The Impact of Risk-Modeling Disclosures on the Market Perception of Banks Estimated Fair Value Gains and Losses for Financial Instruments

Transcription

1 Southern Methodist University SMU Scholar Accounting Research Accounting 2015 The Impact of Risk-Modeling Disclosures on the Market Perception of Banks Estimated Fair Value Gains and Losses for Financial Instruments Gauri Bhat Southern Methodist University, Stephen Ryan New York University, Follow this and additional works at: Part of the Accounting Commons This document is brought to you for free and open access by the Accounting at SMU Scholar. It has been accepted for inclusion in Accounting Research by an authorized administrator of SMU Scholar. For more information, please visit

2 The Impact of Risk Modeling on the Market Perception of Banks Estimated Fair Value Gains and Losses for Financial Instruments Gauri Bhat Cox School of Business Southern Methodist University 6212 Bishop Blvd. Dallas, TX (214) Stephen G. Ryan Leonard N. Stern School of Business New York University 44 West 4th Street, Suite New York, NY (212) April 2015 This paper was presented at the Accounting, Organizations & Society Conference on Accounting Estimates on October 23-25, 2014 sponsored by the Deloitte Foundation. We are grateful for helpful comments from the editor, Hun Tong Tan, two anonymous reviewers, and conference participants, as well as accounting workshop participants at Rice University.

3 The Impact of Risk Modeling on the Market Perception of Banks Estimated Fair Value Gains and Losses for Financial Instruments ABSTRACT We examine whether and how measures of market and credit risk modeling identified from banks financial reports enhance the returns-relevance of their estimated annual unrealized fair value gains and losses for financial instruments. To capture differences in market liquidity and fair valuation difficulties across types of financial instruments, we distinguish unrealized gains and losses that are recorded in net income versus recorded in other comprehensive income versus calculable using financial statement note disclosures. We predict and generally find that banks market (credit) risk modeling enhances the returns-relevance of their unrealized fair value gains and losses, more so for less liquid instruments subject to greater market-risk-related (credit-riskrelated) valuation difficulties and during periods for which market (credit) risk is higher. We obtain these findings both for banks unadjusted risk modeling measures and for the portions of these measures that we model as attributable to banks risk modeling activities, but not for the portions we model as attributable to banks disclosure of these activities. Keywords: Risk modeling, market risk, credit risk, chief risk officer, fair value, estimates disclosure. JEL classification: G14, G21, G32, M41

4 I. Introduction We examine whether and how banks risk modeling enhances the returns-relevance of their estimated annual unrealized fair value gains and losses (FVGL) on financial instruments. FVGL are changes in fair value during periods that are not yet realized through cash received or paid. When the markets for banks financial instruments are sufficiently illiquid that observable market inputs do not suffice to determine the fair values of those instruments, banks must estimate FVGL by developing valuation models and identifying the inputs necessary to implement those models. These activities require risk modeling both to predict uncertain future cash flows and to determine appropriate rates to discount those cash flows. To conduct risk modeling effectively, banks must invest in adequate personnel and information systems and apply managerial judgment appropriately and with discipline, with inadequate investment (selfinterested application of judgment) introducing unintentional (intentional) noise and bias in FVGL. Banks investment in risk modeling and other risk management activities that discipline fair value estimation appears to vary considerably across banks and time. 1 Banks and their financial instruments exhibit two primary types of risk, market risk and credit risk. Banks engage in two corresponding types of risk modeling activities, market risk modeling (MRM) and credit risk modeling (CRM). Market risk is variability in the value of a position attributable to changes in market prices. Interest rate risk is the primary market risk for most banks. This risk manifests through: (1) discounting effects, which are larger for longer duration positions; (2) prepayment of fixed-rate mortgage-related assets (both securities and loans); and (3) the exercise of other interest rate options, which may be standalone derivatives or 1 For example, Mikes (2011) discusses detailed case studies of two banks and extensive interviews at five additional banks indicating that the quality of risk management varies considerably across banks and time. Only 28% (2%) of our sample banks disclose in their Form 10-K filings that they employed a chief risk officer in 2013 (2002).

5 embedded in traditional financial instruments. MRM involves analyzing the durations of banks financial instruments and the resulting sensitivity of their net interest income and value of equity to interest rate movements. It also involves simulating the effects of interest rate movements on the prepayment of fixed-rate mortgages and exercise of other interest-rate options. Credit risk is variability of the value of funded assets attributable to uncertainty about default probabilities, losses given default, and timing of default, as well as variability of the value of unfunded loan commitments due to uncertainty about draws on those commitments, which are more likely to occur during worse economic times. CRM involves analysis of these parameters based on attributes of the borrowers, borrowing contracts, borrowers performance to date on the contracts, and relevant economic conditions. We argue that banks MRM and CRM activities enhance the quality of their estimates of FVGLs on financial instruments when two conditions hold: (1) the relevant markets for those instruments are sufficiently illiquid that prices or other information from these markets do not substantially determine the instruments fair values; and (2) the instruments exhibit features, such as embedded options or complex structuring, that increase the difficulty of estimating the instruments fair values. As a first cut to capture the applicability of these conditions, we distinguish FVGL that are recorded in net income versus recorded in other comprehensive income versus calculable using financial statement note disclosures ( disclosed ). Figure 1 summarizes current relevant U.S. generally accepted accounting principles, under which FVGL are recorded in net income for most trading and risk management instruments and in other comprehensive income for available-for-sale securities and cash-flow-hedge derivatives. FVGL are disclosed for most of banks other primary types of financial instruments, including their largest asset, loans, and largest liability, deposits. We propose three main hypotheses below that 2

