sr663

Transcript

1 Federal Reserve Bank of New York Staff Reports Assessing Financial Stability: The Capital and Loss Assessment under Stress Scenarios (CLASS) Model Beverly Hirtle Anna Kovner James Vickery Meru Bhanot Staff Report No. 663 February 2014 July Revised 2015 This paper presents preliminary findings and is being distributed to economists and other interested readers solely to stimulate discussion and elicit comments. paper are those of the author not necessarily views expressed in this The do and s position of the the reflect Federal Reserve Bank of New York or the Federal . s Reserve System. Any errors or omissions are t he responsibility of the author

2 Assessi ng Financi ility : The Capita l and L oss Assessmen t un der Stres s Scenarios al Stab ) Mode l (CLASS Hirtle, Ann , Jam es V ickery, and Meru Bhanot a Kovner Beverl y of New York Staff Reports , no. 663 Federal Reserve Bank July 2015 February 2014; revised on: G21, G17, G01 JEL classificati Abstract The CLASS model is a top-down capital stress testing framework that uses public data, simple econometric models, and auxiliary assumptions to project the effect of macroeconomic scenarios on U.S. banking firms. Through the lens of the model, we find that the total banking system capital shortfall under stressful macroeconomic conditions began to rise four years before the financial crisis, peaking in the fourth quarter of 2008. The capital gap has since fallen sharply, and is now significantly below pre-crisis levels. In the cross section, banking firms estimated to be most sensitive to macroeconomic conditions also have higher capital ratios, consistent with a “precautionary” view of bank capital, though this behavior is evident only since the crisis. We interpret our results as evidence that the resiliency of the U.S. banking system has improved since the financial crisis, and also as an illustration of the value of stress testing as a macroprudential policy tool. Key words: capital, bank, stress testing, financial stability _________________ Hirtl Vickery: Federal Reserve Bank of e, Kovner, York (e-mai l: bev [email protected], New [email protected], [email protected]). Bhanot: University of Chicago (e-mail: [email protected]). The authors thank Dafna Avraham, Peter Matthew Mazewski, Lev Hull, Lily Ulysses Velasquez for outstanding research assistance, and Menand, Zhou, and particularly for their many Federal Reserve colleagues ideas. The authors also thank suggestions and discussants Ignazio Angeloni and Mark Jensen, as well as participants at the Yale Program on Financial Stability Annual Conference, the Interagency Risk Quantification Forum, and a research seminar to thank Viral Acharya and Robert at the FDIC. Finally, the authors want Engle for providing data and assistance regarding their SRISK measure of capital shortfall. The views expressed in this paper are those of the and do not necessarily reflect the position authors of Reserve System. Reserve Bank of New York or the Federal the Federal

