogous indicator variables (0/1) that take the value one when we identify at least a single sentence that discusses the particular form of risk management for firm i in fiscal year t and denote it as > 0. Thus, for instance Derivatives it > 0 is 1 for a user of derivatives. Table 1 provides summary statistics on the measures of risk management. 11 Note that derivatives, insurance, and credit lines are zero for many observations. In contrast, long-run contracts and diversification are discussed by nearly all firms, but there is substantial variation. These findings highlight that if we only focus on derivatives in an analysis of risk management, we are missing large parts of the picture. The summary statistics also indicate that the 0/1 margin of different ways of managing risk may play an important role.

An alternative approach to textual analysis: Dictionaries
The supervised learning methods that we use are labor intensive because they require manual annotation of substantial bodies of text. So far, textual analysis of 10-Ks has instead largely relied on a dictionary approach where a list of 10 See https://nlp.stanford.edu/software/classifier.html 11 There are 75,626 observations on the text-based measures, but the COMPUSTAT data are not available for all these firms and years. Tables 1 and 2 therefore give summary statistics for the 48,560 observations that form the core regression sample used in Tables 5-7. In some of the later validation tests, the number of observations is lower yet, because the external measures of risk management that we compare our text-based measures against are available for only a subset of observations.  (Tables 5-7) of the share of sentences in Item 1 of 10-Ks that discuss use (positive mentions) of the relative risk-management topic in the heading during fiscal years 1996-2015.
Sentences are identified using a supervised learning algorithm, as described in the text.
words (or terms) is created and the occurrence of those words are counted in each 10-K. We complement our supervised learning methods with such a dictionary approach. If the two approaches yield similar results, this would be a valuable input for future research that could then rely on a simpler method with little loss of precision. In contrast, if the results differ, the expected gain in precision will have to be weighted against an increased cost of data collection.
Using the same sample as above (Item 1 from 10-Ks), we check for words and phrases from a set of dictionaries. The dictionaries are detailed in the Appendix in the Supporting Information. For derivatives, we follow the dictionary from Disatnik et al. (2014), and for insurance, we create a new set of words related to insurance. For aspects of operational hedging, we create a set of words related to operational hedging. In addition, we build a set of words that aim to capture flexibility, the latter inspired by the literature on real options (e.g., Chevalier-Roignant & Trigeorgis, 2011). For access to liquidity, we use a word list that contains a set of words and terms capturing precautionary motives for access to liquidity as well as a set of terms related to credit lines that follow (Sufi, 2009). Finally, we also used a word list to capture if firms engage in enterprise risk management (ERM, Bromiley et al., 2015). Table 2 presents summary statistics on these word counts (expressed as shares of the total number of words in the respective Item 1). In all these cases, there is substantial variation; only in the case of derivatives is the mean greater than the standard deviation. As observed in the right-most column, the share of ERM-related words is minuscule; also at the 95th percentile, the share is zero. 12 The bottom row presents the correlation with the most closely related measure based on supervised learning that we use in the main analysis. The correlations are positive. For insurance the correlation is relatively high at 0.64 but TA B L E 2 Summary statistics on text-based measures of risk management

Other data
We complement the text-based data on risk management with financial information on these firms from COMPUSTAT and match it by fiscal year. 13 We thus calculate cash holding, cash flows, and leverage, all scaled by total assets. 14 We also include market-to-book in regressions below. 15 In addition, we use COMPUSTAT segment files to generate information on the number of four-digit Standard Industrial Classification (SIC) segments in which a firm is active and information on the number of geographical segments. This is available for fiscal years 1997-2013 inclusive.

VALIDATION EXERCISES
An assumption in this article is that the share of sentences that are identified as relating to a particular form of risk management is informative about the relative importance of managing risk in that way. A foundation for the expectation of a positive correlation between sentence share and intensity of use stems from the evidence that textual risk disclosures in 10-Ks are informative to investors (Campbell et al., 2014;Kravet & Muslu, 2013;Loughran & McDonald, 2016). Furthermore, a useful comparison can be made between some of our text-based measures and alternative measures. In the following, we explore if such comparisons support the notion that the text-based measures are TA B L E 3 Cross-tabulation between text-based measure of derivatives use and COMPUSTAT-based measure ond, the share of users of derivatives is higher in the text-based definition (38%) than in the COMPUSTAT definition (27%). This is expected as a firm may use derivatives but may not have any unrealized gains or ineffective hedges, as defined by AOCIDERGL or HEDGEGL. This also means that 22% of observations in the lower left-hand corner are not an indictment of the textual analysis. Somewhat more worrying is that 11% of observations use derivatives according to COMPUSTAT but not according to our algorithm.