6 we test by examining whether and how MRM and CRM enhance the returns-relevance of these three types of FVGL from Our first and most general hypothesis is that banks MRM and CRM enhance the returnsrelevance of their FVGL, more so for FVGL on less liquid and more difficult-to-fair-value financial instruments. In testing this hypothesis, we exploit the fact that banks financial instruments for which FVGL are disclosed, such as loans and deposits, usually are less liquid and more difficult to fair value than their other instruments. Our second hypothesis is that banks MRM also enhances the returns-relevance of their FVGL recorded in other comprehensive income. Available-for-sale securities and cash-flow-hedge derivatives typically are near credit riskless. Moreover, to the limited extent that banks experience credit losses on these instruments, banks typically record these losses in net income under impairment accounting rules. Hence, interest rate risk is the primary risk reflected in FVGL recorded in other comprehensive income. We expect this hypothesis to hold only for available-for-sale securities and cash-flow-hedge derivatives that are both less than highly liquid and exhibit fair valuation difficulties, such as mortgage-backed and asset-backed securities, so that MRM is essential to estimate the fair values of the instruments accurately. Our third hypothesis is that banks CRM primarily impacts the returns-relevance of their disclosed FVGL, because banks assume credit risk primarily through their funded loans and unfunded loan commitments. To test these hypotheses, we identify banks risk modeling activities from disclosures in their Form 10-K filings. As described in Appendix A, we hand collect disclosures of five MRM activities (interest rate gap analysis, interest rate sensitivity analysis, Value-at-Risk analysis, stress testing, and backtesting) and four CRM activities (statistical credit risk measurement, credit scoring, internal credit risk rating, and stress testing). We equally weight these activities to 3

7 construct indices of banks MRM and CRM. This approach raises the issue that many bank-year financial reports include little about risk modeling activities, particularly CRM early in our sample period. Since all banks must engage in at least minimal levels of MRM and CRM to make investment and financing decisions and to estimate the fair values of financial instruments for which market data do not suffice for the task, it appears that some banks do not disclose these activities. Hence, non-disclosure of a risk modeling activity does not imply absence of the activity. We assume, however, that our MRM and CRM measures capture meaningful variation in risk modeling intensity across banks and time. We test all hypotheses using both one-stage and two-stage approaches. The one-stage approach regresses returns for the twelve months ending four months after the fiscal year end on net income before FVGL recorded in net income 2 and the three types of FVGL (recorded in net income, recorded in other comprehensive income, and disclosed), separately and interacted with the unadjusted MRM and CRM measures, as well as control variables. We frame and test our hypotheses as restrictions on the one or more coefficients on the interactions of banks unadjusted MRM and CRM measures with specific types of FVGL. Empirical results using this approach support our main hypotheses with one explainable exception. We use the two-stage approach to help ensure that the one-stage approach results are attributable to banks risk modeling activities rather than to their choice to disclose these activities. In this approach, we first regress banks unadjusted MRM and CRM measures on 2 Net income before FVGL recorded in net income includes realized gains and losses that are distinct from FVGL except for two types of impairment write-down that are included in net income and thus are accounted for in the same fashion as realized losses. First, net income includes all or the credit loss portion of other-than-temporary impairment write-downs of available-for-sale and held-to-maturity securities. Second, net income includes losses on loans held for sale recognized at fair value under the lower-of-cost-or-fair-value measurement basis. 4

8 proxies for their discipline over risk modeling, technical sophistication, risk exposures, and risk tolerance, which we expect primarily indicate banks risk modeling activities rather than their disclosure of those activities. We use the explained (unexplained) portions of banks unadjusted MRM and CRM measures from these first-stage models as measures of banks risk modeling activities (disclosure of these activities) in second-stage returns models. The estimated coefficients on the MRM and CRM activity measures in the two-stage approach yield the same inferences as the estimated coefficients on the unadjusted measures in the one-stage approach, whereas the estimated coefficients on the MRM and CRM disclosure measures generally are insignificant. These results are consistent with the one-stage approach results being driven by banks risk modeling rather than their disclosure of that modeling. We further hypothesize that MRM more strongly impacts the returns-relevance of FVGL that are recorded in other comprehensive income or disclosed in years with high interest rate volatility, and that CRM more strongly impacts the returns-relevance of disclosed FVGL during the financial crisis. To test these predictions, we interact the primary test variables with indicator variables for years with above-median interest rate volatility or the crisis period Empirical results for the unadjusted MRM and CRM measures and the MRM and CRM activity measures generally support these further hypotheses. Our study contributes to the extensive literature beginning with Barth (1994) that empirically examines the extent and determinants of the value-relevance of fair values and the returns-relevance of FVGL for financial instruments. Our study is most related to recent papers examining disclosures of fair valuation inputs and other measures of the reliability of recognized fair value estimates under Statement of Financial Accounting Standards (FAS) 157 (2006, Accounting Standards Codification (ASC) 820), which became effective in In particular, 5