3 1. Introduction Central and bank supervisors have increasingly relied on capital stress testing as a supervisory banks macroprudential and The recent financial crisis highlighted the importance of the amount and tool. and capital in of public confidence in individual financial institutions bank in the quality ensuring system as a whole. Stress test financial have been used by central banks and supervisors to assess s resilience the of individual banking companies to adverse macroeconomic and financial market as a way of gauging additional capital needs at individual firms and as means of assessing conditions the overall capital adequacy of the banking system. In United States, the first formal ban the k stress tests – the Supervisory Capital Assessment Program (SCAP) supervisory – were performed during 2009, and stress tests have since been made permanent through the implementation of the provisions of the Dodd ‐ Frank Act (Dodd ‐ Frank Act Stress Tests, or DFAST) and the stress test 1 the Comprehensive Capital Analysis and of Review (CCAR) introduction European banking . conducted stress tests of the largest European banking companies in 2009, 2010, 2011 supervisors 2 2014, with an additional round of tests planned for 2016. and banks A of central number have also system ‐ wide stress test frameworks to assess constructed robustness of their banking systems the 3 adverse macroeconomic environments and stressed funding conditions. to In this paper, we describe a framework for assessing the impact of macroeconomic conditions on banking the U.S. system – the Capital and Loss Assessment under Stress Scenarios (CLASS) model. The model is a “top ‐ down” model of the U.S. commercial banking industry that generates CLASS and of bank commercial ba projections holding company (BHC) income and capital under nk regression scenarios. These projections are based on models of components of macroeconomic income, expense and loan performance, combined with assumptions about provisioning, bank other growth and factors. asset dividends, 1 Board of Governors of the Federal Reserve System (2009a, 2009b, 2012, 2013a, 2013b) for more detail on the See Greenlaw and DFAST stress tests. Bookstaber et al. (2013) and CCAR et al. (2012) discuss use of supervisory SCAP, e housing macroprudential purposes. Pre ‐ dating the SCAP, regular supervisory stress tests of th stress tests for conducted government enterprises Fannie Mae and Freddie Mac were by their regulator, the Office of sponsored Housing Enterprise Oversight (OFHEO). Frame, Gerardi and Willen (2015) present a detailed analysis of Federal c in 2008. reasons why they failed to forsee the insolvencies of Fannie Mae and Freddie Ma these tests and the 2 Committee of European Banking Supervisors (2010) and See European Banking Authority (2011) for details and results of the early European stress tests. 3 For instance, Kapadia et al. (2012) describe the RAMSI developed by the Bank of England and Wong and model risk. Hui (2009) describe a model developed at the Hong Kong Monetary Authority to assess liquidity 1

4 from the CLASS model provide insight into the capital resiliency of the U.S. banking Projections against market severely stressed economic and financial conditions and thus into the system CLASS of the broader financial system. Specifically, the projections suggest that the U.S. stability banking industry’s vulnerability to undercapitalization has declined, not only relative to the financial projections 2007 ‐ 09, but also relative to the period preceding the crisis CLASS model of crisis. indicate an increasing capital “gap” (a shortfall of capital under stressed economic conditions) starting as early as 2004, well before most market ‐ based measures of capital adequacy began to deteriorate. on ‐ sectionally, CLASS model projections based cross current industry data suggest that Looking that are projected to experience large declines firms capital under stressful economic conditions in also tend to have higher current capital ratios. This relationship is consistent with a “precautionary” view of bank capital. That is, banking firms holding assets or engaged risky in risky income ‐ producing activities also hold higher capital buffers to limit the likelihood of financial distress. This relationship has evolved over time, however. CLASS model results for years prior to the financial relationship and crisis do not show a consistent cross ‐ sectional ratios between capital projected capital under stress – instead we find evidence of this precautionary behavior only in the declines in further last of our sample (2011 ‐ 13). This finding part supports the idea that the capital strength and stability of the U.S. banking industry have improved relative to both the financial crisis period into period leading and the the crisis. The CLASS model’s top ‐ down approach is intended to complement more detailed supervisory models of components of bank revenues and expenses, such as those used in the DFAST, CCAR, and model European stress tests. Unlike such models, the CLASS relies only on public information, and financial market data combined with bank and BHC regulatory report namely, macroeconomic The use filings. of regulatory report data allows the model to compute projections easily for a much larger number of firms and with greater frequency than is practical from detailed bottom ‐ up analysis supervisory data using from BHCs. collected In addition, directly the CLASS framework is relatively to understand, and can produce income and capital projections in only a couple of simple to minutes for a single macroeconomic scenario. As a result, it can be used either for simulations or 2