We proceed to examine the intensity of derivatives use and compare it with the dichotomous COMPUSTAT-based variable. The top row of Panel A in Table 4 first provides the mean (standard errors in parentheses) of our main text based measure of using derivatives for all observations; next, in column 2 for observations that are not defined as hedgers under the COMPUSTAT-based definition and in column 3 for observations that are defined as hedgers following the COMPUSTAT-based definition. The share of sentences that are defined as pertaining to using derivatives is higher for firms that are defined as hedgers following the COMPUSTAT-based definition and the difference is both statistically and economically significant, thus confirming the impression from the cross-tabulation in Table 3.
The second row of Panel A in Table 4 makes the same comparison for our dictionary-based measure of derivatives, and as can be observed, the relation with the COMPUSTAT measure is much weaker. The t-statistic is sufficiently high that equality can be rejected at the 1% level; however, economically, the difference is trivial. Although the COMPUS- TAT measure of using derivatives should not necessarily be seen as "correct," the first two lines nevertheless strongly support our supervised learning over a dictionary approach. We also trained the algorithm to explicitly predict the nonuse of risk management. This proved less successful than predicting use and we did not use these data in the present article. However, it provides yet another opportunity to check the algorithm. The third row, therefore, gives the share of sentences explicitly discussing the non-use of derivatives and, as observed, the share is much lower than the share of sentences discussing the use of derivatives. The share of non-use sentences is also lower among firms that Panel B uses another potential comparison with COMPUSTAT data and compares our text-based measure of diversification across firms that operate in only one four-digit SIC industry (following COMPUSTAT segment files) to those 16 Note that a nonzero share here may at first appear as a predictive failure, which it need not be. For instance, a firm may use one sentence to say, "We do not use weather derivatives to hedge exposure," in the same document as it states that "We use forward contracts to hedge euro exposure." A single positive sentence is sufficient to make it a derivatives user, but one negative sentence is not enough to make it a nonuser.
that operate in several industries. As expected, the share of diversification sentences is higher in the latter group, and the difference is statistically significant at the 1% level.
Panel C uses of a list of 42 six-digit North American Industry Classification System (NAICS) industries with traded futures from Almeida et al. (2017), Appendix B). As expected, when using our text-based measure, we observe a marked difference between the share of derivative sentences among firms in industries with traded futures and those without, and the average share is about five times higher in the former group. Instead, if we consider word counts, we see the same qualitative pattern, but it is much less pronounced.
Panel D splits the sample by an indicator variable from Sufi (2009)  Regulations also require that the 10-Ks provide information about market risks and the use of derivatives to manage, for example, commodity and currency risk. 17 The related information is to be found in Item 7A, but an initial manual reading indicated that the information is often cross-referenced to other sections and included in tables, which make it harder to apply automated textual analysis. Panel E relies on a manual reading of a stratified random sample of 100 10-Ks to identify firms that use currency or commodity derivatives. Our supervised learning-based measure of derivatives use is significantly higher for firms for which a manual reading of the information (referenced to) in Item 7A indicates that they use commodity or currency derivatives (see the Appendix in the Supporting Information for further discussion).
We draw two conclusions from the analysis in this section. First, text-based analysis of Item 1 in 10-Ks provides information on firms' risk management practices. As such, it echoes findings from literature that examines the impact of 10-K text on investor perceptions and actions (e.g., Kravet & Muslu, 2013). We expect the information to be noisy, but it provides a useful complement to other noisy measures of risk management that are potentially available, such as evidence from questionnaires with low response rates. Second, at least in the case of risk management, it appears that supervised learning methods provide more precise results than a dictionary approach.

Empirical specification
We rely on regression analysis to examine the various ways in which firms manage risk. The regressions follow the general format of The dependent variable Y it is either cash holding or one of the text-based measures of risk management (Derivatives it , Insurance it , Credit lines it , Long-run contracts it , or Diversification it ). We consider both regressions in levels, and regressions where risk management is captured by indicator variables that take the value one if this form of risk management is discussed, and zero otherwise.