9 Chung et al. (2014) develop measures of the reliability of recognized fair value estimates from textual analysis of banks and insurers financial statement notes. Chung et al. provide evidence that these measures are associated with enhanced market pricing and lower information risk for recognized level 3 fair values. Our study complements Chung et al. by examining an alternative measure of the reliability of fair value estimates, a longer and more diverse sample period, disclosed as well as recognized FVGL, and returns-relevance rather than value-relevance. II. Financial Report Information about Estimated Fair Values of Financial Instruments, Literature Review, and Hypothesis Development In this section, we first describe the various bases in accounting standards, accounting research, and other sources for our main hypotheses about the effects of MRM and CRM on the returns-relevance of FVGL. We then support and state our further hypotheses about the distinct conditions under which we expect MRM and CRM to be particularly useful. Required Information about Estimated Fair Values and FVGL Under current U.S. generally accepted accounting principles, firms recognize certain types of financial instruments at fair value on the balance sheet. They record FVGL in net income for some of these instruments and in other comprehensive income otherwise. Banks primary types of financial instruments currently recognized at fair value with unrealized gains and losses recorded in net income are: trading securities under FAS 115 (1993, ASC 320); nonaccounting-hedge and fair value hedge derivatives as well as fair value hedged items under FAS 133 (1998, ASC 815); most other trading instruments under industry accounting principles or practices; and financial instruments for which the fair value option is selected under FAS 155 (2006, ASC ) and FAS 159 (2007, ASC ). The primary types of financial 6

10 instruments that currently are recognized at fair value with unrealized gains and losses recorded in other comprehensive income are available-for-sale securities under FAS 115 and cash-flowhedge derivatives under FAS We denote FVGL that are recorded in net income (other comprehensive income) by NIGL (OCIGL). FAS 107 (1991, ASC ) requires firms to disclose the fair and carrying values of most types of financial instruments in the notes to financial statements. FAS 107 disclosures are the only information in banks financial reports about the fair values of their on-balance sheet loans, deposits (excluding core deposit intangibles), and debt as well as off-balance sheet instruments such as loan commitments; FAS 115 requires these and other disclosures for held-tomaturity securities. We denote FVGL that can only be calculated from disclosures by DISCGL. FAS 107 also requires firms to disclose the method(s) and significant assumptions used to estimate the fair value of financial instruments. Banks fulfilled this requirement in boilerplate and minimally informative fashions until the issuance of FAS 157. FAS 157 defines fair value as the price that would be received to sell an asset or paid to transfer a liability in an orderly transaction between market participants at the measurement date, and it creates a three-level hierarchy of fair value inputs. 4 Level 1 (highest quality) inputs are quoted prices in active markets for the identical item. Level 2 inputs are quoted prices in 3 Other standards require certain types of financial instruments to be recognized at fair value at either: (1) inception, e.g., FAS 166 (2009, ASC 860) for retained interests in securitizations; or (2) the time of (other than temporary) impairment write-downs, e.g., FAS 65 (1982, ASC ) and SOP 01-6 (2001, ASC ) for held-forsale loans and FAS 115, as amended by FSP FAS & FAS (2009, ASC ), for available-for-sale and held-to-maturity securities. 4 Paragraphs C21-C22 of FAS 157 indicates that generally accepted accounting principles include various practicability exceptions to fair value measurement that FAS 157 did not change. While banks do not often invoke these exceptions in their financial reports, the exceptions may have some effect on our FVGL measures or be correlated with our risk modeling measures; e.g., a bank with better risk modeling may have less need to invoke a practicability exception. 7

11 markets that are not active for the identical item or in active markets for similar items and most other observable information. Level 3 (lowest quality) inputs are unobservable reporting-firmsupplied inputs. The level of a fair value estimate is determined by the level of its lowest quality significant input. FAS 157, as amended by FASB Staff Position (FSP) FAS (2009, ASC ) and Accounting Standards Update (ASU) (2010, ASC ), requires quarterly disclosures of the amounts of each major category of assets and liabilities that is recognized at fair value on the balance sheet in each of the three levels, separately for items fair valued each reporting period versus on a non-recurring basis. For level 2 and level 3 fair values, firms must disclose their valuation techniques and inputs. For level 3 fair values, firms must disclose rollforwards of the fair values from the beginning to the end of the period, distinguishing total gains and losses, purchases, sales, issuances, settlements, and transfers in and out of level 3. 5 Prior Empirical Research on the Differential Value-Relevance of Fair Values and Returns- Relevance of FVGL by Type of Financial Instrument and by Recognition versus Disclosure An extensive literature beginning with Barth (1994) empirically examines the valuerelevance of estimated fair values and returns-relevance of FVGL on financial instruments. Most of these studies limit their samples to banks because financial instruments dominate banks balance sheets, banks financial instruments exhibit varying levels of liquidity and other factors associated with the need or ability to exercise discretion over fair value estimation, and analysis 5 While a massive improvement over FAS 107 disclosures, FAS 157 disclosures exhibit three limitations that make them unsuitable for use in our study. First, these disclosures are available for the bulk of our sample beginning in 2008, less than half our sample period and a period dominated by the financial crisis and its aftermath. Hence, controlling for these disclosures would essentially eliminate our ability to compare the crisis and non-crisis periods. Second, the most important FAS 157 disclosures pertain to level 3 recognized fair values for financial instruments, which constitute only about 2% of banks recognized financial instruments. Third, these disclosures do not encompass disclosed fair values for the bulk of banks financial instruments. 8