5 provide immediate back ‐ of ‐ the envelope estimates of the effect of a particular macroeconomic on the U.S. banking system. shock “top these advantages, Balanced CLASS model’s against ‐ down” approach also has some the limitations. For example, it abstracts from many idiosyncratic differences between significant individual For this reason, while the institutions. model reasonably be used can to model aggregate income and capital, net and the overall distribution of capital across institutions, caution should be exercised in using the model to project the capital of a specific bank or BHC. In addition, the model does currently incorporate any feedback from the not to the macroeconomy or to banking system markets. Instead, the macroeconomic financial used as inputs to the model are treated projections as exogenous. In spite of these limitations, we show that the CLASS model’s projections of revenues, loan losses, significantly and income are positively and statistically net correlated with the Federal Reserve’s DFAST projections, which are based on more detailed models and extensive confidential supervisory data. CLASS model projections for the financial crisis period are also positively that These correlated with actual outcomes for individual BHCs during this period. results suggest that CLASS model is capturing some of the important the ‐ firm and economy specific wide factors ‐ generate differences in bank performance under stress. that rest The of this paper describes the CLASS model in more detail and presents model projections that provide insight into the evolution of the capital strength and financial stability of the U.S. banking system over time. Section 2 provides an overview of the CLASS model’s framework and analytical approach and presents projections of industry aggregate revenue, losses, net income and capital a ratios under range of hypothetical scenarios, based on U.S. banking system data as of 2013:Q3. Section shows how the CLASS model can be used 3 to analyze trends in financial stability. Section 4 detailed contains of the data, specifications a of the CLASS model equations and discussion describes the auxiliary assumptions needed to complete the model. Section 5 reports specification by tests CLASS model results to those generated comparing the Federal Reserve in DFAST 2014 and crisis to BHCs’ actual experiences during the financial and examines how different elements and concludes. assumptions of the CLASS model affect model output. Section 6 3

6 Overview of the CLASS Model Framework and Results 2. Framework and Analytical Approach 2.1 CLASS model is designed to project net The income and capital for individual banks and BHCs over a period of two to three future (the “stress test horizon”) under different macroeconomic and years financial market scenarios. scenarios are defined by a set of economic and The macroeconomic market financial – such as GDP growth, the unemployment rate, housing prices, equity variables prices, short ‐ term and long ‐ term interest rates, and credit spreads – that are likely to influence the profitability banking institutions. The key outputs of of projections the model are CLASS of net and capital given assumed paths income these economic and financial market variables over the for stress test horizon. Figure 1 summarizes the CLASS model’s structure and the main steps involved in generating income core and projections. The model’s capital is a set of regression equations that are used to project margin how various financial ratios (e.g. the net interest (NIM), net charge ‐ off rates on different the types of loans) evolve over time, conditional on macroeconomic conditions, the lagged value of 4 financial and other controls. ratio, projections are converted to dollar values by These ratio by loan balances (in multiplying the case of loan loss rates), securities balances (in the case of losses), or assets (in the case revenue and expense items). The securities loss, revenue, and of expense projections are then combined to compute projected Changes ‐ net income. in pre tax capital and regulatory capital regulatory are derived by combining these pre ‐ tax net income ratios projections with assumptions about dividends, taxes, and regulatory capital rules, along with growth assumptions about of risk ‐ weighted assets (RWA). Details of the design and specification of the model CLASS Section auxiliary assumptions are presented equations in and 4 and the Appendix. income and capital projections are computed for each of the 200 largest U.S. banking Net st (BHCs and independent banks) and for a hypothetical 201 organizations representing the firm aggregate of the rest of the U.S. banking system. Individual firm projections are summed to generate system ‐ wide results. 4 structure. As discussed in greater detail in Section 4, the regression equations have an AR(1) 4