Typically, we include firm fixed effects i , which imply that identification comes from within-firm changes across fiscal years in how much the respective firm discusses a particular form of risk management. We use fiscal year fixed effects t to capture overall developments in how firms discuss risk management. We also include a set of financial variables that are common control variables in corporate finance: cash flows, leverage, market-to-book, and the natural logarithm of total assets. To limit endogeneity concerns, explanatory variables are lagged one fiscal year.
17 This is stipulated in Item 305 of Regulation S-K.
A potential concern is the likely measurement error in the text-based variables. For instance, although it seems imminently plausible that a firm that assigns more importance to insurance in its management of risk spends a larger share of Item 1 discussing insurance, it is also clear that we only rarely expect a one-to-one relationship between the amounts covered by insurance contracts and the share of sentences discussing insurance. 18 To the extent that measurement error impacts the dependent variable, ordinary least squares (OLS) will remain consistent, but standard errors on coefficients will tend to increase. To the extent that measurement errors affect the explanatory variables, it is likely that there will be some attenuation bias-that is, coefficients will be biased toward zero. As our primary interest is the sign of relations (rather than magnitudes per se), the prime effect of measurement error is, if anything, to make it too easy to reject a hypothesis (by potentially biasing coefficients toward zero and inflating standard errors).
To consider the potential correlation of the error term within firms, we cluster standard errors at the firm level. Table 5 presents the results of a regression analysis of the interrelations between the different continuous measures of risk management that we consider.

Analysis with continuous measures of risk management
Column 1 indicates that firms hold less cash when they use more derivatives, credit lines, or are more diversified.
The literature on cash holding is particularly rich, and the results here are consistent with previous findings (see, e.g., Bates et al., 2009;Duchin, 2010). The other regressions are more novel, and rather than commenting on every individual coefficient, let us consider the overall picture that they paint with regards to the coefficients on risk management.
Out of a total 30 coefficients, 14 are negative and statistically significant at the 5% level or greater. An additional two coefficients are negative and statistically significant at the 10% level or greater. Only four coefficients are positive and statistically significant. This pattern indicates that by and large, different means of managing risk are alternatives; if a firm uses one form of risk management more, it tends to use less of the other forms.
The four positive and statistically significant coefficients indicate positive relations between credit lines on the one hand and derivatives and insurance on the other hand. We provide an extended discussion of the documented patterns below, but for now, note that the positive relations involve risk management instruments that are developed in association with the financial industry.
Any causal interpretation of the patterns that we document should be done with care. Our preferred interpretation is that our tables report partial correlations, after potentially confounding factors have been controlled for via the inclusion of firm fixed effects, fiscal year fixed effects, and some key accounting variables. Overall, the economic significance of the estimated coefficients is nontrivial. Using results in column 2 to exemplify the economic significance, we find that a one-standard-deviation increase in the share of diversification sentences is associated with a decrease in the share of derivatives sentences by 13.2% and a one-standard-deviation increase in the share of long-run contract sentences with a decrease in derivatives sentences by 6.6%. A one-standard-deviation increase in the share of credit line sentences is associated with an increase in the share of derivative sentences by 14.5%.
One puzzling finding in the prior literature is that large firms are more likely to use financial hedges (see, e.g., Geczy et al., 1997 for an early reference). This is surprising because factors leading to a value of risk management explored by theory, such as nondiversified ownership and credit constraints, are typically deemed to be more important for smaller firms. 19 A potential solution to the conundrum may be that smaller firms manage risk using other tools instead. Using ln(Assets) as a proxy for size we see no such pattern in Table 5. 18 Presuming that we want to consider the insurance variable as capturing the insured amounts. Clearly, the set of risks that are covered, the extent of deductibles, and other variables will potentially affect the extent to which insurance is discussed. 19 The Fisher separation theorem implies that a firm's goal should be to maximize its discounted net present value, irrespective of the owners' risk or time preferences (see, e.g., Stein, 2003). Even though it is an important benchmark result note that risk aversion creates a value of risk management because the expected value of a concave function is decreasing in the variance of shocks (and risk aversion is tantamount to a concave utility function). However, risk aversion is only one reason for why the value of the firm can be a concave function of shocks. Other reasons for concavity, which are not directly affected by a diversified ownership, are transaction costs of financial distress or convex tax schedules (Smith & Stulz, 1985) as well as the need to maintain sufficient funds Note: This table presents OLS regressions with, as dependent variables, text-based measures of risk management using the share of sentences in Item 1 of 10-Ks that discuss use (positive mentions) of the particular risk-management topic in the heading during fiscal years 1996-2015. Sentences are identified using a supervised learning algorithm as described in the text. Cash, cash flow, leverage, market-to-book, and ln(Assets) calculated based on COMPUSTAT and winsorized at 1, and 99%. ***, **, and * indicate statistical significance at the 0.01, 0.05, and 0.1 levels, respectively. Abbreviation: FE, fixed effect.