12 of a single industry mitigates concerns about unmodeled heterogeneity. Given that several recent and extensive surveys of this literature exist (e.g., Ryan 2011, Section 4.5), we briefly summarize the four main findings in these studies that bear on the hypotheses developed below. First, for investment securities and other similarly liquid assets, fair values are valuerelevant and FVGL are returns-relevant, although the returns-relevance results are more sensitive to model specification and sample (Barth 1994; Ahmed and Takeda 1995; Carroll et al. 2003; and Danbolt and Rees 2008). Second, inconsistent (essentially no) evidence has been generated as to whether the fair values of loans and derivatives (other financial instruments) are valuerelevant (Barth et al. 1996; Eccher et al. 1996; Nelson 1996; and Venkatachalam 1996). Third, banks disclosed fair values of financial instruments appear to be both noisy (i.e., unexplainable based on contemporaneous economic events) and managed to make less solvent banks appear more so. Surprisingly, the value-relevance of disclosed fair values does not appear to be affected by noise, but it is reduced by discretionary behavior (Beaver and Venkatachalam 2003 and Nissim 2003). Fourth, fair values that are recognized (i.e., more prominent) are more valuerelevant than those that are disclosed (Hirst et al. 2004; Ahmed et al. 2006; Chambers et al. 2007; Badertscher et al. 2014; and Dong et al. 2014). Collectively, these findings suggest that estimated fair values and FVGL for which the signal-to-noise ratio is (perceived to be) higher have higher value-relevance and returns-relevance, respectively. We examine returns-relevance rather than value-relevance because the latter is both more extensively studied in the prior literature and more subject to omitted variables (but less subject to measurement error). Barth, Beaver, and Landsman (2001) and Holthausen and Watts (2001) discuss these methodological trade-offs. 9

13 Prior Empirical Research on the Differential Value-Relevance of Fair Values of Financial Instruments by FAS 157 Level and Other Reliability Disclosures Using FAS 157-required disclosures, several studies provide evidence that banks recognized fair values of financial instruments estimated using higher quality inputs are more value-relevant (Kolev 2009; Song et al. 2010; Goh et al. 2015). These studies all find that level 1 and 2 fair values are more value-relevant than level 3 fair values. Goh et al. find that level 1 fair values are more value-relevant than level 2 fair values. Song et al. find that level 3 fair values are more value-relevant for banks with better corporate governance. Chung et al. (2014) develop binary (disclosure made versus not) and continuous (number of words in the disclosure divided by total number of words in the Form 10-K filing) measures of FAS 157-encouraged disclosures in banks and insurers financial statement notes about the controls, processes, and procedures used to assure the reliability of their recognized fair value estimates. Chung et al. (2014) provide evidence that these measures are associated with enhanced market pricing (higher share price and lower priced risk) and lower information risk (higher analyst consensus) for recognized fair values estimated using significant level 3 inputs. Collectively, these findings suggest that disclosures indicating that banks more reliably estimate the fair value of their financial instruments are associated with enhanced valuerelevance of these estimates. Hypotheses We discuss banks MRM and CRM activities in the introduction and repeat this discussion here only insofar as it pertains directly to specific hypotheses. All of our hypotheses reflect the view that risk modeling typically improves fair value estimation. We acknowledge that risk modeling may instead deteriorate fair value estimation if it crowds out the appropriate 10

14 use of judgment or devolves into a compliance exercise, as Mikes (2011) and Kaplan and Mikes (2012) discuss occurred at specific banks. Our first and most general hypothesis is that banks MRM and CRM enhance the returnsrelevance of their FVGL, particularly for less liquid financial instruments that exhibit lower quality market information (i.e., less price transparency ) and thus greater need for risk modeling to estimate FVGL reliably. This hypothesis is supported by four prior theoretical and empirical literatures: (1) the theoretical literature indicating that market prices respond more strongly to more precise information (e.g., Holthausen and Verrecchia 1988); (2) the empirical literature on the value-relevance of fair values and returns-relevance of FVGL by type of financial instruments and by recognition versus disclosure discussed above; (3) the empirical literature that demonstrates that FAS 157 fair value input level and voluntary fair value reliability disclosures affect this value-relevance discussed above; and (4) Bhat et al. s (2014) findings that banks CRM is positively associated with the timeliness of their loan loss provisions. We formally state this hypothesis and all subsequent hypotheses in alternative form. H1: MRM and CRM enhance the sensitivity of returns to FVGL, more so for FVGL on less liquid and more difficult-to-fair-value financial instruments. In testing Hypothesis H1, we exploit the fact that banks financial instruments generating DISCGL, such as loans and deposits, usually are less liquid and more difficult to fair value than their other financial instruments. As discussed below, the financial instruments that generate OCIGL are a mixed bag of highly liquid and thus easy-to-fair-value instruments and less liquid and more difficult-to-fair-value instruments. The instruments that generate NIGL usually are highly liquid and easy to fair value. Hence, we expect Hypothesis H1 to hold most strongly for DISCGL, less strongly for OCIGL, and not at all for NIGL. 11