7 time model projects net income and regulatory capital ratios as they would occur over The CLASS particular macroeconomic scenario, rather than the generating estimates of marked ‐ to ‐ under values of the banks’ assets or capital or market estimating the impact of an instantaneous roll ‐ forward of peak ‐ to ‐ trough scenario U.S. such, the CLASS model projections follow conditions. As regulatory accounting generally principles (GAAP) and U.S. accepted capital rules. In particular, loss and revenue projections reflect the U.S. GAAP treatment of the underlying positions. The CLASS model uses 22 regression equations to project the components of pre ‐ tax net income. The provision first major component of income is pre net revenue (PPNR), an accounting measure ‐ defined as: 1) net interest income (interest income earned minus interest expense) plus 2) non ‐ interest (including trading income, as well as income non ‐ trading noninterest income earned from fees and other sources), minus 3) non ‐ interest (compensation, expenses expense related to premises and fixed assets, and other non ‐ credit ‐ related expenses). model The next net income component is provision expense for loan and lease losses. The CLASS NCO first projected net charge ‐ offs (NCOs) based on rates on 15 different categories of computes CLASS includes loans. offs a rule that then translates current net charge ‐ and the level of loan loss into provision expense, since under reserves U.S. GAAP, it is provision expense rather than charge ‐ described offs that directly affects net income. This provisioning rule is in Section 4. ‐ tax net Pre income equals PPNR minus provision expense for loan losses plus projected gains or (AFS) losses on investment securities held in the firm’s available ‐ for ‐ sale maturity and held ‐ to ‐ (HTM) The model includes an econometric model portfolios. for AFS returns. Returns on HTM portfolios, which are generally small for most firms, are assumed to zero. After ‐ tax be net income is calculated using a constant, assumed tax rate applied and to all banks BHCs. CLASS allows firms to accumulate deferred tax assets (DTAs) as a result of pre ‐ tax losses incurred. However, since U.S. regulation limits the extent to which these DTAs can be recognized for regulatory capital purposes, limits. CLASS includes an adjustment to recognize these 5

8 In the final step, CLASS computes the evolution of capital for the firm, based on the path of net 5 and with a behavioral rule for dividends other distributions. income combined Following practice in DFAST and CCAR stress tests, the primary capital metric in the CLASS model projections is Tier 1 the defined as common equity minus common the deductions from Tier 1 capital (such as equity, rules. assets) required under U.S. regulatory capital certain Capital ratios are calculated intangible using U.S. regulatory capital rules prevailing at the the of” date of the projections (the last “as historical observation), including the definitions of regulatory capital and rules for calculating risk ‐ weighted The CLASS model results presented in this paper primarily reflect Basel 1 risk assets. 6 weights and regulatory capital definitions, since these are the rules under which U.S. banks and BHCs their risk ‐ weighted regulatory capital ratios in calculated 2013:Q3, the “as of” date of the projections. Future versions of the CLASS model will incorporate Basel 3 risk ‐ weighted asset and regulatory definitions, as those capital force in the come United States. into 2.2. Net Income Capital Projections and This section presents CLASS model net income and capital projections under two macroeconomic or scenarios: a “baseline” scenario representing an expected median path for the economy and experienced financial conditions markets, and a “crisis redux” scenario that replicates during the scenario ‐ financial crisis. The baseline scenario is the developed by the Federal Reserve for 2007 09 The crisis redux CCAR. represents a repeat of the actual path of economic conditions scenario 7 from the third of 2007 onwards. experienced quarter We seed the model with BHC and bank balance and income data as of 2013:Q3. From sheet this starting point, we use the CLASS framework to compute income and capital projections over the subsequent nine quarters under 5 The behavioral rule for dividends is described in Section 4. 6 largest An important exception is trading ‐ related risk ‐ weighted assets at the BHCs, which are calculated under “Basel rules starting with the first quarter of 2013 and for all subsequent quarters. These rules significantly 2.5” weighted ‐ related risk ‐ increase assets at these firms. trading 7 the crisis scenario uses the historical Specifically, path for the transformation of each macroeconomic redux variable as it is used in the CLASS model. For example, one of the macroeconomic forcing variables in the CLASS we set the quarterly change in the unemployment rate. Correspondingly, for the crisis redux scenario, model is the change in the unemployment rate from 2013:Q2 onwards equal to the historical change in the unemployment rate onwards. from 2007:Q3 6