Note, however, that with firm fixed effects, the coefficient on size is identified only from year-to-year variation within firms. To examine the relation between size and risk management more broadly, we estimate the corresponding regressions as above but with industry fixed effects (48 industries following the Fama-French classification 20 ) instead of firm fixed effects. The results are reported in Table 6. Our data do not support the hypothesis that large firms manage risk to a lesser extent; with the exception of cash holding, none of the coefficients on the proxy for size, ln(Assets), is negative and statistically significant, which is what we would expect if large firms managed risk to a lesser extent. The to be able to access capital markets when there are financial constraints (Froot et al., 1993). Clearly, motivations may differ across firms and across time, and research has also established cases where other motivations play a role. For instance, in an early influential paper, Tufano (1996)   Note: This table presents OLS regressions with, as dependent variables, text-based measures of risk management using the share of sentences in Item 1 of 10-Ks that discuss use (positive mentions) of the particular risk-management topic in the heading during fiscal years 1996-2015. Sentences are identified using a supervised learning algorithm as described in the text. Industry fixed effects following Fama-French classification with 48 industries. Cash, cash flow, leverage, market-to-book, and ln(Assets) calculated based on COMPUSTAT and winsorized at 1, and 99%. Robust standard errors clustered by firm. ***, **, and * indicate statistical significance at the 0.01, 0.05, and 0.1 levels, respectively. Abbreviation: FE, fixed effect.
results instead support the notion that larger firms engage more in risk management, in some sense deepening the puzzle of the relation between risk management and size. We will return to this issue in our model-based discussion below.

Analysis with indicator variables of risk management
We may be interested in not only understanding how much a particular risk management practice is discussed, but also if it is discussed at all. To examine this, we use regressions that rely on indicator variables that take the value 1 if a particular form of risk management is discussed and 0 if it is not. The estimation uses linear probability. Column 1 TA B L E 7 Relation between different measures of risk management using indicator variables (0/1) to capture use of risk management tools Note: This table presents OLS regressions with, as dependent variables, text-based measures of risk management using the share of sentences in Item 1 of 10-Ks that discuss use (positive mentions) of the particular risk-management topic in the heading during fiscal years 1996-2015. Sentences are identified using a supervised learning algorithm as described in the text. All text-based measures defined as one if the topic is discussed by firm i in fiscal year t, and 0 otherwise. Cash, cash flow, leverage, market-to-book, and ln(Assets) calculated based on COMPUSTAT and winsorized at 1, and 99%. Robust standard errors clustered by firm. ***, **, and * indicate statistical significance at the 0.01, 0.05, and 0.1 levels, respectively. Abbreviation: FE, fixed effect.
of Table 7 uses the continuous measure of cash holding as the dependent variable. We note that a firm reporting the use of derivatives, credit lines, or diversification is associated with a lower level of cash holding. On average, a firm that discusses use of derivatives holds around 0.6 percentage points less cash-that is, a cash over asset ratio of 20.9% rather than 21.5%. The negative relationship between cash holding and the other means of managing risk is also seen by the negative coefficient on cash holding in columns 2-6.
In contrast, the analysis suggests that relationships between text-based measures of risk management are mostly positive. Of 20 coefficients, 13 are positive and statistically significant at the 5% level or higher (an additional two are positive and statistically significant at the 10% level). For instance, a firm that reports using insurance is around 3.8 percentage points more likely to report using derivatives. Table 7 is estimated with linear probability, which has the advantage of easy interpretation. In the Appendix in the Supporting Information, we report logit and probit estimation results of corresponding regressions as in Table 7 and note that the pattern with a positive relation between text-based measures of risk management are clear also in these alternative estimations.

Relation between different ways of managing risk: A simple model-based discussion
Summarizing our investigation, results indicate that (i) firms that use one form of managing risk are more likely to also use other forms of risk management, (ii) a more intensive use of one form of managing risk is associated with less intensive use of other forms of managing risk (with the exception of insurance, derivatives, and credit lines that tend to be positively associated), and (iii) larger firms use all types of risk management tools more (apart from cash holding).