15 Our second hypothesis is that banks MRM enhances the returns-relevance of their OCIGL, particularly for mortgage-backed and asset-backed securities. This hypothesis has three interrelated bases. First, interest rate risk is the primary risk reflected in OCIGL, which are generated by available-for-sale securities and cash-flow-hedge derivatives. Banks most common types of available-for-sale securities are U.S. Treasuries, other U.S. federal government securities, and agency-guaranteed mortgage-backed securities. Their most common type of cashflow-hedge derivatives are interest rate derivatives with highly creditworthy counterparties. These instruments typically are near credit riskless. Moreover, to the limited extent that banks experience credit losses on these instruments, they typically record these losses in net income under impairment accounting rules. Second, interest rate risk manifests through: (1) discounting, which more strongly affects longer duration positions; (2) prepayment of fixed-rate mortgages underlying mortgage-backed and other asset-backed securities; and (3) the exercise of other interest rate options, which may be standalone cash-flow-hedge derivatives or embedded in available-for-sale securities (Ryan 2011, Section 2.4). Banks must use considerably more sophisticated MRM (e.g., interest rate simulation) to determine the effects of prepayment and other interest rate options than to determine discounting effects. Relatedly, it is considerably more difficult for banks to hedge options than discounting effects. Third, compared to banks other types of available-for-sale securities, mortgage-backed securities and other types of asset-backed securities typically are both less liquid and subject to greater fair valuation difficulties, due to their embedded prepayment options and complex structuring (e.g., waterfalls). Even agency-guaranteed mortgage-backed securities generally 12

16 exhibit low price transparency unless and until specific securities are identified in to-beannounced trades (Vickrey and Wright 2013). 6 H2: MRM enhances the sensitivity of returns to OCIGL, especially OCIGL on less liquid and more difficult-to-fair-value mortgage-backed and asset-backed available-for-sale securities. Our third hypothesis is that banks CRM primarily impacts the returns-relevance of their DISCGL. This hypothesis is motivated by the fact that banks assume credit risk primarily through their funded loans and unfunded loan commitments for which estimated fair values are disclosed (Ryan 2011, Section 2.5). H3: CRM primarily enhances the sensitivity of returns to DISCGL. Our fourth and fifth hypotheses are motivated by the intuition that banks risk modeling should be more important during periods when the modeled risk is higher. In particular, MRM (CRM) should be more important during years when interest rates are more variable (credit losses are unexpectedly higher). H4: The effect of MRM is stronger during years with above-median interest rate volatility than in other years. H5: The effect of CRM is stronger during the financial crisis than in other years. III. Measurement of Variables and Methodology Fair Value Gains and Losses 6 Vickrey and Wright (2013) provide evidence that to-be-announced agency mortgage-backed securities trades (which are accounted for as derivatives, not investment securities, until settled) exhibit high volume and price transparency. They also explain, however, that these trades exhibit various features (e.g., settlement at a single date each month and the cheapest-to-deliver option) that reduce the price transparency these trades provide for existing pools of agency-guaranteed mortgage-backed securities. 13

17 We measure NIGL as trading revenue reported on line 5.c of Schedule HI and OCIGL as other comprehensive income reported on line 4 of Schedule HC-R of Y-9C filings. Unfortunately, both NIGL and OCIGL include items other than FVGL; NIGL includes fee income and realized gains and losses and OCIGL includes other comprehensive income from pensions and foreign currency. 7 For financial assets (liabilities) for which fair values are disclosed, we calculate DISCGL as (minus) the change during the year of the difference between the fair value and carrying value of the instrument. 8 We obtain disclosed fair and carrying values from SNL Financial. All FVGL components are after taxes calculated using a statutory tax rate of 35% and divided by beginning-of-year market value of equity. Unadjusted Risk Modeling Measures To construct our MRM and CRM measures, we had to decide which risk modeling activities to include in the measures and also the disclosure medium to examine for these activities. To make the first decision, we considered all of the market and credit risk modeling activities (including related risk management processes) mentioned in a 2001 public disclosure survey of large international banks conducted by the Basel Committee on Banking Supervision. 9 To accommodate the U.S. banks in our sample, we excluded risk modeling activities for which disclosures are required for U.S. banks or that are not provided by any sample bank in any 7 In untabulated analysis, we tried to remove fee income from NIGL by subtracting its average over the prior three or five years. This transformation of NIGL should mostly remove fee income, which is much more persistent than fair value gains and losses. No inferences are affected by this alternative specification of NIGL. 8 Because provisions for credit losses on loans and loan commitments are recorded in net income, these recognized accruals are not included in DISCGL. This likely reduces the power of our tests of hypotheses H1, H3, and H5, because prior research shows that CRM is associated with these accruals (Bhat et al. 2014). 9 While these documents also consider operational risk modeling activities, we do not deem this type of activity to be directly related to the estimation of fair values and FVGL for financial instruments. We acknowledge that Chernobai et al. (2011) provide evidence that operational risk is positively correlated with credit risk. 14