9 8 scenario. each Macroeconomic and financial conditions under the baseline and crisis redux are summarized in Table 1. scenarios projections 2.2.1 Income 2 presents the industry ‐ wide CLASS Figure projections under these two scenarios for components of ‐ provision net income, and for loan performance as measured by the net charge ‐ pre rate. off Recall and model projections are computed firm ‐ by ‐ firm quarter ‐ by ‐ quarter; the model then that the industry calculates projections by summing all dollar projections across firms, and computing ratios based on these sums. industry The upper panels of Figure 2 present projections for different PPNR components: net interest margin, return noninterest on assets, and non ‐ trading trading income and noninterest expense by total assets. The green line in scaled each graph represents baseline scenario projections, while the yellow line represents projections under the crisis redux scenario. As the figure illustrates, the CLASS model projections are quite sensitive to the scenario, with the and financial market conditions stressed of economic the crisis redux scenario generating projections of losses, revenue and expenses that are significantly more severe than those under the deteriorates baseline scenario. In particular, with the exception of NIM, each component of PPNR significantly the crisis redux scenario. Projected trading under and income volatile, is significantly in the worst quarters of the scenario, approximately matching its behavior negative during the financial crisis. Non ‐ trading noninterest income also but is less volatile quarter ‐ to ‐ deteriorates, quarter due to the more highly autoregressive statistical model used for this category. In addition, noninterest expense total assets is significantly scaled elevated by under the crisis redux scenario. Aggregate PPNR (bottom left panel of Figure 2) falls sharply in the crisis redux scenario and is actually projected to be negative at of the worst point the scenario, an outcome not observed at any over our historical point sample period. 8 As explained in Section 4, our approach to loan loss provisions uses projected future net charge ‐ offs in modeling the subsequent four quarters as an input into computing the value of ALLL at each point in time. Correspondingly, order to in calculate provision expense a longer thirteen actually quarter horizon, we project net charge ‐ offs over ALLL over the nine quarters of the scenario proper. For this reason, each macroeconomic scenario is actually and length. specified to be thirteen quarters in 7

10 right panel of Figure 2 plots the projected industry net charge ‐ off ratio, a summary ‐ The bottom realized credit losses. This ratio rises of sharply under the crisis redux scenario, measure (albeit not reaching) the peak NCO rate realized during the financial crisis. The approaching NCO rate is essentially flat baseline scenario, implying that the NCO ratio as of 2013:Q3 is close to in the long ‐ term steady state its value. Although not shown in the figure, provision expense, which is closely linked to NCOs, these patterns. mirrors Figure 3 plots annualized industry ‐ level projected return on assets (ROA), defined as annualized net a percentage of total assets. Final net income reflects the sum of the income components income as Figure 2, as well as projections for other presented components of net income such as the model in for AFS returns. ROA falls sharply under the crisis redux scenario, mirroring its realized during path crisis itself, although with the differences. financial This variation between the historical crisis some ROA and the projected ROA path under a repeat of the same macroeconomic conditions reflects crisis two factors: first, some losses experienced during the are not fully captured by the CLASS example framework, for because they occurred during quarters when the macroeconomic forcing not deteriorate significantly, and second, the set of banking data that variables is used to seed did banking the is different, due to changes in the model system between 2007 and 2013 (e.g. firm entry and exit, changes in the composition loan of system assets and income, and shifts in banking performance, ALLL, and income expense ratios). and 2.2.2 Capital projections Tier Figure 4 presents CLASS model projections for the 1 common equity capital ratio (Tier 1 common capital as a percent of risk ‐ weighted assets) for the U.S. banking industry. Panel A of the the industry ‐ level ratio, calculated as the weighted average for the BHCs and figure banks presents in CLASS framework, using risk ‐ weighted assets as the weights. As illustrated in the panel, the industry ‐ level Tier 1 common ratio rises slowly and steadily under the baseline scenario. This ratio from under the crisis redux scenario, however, declines sharply a historical value of 11.9% in 2013:Q3 to a level of 10.1% after the ninth quarter of the scenario. This drop approximately industry matching the magnitude of the decline in capitalization experienced during the 2007 ‐ 09 period. financial crisis 8