In the following section, we sketch a simple model to show how such a combination of findings could arise. Let us first present the motivation underlying the different ingredients of the model. Economies of scale and scope in building up the ability to manage risk in a particular way is one part of the model. Several researchers have previously hypothesized that fixed costs of building capacity for using derivatives can explain why large firms are more likely to use derivatives (see, e.g., Bodnar et al., 1998). We build on this intuition and also include economies of scope, which would imply, for instance, that if a firm takes the fixed cost associated with the capacity to use (and monitor) derivatives, this lowers its fixed costs of having the ability to apply other ways of managing risk.
The fixed costs for building capacity to manage risk with economies of scope across different ways of managing risk is consistent with the findings that large firms manage risk to a greater extent and cash holdings decrease with firm size. Cash holdings mostly have a negative relation with the dichotomous measures of risk management, which may reflect cash holding as a sort of default form of risk management: higher levels of cash holding is a risk management strategy that is available to all firms. On the one hand, cash holding is an attractive form of risk management as it has value across many states of the world, which limits the need for sophisticated analysis. 21 On the other hand, cash may be a costly form of managing risk as it ties up resources and is likely to lower investment.
Next, we focus on the cost of using a specific tool for a particular risk. In the analysis above, we did not attempt to model the various sources of risk a firm is exposed to. Instead we presented a bird's-eye perspective of different ways through which risk management methods are combined. Clearly, many nuances are glossed over at this level of detail, including that some risks may be handled in multiple ways with relative ease (for instance, exchange rate risk may be handled with derivatives, with long-term contracts with customers or suppliers, or by diversifying), all of which may be more or less appropriate for different horizons. Other risks, such as major property damage, may be much more suited for only one way of managing risk (insurance in this case). A broad implication of this is that some risks are relatively cheap/easy to manage with a specific tool, and the more risks that one attempts to cover with this particular tool, the more costly it becomes.
We formalize the intuition from the previous paragraph by assuming that there are convex costs of using a particular way of managing risk. This is consistent with a situation where it becomes increasingly costly to use a specific form of risk management the more risks that a firm seeks to use it for. Consider exchange rate risk, for instance. It may be easy for a US firm to hedge transaction exposure to the euro at the 90-day horizon; however, longer horizons, less liquid currencies, and other factors may escalate marginal costs as the firm tries to hedge increasingly complex exposures. Increasing marginal costs may be linked to the costs of increasingly nonstandard contracts but also to a need for more accurate predictions and more sophisticated analysis. 22 Increasing marginal costs would imply that in a situation with multiple risks, and conditional on having taken the fixed cost of the capacity to use a particular tool of risk management, it is less costly to use several means of managing risk in conjunction, rather than exclusively relying on only one way of managing risk. As a specific form of risk management becomes more costly, we expect some substitution away from this way. 23 21 A currency hedge, for instance, only hedges against exchange rate exposure, whereas a large war chest in the form of greater cash holdings will loosen constraints almost irrespective of the source of shocks. 22 Such logic is also consistent with evidence that relatively modest amounts are hedged, as shown in the seminal work of Guay and Kothari (2003). 23 As noted, the exception to this pattern is the positive relation between credit lines, insurance, and derivatives. These ways of managing risk are purchased from financial institutions have a positive relation: if a firm uses derivatives, it is not only more likely to use insurance but also more insurance. A possible reason is that stronger financial sophistication may affect all these costs.

F I G U R E 1 Distributions of ex post profits under different hedging schemes
Note: This figure presents density plots for ex post profits with different hedging practices across 10,000 draws on price and quantity, as described in the text The second-order conditions for utility maximization require that the determinant of the denominator be negative, and hence the sign of the derivative will be determined by the sign of the determinant in the numerator. Thus, As h q > 0 (by assumption) and > 0, the two hedging instruments are substitutes if the cross-differential of variance with respect to hedging instruments is negative. Intuitively, this holds if the more that a firm hedges price risk, the less does additional hedging of quantity risk decrease variance.
To examine the 0/1 margin of hedging we present a simple numerical illustration. We are not attempting to solve for optimal hedges but simply want to illustrate the effects of different hedging choices on the mean and variance of posthedging profits. Consider a firm with operating profits given by Π = (p − mc)q − F, where F = 5000, mc = 20, and p ∼ N(120, 20) and q ∼ N(100, 20), respectively. Both price and quantity are thus taken to be exogenous and random. Using the utility function specified in Equation (2) We illustrate the probability density function of profits after hedging in Figure 1. The solid blue line represents the case where there is no hedging, the red dashed line represents the case when there is full hedging of price (such that p = 120), and the green dash-dot line is the case where there is some hedging of both quantity and price.