18 sample year. This yields five market risk modeling activities (interest rate gap analysis, interest rate sensitivity analysis, Value-at-Risk analysis, stress testing, and backtesting) and four credit risk modeling activities (statistical credit risk measurement, credit scoring, internal credit risk rating, and stress testing). We describe these activities in detail in Appendix A. We choose banks annual Form 10-K filings as the sole disclosure medium for three reasons. First, banks provide the most extensive disclosures of their risk modeling activities in these filings, primarily due to SEC requirements. For example, SEC FRR 48 requires that banks make extensive disclosures of their market risk in these filings. Second, the extents of disclosure in Form 10-K filings and other media are positively correlated (Lang and Lundholm 1993). Third, our analysis of samples of banks financial analyst reports and conference call transcripts yielded no mention of risk modeling. We hand collected banks risk modeling disclosures from their Form 10-K filings from We electronically searched those filings for a large number of relevant words or phrases, such as risk management, risk model, scoring, risk rating, etc. We read each identified disclosure to verify that the risk modeling activity was disclosed. We construct an indicator for each activity that takes a value of one if it is disclosed and zero otherwise. Our unadjusted market risk modeling measure, MRM (credit risk modeling measure, CRM), is the simple average of the indicators for the five market risk activities (four credit risk activities). As averages of indicator variables, MRM and CRM take values from zero to one. Several caveats are in order regarding our risk modeling measures. First, we observe banks disclosures of risk modeling activities, not the activities themselves. Presumably the disclosure of risk-modeling activities implies banks engage in the activities, because there does not appear to be any reason for banks to lie in these disclosures, which generally are too terse to 15

19 divulge meaningful proprietary information. 10 However, the non-disclosure of an activity does not imply that the bank does not engage in the activity. As discussed in the introduction and below, we employ a two-stage approach to help ensure that our results are attributable to banks risk modeling activities rather than their disclosure of those activities. Second, our risk modeling measures capture the existence of particular activities more than the quality of those activities. Third, and relatedly, despite robustness checks to incorporate various different weighting approaches, simple averaging of the indicators for risk modeling activities might not reflect the weights investors assign to these activities. Lastly, despite careful reading of each disclosure, our coding of risk modeling activities might be subjective. The large number of disclosures involved makes coding by multiple individuals costly, however. Chief Risk Officer Indicator We use an indicator variable for whether banks employ a chief risk officer, CRO, as our measure of banks discipline over risk modeling. We considered using standard proxies for corporate governance for this purpose, but these proxies do not appear to correspond strongly, if at all, with the estimation of fair values and FVGL. 11 Consistent with this view, Aebi et al. (2012) find that these standard proxies are mostly insignificantly or even negatively related to the banks performance during the crisis, whereas banks that employ chief risk officers that report directly to the board of directors exhibit superior performance during the crisis. Ellul and 10 In addition, disclosures of risk modeling are subject to internal controls, to CEO and CFO certification of financial reports under Section 302 of the Sarbanes-Oxley Act, and to audit (reading for consistency) by banks auditors if the disclosures appear in the financial statement notes (MD&A) sections of financial reports. The observed frequency of disclosures is fairly low, inconsistent with easily imitated cheap talk. 11 More generally, Larcker et al. (2007) argue that construct validity issues with corporate governance measures cause mixed and hard-to-interpret results in many studies, in part due to the absence of theory indicating which corporate governance variables matter in what contexts. 16

20 Yerramilli (2011) find similar results for their risk management index, which depends strongly on the employment and status of the chief risk officer. CRO takes a value of one if the bank discloses that it employs a chief risk officer in its Form 10-K filing and zero otherwise. We searched filings to determine whether the words chief and risk appear within five words of each other and, if so, whether the words risk and either officer or director appear within five words of each other. For filings that met both filters, we read the relevant passages to confirm the existence of a chief risk officer. CRO takes a value of one for these bank-year observations and zero otherwise. 12 First-Stage Models Explaining Risk Modeling In the two-stage approach, the first-stage models regress each of the unadjusted MRM and CRM measures (collectively denoted XRM) on seven proxies for banks discipline over risk modeling, technical sophistication, risk exposures, and risk tolerance, all of which we expect primarily to affect the extent of banks risk modeling activities rather than their disclosure of those activities. These proxies are: CRO, a proxy for banks discipline over risk modeling; log of total assets (SIZE) and a trading portfolio indicator (TRADD), proxies for banks technical sophistication; 0-1 year repricing gap (INTSEN), the proportion of assets that are commercial loans (COMMLOAN), and the standard deviation of fair value income deflated by beginning market value of equity over the past ten years (FVISTD), proxies for banks interest rate risk, credit risk, and overall risk, respectively; and tier 1 risk-based capital ratio (TIER 1), a proxy for 12 Using data available on banks top five officers from SNL Financial, we endeavored to determine whether CROs were one of these officers, but this determination cannot be made from these data. 17