11 projections in Figure 4 and elsewhere in this paper are point estimates that do not reflect the The of The statistical uncertainty around our conditional forecasts. width of the confidence degree will depend significantly our estimates or assumptions about the joint variance ‐ intervals on of the matrix regression coefficients covariance across 22 CLASS model regression models. all Currently, we do not estimate this joint matrix, since we estimate each equation separately, rather intervals a system. Exploring these confidence than and the correlation of the equations as methods. avenue for future work, using bootstrap represents an Figure of Panel B 4 looks at the distribution of projected capital across the cross ‐ section of BHCs and banks. it plots the cumulative distribution function of capital: the percentage of Specifically, that are held in industry firms with a Tier 1 common assets ratio lower than different banking thresholds between 0% and 15%, as plotted on the x ‐ axis of the figure. For each scenario, we the present function during the “worst” quarter, that is, this quarter of the scenario in which the projected industry capital ratio is minimized. In practice, this is the first quarter of the baseline of scenario and the ninth quarter the crisis redux scenario. the The cumulative distribution of the Tier 1 common ratio is shifted significantly to the left under crisis scenario relative to the baseline scenario. Reading off the figure, at the low point of the redux scenario, around one ‐ tenth of industry assets are owned by firms baseline Tier a common with 1 of less than 10%. But under the crisis redux scenario, more than three ‐ quarters of industry ratio held in firms with a Tier 1 common ratio below this same threshold. Even under the crisis assets are redux scenario, however, only a small fraction of industry assets held in firms are with a projected Tier 1 common ratio below 5%, the threshold referenced in the Federal Reserve’s 2011 Capital Plan 9 Rule. Note that the leftward shift in the distribution of capital under the crisis redux scenario (relative to baseline) not entirely parallel ‐‐ projected capital declines more significantly for some firms than is 9 The Capital Plan Rule requires bank holding companies to demonstrate in their capital plans how the firm will maintain a minimum tier 1 common ratio of more than 5% under stressful conditions, and provides that the plan. This capital rule applies to ability to do so assessing the firm’s Federal Reserve will evaluate the firm’s in (2011). banking firms with at least $50 billion in total assets. See Board of Governors of the Federal Reserve System 9

12 others. Reflecting this, the variability in the final projected Tier 1 common ratio across firms is more under the crisis redux scenario than under the baseline scenario. diffuse CLASS to Analyze Using in Financial Stability 3. Trends this section we use the CLASS model as a tool to analyze trends In in financial stability, with a focus adequacy under stress. In the time series, we evaluate how on banking system has capital the evolved in terms of being able to a severe macroeconomic downturn without banks withstand becoming undercapitalized or shrinking in size. We then look across the cross ‐ section at the exposed banking firms that are particularly to a macroeconomic downturn characteristics of through lens of the CLASS framework. the of 3.1 Evolution capital “gap” the As a summary measure of system ‐ wide undercapitalization, we use the CLASS projections described “gap” above compute an estimate of the total capital to – that is, the projected dollar capital required to bring each BHC and bank up to a given threshold capital ratio under the injection this industry scenario in question (or equivalently, the total dollar capital shortfall relative to threshold). calculate this capital gap firm ‐ by ‐ firm, and then sum across We firms, reflecting the fact capital is not fungible across institutions, and compute the gap in the quarter in that which the industry capital ratio is minimized over the stress test horizon. capital 5 plots the time series evolution of the Figure gap under the crisis redux scenario, relative to two Tier 1 common / RWA thresholds, 5% and 8%. This figure is constructed by computing the data CLASS projections repeatedly using different historical quarters of banking model to “seed” the 2013:Q3). (we this every quarter between 2002:Q1 and vary We hold the model parameters and macro scenario constant across these runs, only variation in the results so reflects changes in the time characteristics of the banking system over time. The of series path the resulting capital gap can be viewed as an index of how the vulnerability to undercapitalization of the US banking system has evolved, under a measured given stressful macroeconomic (i.e., in this case, the conditions scenario crisis). experienced during the 2007 ‐ 09 financial 10