Another examination of the same distributions is presented in

CONCLUSIONS
Starting from a broad list of the different ways through which firms can manage risks, this article used textual analysis of US annual reports in the 10-K format to identify the prevalence of these different ways of managing risk. Manual reading of a sample of 10-Ks and annotating sentences yielded a data set that was used to train a classifier, which was then used to classify all the sentences appearing in Item 1 of the 10-Ks for fiscal years 1996-2015. The main tools of risk management that we examine (use of derivatives, insurance, credit lines, long-run contracts, and diversification) are positively related, in the sense that use of one tool makes the use of another tool more likely. In contrast, higher cash holdings make it less likely that these methods are used. As to the overall patterns, a negative relation dominates such that more intensive use of one way of managing risk is associated with less intensive use of alternative ways of risk management, apart from insurance, credit lines, and derivatives, which tend to move in tandem. The previous literature has mostly examined one or two ways of managing risk in isolation, and we believe that the broad evidence is useful in, for instance, understanding what previous results imply for the relation between firm size and risk management as well as for understanding the motivations for risk management.
A comparison with dictionary-based word counts indicates substantial promise for supervised learning methods as applied to 10-Ks. We conclude with a discussion of the potential lessons for future use of textual analysis of 10-Ks to examine issues of risk management. We found that even though there is no explicit requirement to discuss risk management in Item 1 in 10-Ks, in several dimensions, firms discuss risk management in a way that is amenable to automated text analysis. Topics such as derivatives, credit lines, insurance, and diversification were not only discussed relatively frequently but also often appeared in similar texts, which made it easier for the algorithm to identify them.
A possible reason for this is that a discussion of specific risk factors, as mandated in Item 1A, quite naturally invites a discussion of the means used to manage the associated risk. Thus, a deliberation on using derivatives follows the discussion on exposure to the euro; discussion on liability insurance follows naturally from an examination of legal risks.
In line with this reasoning, manual reading suggests that much of the risk management discussion that we documented surfaces in the discussion on risk factors with the proviso that it may not be enough. For instance, Winston (2006) state that "although we have entered into an interest rate cap agreement that limits our interest rate exposure to increases in 30-day LIBOR over 9.17% on principal balances up to $150 million outstanding debt under the GE Line, increases in interest rates up to this limit could increase our interest expense and adversely affect our cash flow." This is an example of how, in a discussion that is required to examine well-specified risks, it is quite natural to also consider ways in which the respective risk is managed.
Risk management tools that are less likely to be linked to the type of reasoning "we are faced with this risk, we do 'that' , but 'that' may not be enough" are more difficult to identify and, by impression from our reading, are less likely to be discussed at all. For instance, this concerns some aspects of operational hedging and of flexibility, which may be less clearly linked to a specific risk factor and more broadly linked to resilience and keeping options open. These may also be amenable to textual analysis and, possibly, one way to identify such sentences would be to examine the sources of risk in more detail. One indication suggesting potential in such a strategy is a positive correlation between our dictionary-based measure of flexibility and the text-based measure of product market fluidity as developed by Hoberg et al. (2014)). It may also be useful to focus manual reading on specific industries as the routes via which flexibility and operational hedging function as risk management are likely to depend on the industry under analysis. For airlines, operational hedging may, for instance, be about fuel efficiency (see, e.g., Treanor et al., 2014). We chose to manually annotate two years quite far apart to allow for a long view, but it may well be that future work will aim to focus the manual annotation and analysis on narrower samples. Our first impression was that different sections of the 10-K are quite different in terms of wording and we therefore examined only Item 1. Applying supervised learning to the management discussion and analysis section of 10-Ks should be valuable in future work. The use of automated textual analysis to examine the decision environment of firms is still at a nascent stage, and applying the type of tools used here to other sources such as earnings conference calls or annual reports in other jurisdictions are just two examples of interesting future applications.

ACKNOWLEDGMENTS
Edvin Hertegård provided excellent research assistance. We're grateful to Niklas Amberg, the editor Srini Krishnamurthy and two anonymous referees for valuable comments, and to Jan Wallander and Tom Hedelius Stiftelse for financial support.