21 banks solvency or risk tolerance. We also include year fixed effects to capture changes in the risk modeling measures, particularly CRM, over time. This model is: XRM CRO SIZE TRADD INTSEN COMMLOAN 1 FVISTD TIER1 year fixed effects (1) We suppress time subscripts except when necessary for clarity. We use the portions of the risk modeling measures explained by the seven proxies, denoted with hats, i.e., XRM, as measures of risk modeling activities. We use the remaining portions of the risk modeling measures, i.e., XRM - XRM, as measures of risk modeling disclosures. 13 Models of the Impact of Risk Modeling on the Returns-Relevance of FVGL and Hypotheses Restated as Coefficient Restrictions The returns models using the unadjusted risk modeling measures (i.e., in the one-stage approach) regress share returns for the 12 months ending four months after the fiscal year end (R) 14 on net income before FVGL recorded in net income (NIBNIGL), the three types of FVGL (NIGL, OCIGL, and DISCGL), the risk modeling measure under consideration (either MRM or CRM), interactions between that risk modeling measure and each of NIBNIGL and the three types of FVGL, and year fixed effects. 15 These models are: 13 In untabulated robustness analyses, we use an alternative first stage approach that does not alter any inference in the second stage. Specifically, we include firm attributes, that the disclosure literature surveyed by Healy and Palepu (2001) finds to be associated with disclosure, as additional explanatory variables in equation (1): an indicator for issuance of debt or equity, i.e., capital market transactions that motivate disclosure; an indicator for negative net income, a proxy for performance; the book-to-market ratio, a proxy for growth; number of analysts forecasting earnings and percentage institutional ownership, proxies for information environment. We use the portions of the risk modeling measures explained by these proxies as measures of risk modeling disclosures. 14 We use this return window because some sample banks are non-accelerated filers that file their Form 10-K filings on the last day of the third month following the end of the fiscal year. 15 In equations (2A) and (2B), we interact NIBNIGL with the unadjusted risk modeling measures because certain components of NIBNIGL should be affected by risk modeling, e.g., provisions for loan losses should be affected by 18

22 R a b NIBNIGL b NIGL b OCIGL b DISCGL b ( NIBNIGL* XRM) b ( NIGL* XRM) b ( OCIGL* XRM) b ( DISCGL* XRM) b XRM 6 1 year fixed effects e (2A) The returns models using the risk modeling activity and disclosure measures from the first-stage model equation (1) (i.e., in the two-stage approach) follow directly from substituting XRM and XRM-XRM for XRM in equation (2A). These models are: R a b NIBNIGL b NIGL b OCIGL b DISCGL b ( NIBNIGL* XRM ˆ ) 6 b b b 1 b ( NIGL* XRM ˆ ) b ( NIBNIGL*{ XRM XRM ˆ }) b ( OCIGL*{ XRM XRM ˆ }) b 7 ( OCIGL* XRM ˆ ) b ( DISCGL* XRM ˆ ) b { XRM XRM ˆ } year fixed effects e ( NIGL*{ XRM XRM ˆ }) ( DISCGL*{ XRM XRM ˆ }) 5 9 XRM ˆ (2B) We restate Hypotheses H1-H5 as coefficient restrictions in equations (2A) and (2B). Hypothesis H1 predicts that b8 is positive and larger than b6 and b7 for both MRM and CRM Hypothesis H2 predicts that b7 is positive for MRM, especially when OCIGL are for mortgagebacked and asset-backed securities. Hypothesis H3 predicts that the primary effect of CRM is manifested in b8. Hypothesis H4 predicts that Hypotheses H1 and H2 pertaining to for MRM hold more strongly in years with above-median interest rate volatility. Hypothesis H5 predicts that Hypotheses H1 and H3 pertaining to for CRM hold more strongly during the financial crisis. IV. Sample Selection and Data Sample Selection CRM. We expect FVGL to be more strongly affected than NIBNIGL by risk modeling, however; to this extent, the interaction of NIBIGL with the unadjusted risk modeling measures can be viewed as a placebo test. 19

23 We study banks for the same reasons as the prior literature discussed in Section II. In addition, the cost of hand collecting our risk modeling and CRO variables limits the sample size. As summarized in Table 1, Panel A, the sample comprises all public U.S. bank holding companies with data available from from the following five sources: most financial data from Y-9C regulatory filings on the Federal Reserve Bank of Chicago website, disclosed fair values and carrying values of financial instruments from SNL Financial, risk modeling disclosures from Form 10-K filings on 10-K Wizard, and stock information from CRSP. 16 The total assets threshold at which bank holding companies are required to file Y-9Cs increased from $150 million to $500 million in 2006; we impose the latter size restriction for all bank-year observations to yield a more comparable sample through time. Some bank holding companies that file Y-9Cs differ slightly from the corresponding public companies that file Form 10-Ks. Table 1, Panel A reports the construction of the final sample of 238 unique banks and 2,413 bank-year observations for the years To mitigate the effects of outliers, we winsorize all continuous model variables at the 0.5% and 99.5% levels of their distributions. Descriptive Statistics Table 1, Panel B presents descriptive statistics for the main variables. As is typical in banking studies, the size-related variables exhibit considerable variation and right skewness. Despite the financial crisis, the vast majority of bank-year observations are profitable and well- 16 We start the sample in 2002 because FAS 133 became effective in 2001 and two consecutive years of disclosed fair and carrying values are necessary to calculate DISCGL. FAS 133 requires all derivatives to be recognized at fair value on the balance sheet, with unrealized gains and losses recorded in net income for non-accounting hedge and fair-value-hedge derivatives and in other comprehensive income for cash-flow-hedge derivatives. Compared to prior accounting principles and accounting practices, FAS 133 significantly increased the amount of FVGL recorded in net income and decreased the amount of FVGL disclosed; the standard also affected the FVGL recorded in other comprehensive income in a less obviously directional fashion. 20

24 capitalized, with the 25 th percentile of NIBNIGL equaling 4.9% of beginning-of-year market value of equity and of the Tier 1 risk-based capital ratio equaling 10.3%. The distribution of R is quite spread, however, reflecting that fact that banks experienced good stock performance both early and late in the sample period and poor performance during the crisis. All three types of FVGL have means and medians of approximately zero. The spread of DISCGL is much higher than the spread of NIGL and, to a lesser extent, OCIGL, however, because fair values for banks primary financial asset (loans) and financial liability (deposits) are disclosed and only 21.8% of banks (mean TRADD) hold any trading positions. The median bank-year discloses two out of five MRM activities and one out of four CRM activities. 13.6% of bank-years employ a chief risk officer. Table 1, Panel C reports the number of observations, the means of MRM, CRM, and CRO, and the frequency with which each of the indicators for the risk modeling activities underlying MRM and CRM take a value of one in each year from 2002 to At least 170 banks appear in the sample each year. The number of banks is highest in the middle sample years, peaking at 221 in The lesser number of banks in the early (later) years is attributable to our requirement that observations have at least $500 million total assets (bank failures and mergers and acquisitions resulting from the crisis and passage of time). The frequency of banks disclosures of MRM activities is flat over the sample period due to the 1998 effective date of the main market risk disclosure requirement, FRR 48. In contrast, the frequency of banks disclosures of CRM activities and CRO increases strongly over the sample years. Table 2 reports the Pearson (Spearman) correlations for the main variables in the regression; we discuss only the Pearson correlations. Three insights emerge from this table. First, MRM and CRM are both significantly positively associated with most of the variables capturing 21

25 size and sophistication, such as CRO, SIZE, NIGL, and TRADD. Second, there are interpretable differences between the correlations involving MRM and CRM. For example, NIGL is more positively correlated with MRM than with CRM, consistent with trading-oriented banks primarily trading financial instruments with more market risk than credit risk. In contrast, TRADD is more positively correlated with CRM than MRM, consistent with large and sophisticated banks being relatively more likely to conduct CRM. Consistent with banks chief risk officers primarily being tasked with evaluating credit risk, CRO is more positively correlated with CRM than MRM. Third, unlike NIGL, neither OCIGL nor DISCGL is correlated with any of the risk modeling measures or measures of bank size or sophistication. V. Empirical Analysis First-Stage Regression Results Table 3 reports the estimation of equation (1), the first-stage models used to explain the unadjusted risk modeling measures in terms of risk-related explanatory variables and fixed year effects (untabulated). The first (second) column presents the results for the unadjusted market (credit) risk modeling measure MRM (CRM). The fit is considerably better for the CRM model, primarily because the year fixed effects are more significant in this model due to CRM s strong upward trend over time reported in Table 1, Panel C. The coefficient on CRO is significantly positive at the 5% level in the MRM model and at the 10% level in the CRM model. The coefficient on SIZE is significantly positive at the 1% level in both models. These results indicate that banks discipline over fair value estimation and technical sophistication explain their risk modeling. In the MRM model, the coefficient on INTSEN is significantly negative at the 5% level and the coefficient on FVISTD is significantly 22

26 negative at the 10% level, suggesting that banks with higher interest rate risk tend to seek out or tolerate rather than manage that risk. In contrast, the coefficient on FVISTD is positive at the 10% level in the CRM model. No other explanatory variable is significant in any model. Primary Regression Results Table 4, Panel A reports the estimations of equations (2A) and (2B), which we use to test hypotheses H1-H5 about whether and how banks risk modeling impacts the returns-relevance of their three FVGL components. To link to prior research and provide a benchmark for the main results, column (1) of the panel reports the results for the model including only NIBNIGL, the three FVGL components, and the fixed year effects (untabulated). Column (2) [(4)] presents the results for equation (2A) using the unadjusted market [credit] risk modeling measure MRM [CRM]. Column (3) [(5)] presents the results for equation (2B) using the market [credit] risk modeling activity and disclosure measures MRM and MRM-MRM [CRM and CRM-CRM ]. In the benchmark model reported in column (1), the model fit is good with an R 2 of 41.8%. Such a high R 2 for a returns model obtains in large part from the substantial variation in bank industry performance across the sample years discussed earlier, which is captured by the fixed year effects. The coefficient on NIBNIGL is significantly positive at the 1% level but well below one at 0.249, suggesting the presence of considerable noise in this variable. The coefficient on NIGL is significantly positive at the 5% level and well above one at 3.297, indicating that it includes a permanent component. This reflects the inclusion of some fee revenue in trading revenue on the Y-9C filings and the persistence of FVGL from most dealer and proprietary trading activities due to the incorporation of trading spreads (i.e., day-one profits). The coefficient on OCIGL is significantly positive at the 5% level and somewhat below one at 0.742, consistent with it primarily capturing transitory income items. The coefficient on 23