13 The capital gap relative to an 8% Tier 1 common threshold is approximately $100 billion in 2002, reaching rises over time, particularly during 2007 and 2008, a peak of $540 billion in the and then quarter of 2008. To reiterate, this value implies that if we substitute 2008:Q4 balance sheet fourth and income data for banking firms into the CLASS model and compute capital projections under the of scenario, then by the low crisis of the scenario, CLASS projects a shortfall redux $540 billion point in projected Tier 1 common equity to an 8% threshold. relative This upward trend in the capital is reversed from 2009:Q1 onwards ‐‐ the capital gap falls gap by 2009 and 2013, reflecting equity issuance firms, lower dividends and other sharply between distributions, as well as a capital to profitability for most banks and BHCs. The measured return capital gap as of 2013:Q3, final bar on each the graph, is $8.5 billion relative to an 8% capital ratio threshold. is only about one ‐ tenth of its value This in 2002, even though industry assets have grown significantly over the intervening period. gap similar trends are evident for the capital Broadly measured relative to a 5% threshold, although the 5% of level of the gap is course smaller at each point in time. The capital gap relative to a threshold generally close to zero except in the period between late 2006 and 2011. This gap is $304 billion, also peaks in 2008:Q4. at A notable feature of Figure 5 is that the capital gap begins to increase in 2004, well estimated before the onset of the financial crisis. This increase partially reflects growth in the nominal size of the system, although this is not the banking main explanation: between 2004:Q1 and 2007:Q1 banking system assets increase by 33%, but the capital gap rises by a much larger 83% (from $113bn path to $206bn). This time series of the capital gap implies significant deterioration in the commercial industry’s capital adequacy under stressful economic conditions in the years banking financial to the crisis. leading up caveat is that the capital gap One path in Figure 5 is based presented on the full ‐ sample CLASS model econometric estimates, and thus is not truly “ex ‐ ante” in nature. Would this upward trend in in the capital gap prior to the financial crisis have been identifiable real time using our framework? To answer this question, we ‐ series, computed a “point ‐ in ‐ time” version of this capital gap time 11

14 using regression models estimated only using data up to the quarter in question, rather than the full computed the capital gap as of 2002:Q1 is (e.g. using regression models based on data from sample 1991:Q1 to 2002:Q1 only). A comparison of the “point ‐ in ‐ time” and “full sample” versions of the industry capital shortfall is presented in Figure 6. Note that we observe a very similar build ‐ up in the capital gap using this point ‐ in ‐ time approach to the results based on full ‐ sample estimates. For virtually instance, the estimated real ‐ time capital gap doubles between 2004:Q1 and 2007:Q1 (from $82bn to $163bn), actually a larger percentage increase than the 83% change computed using the full model. ‐ sample level of the measured capital gap prior to the financial crisis is lower under the point ‐ in ‐ time The underlying reflecting that some of the econometric models the CLASS framework are less approach, to macroeconomic conditions when estimated over a sample period that does not include sensitive 10 financial crisis. the Recession the financial crisis and Great period is included, however, the Once projected capital gap based on the point ‐ in ‐ time and full ‐ sample versions of the model are quite similar, and over the last 8 quarters or so of the sample are almost identical. This is consistent with our practical as we have updated the CLASS model progressively in recent periods. The experience and Great Recession period has significant effects on many of financial our regression crisis it represents a coefficients, period of high volatility in earnings and because macroeconomic conditions, helping to identify our estimates. But the models parameter are relatively stable to the addition new data points of in recent years. It is interesting to compare these projected capital gaps with market ‐ based measures of stress capital capital adequacy. In Figure 7, we compare the evolution of the gap from CLASS to the “SRISK” measure of capital shortfall developed by researchers at New York University (Acharya, Engle Richardson, 2012, and Acharya, Engle and Pierret, 2013), and to credit default swap (CDS) and banks, U.S. investment banks and commercial spreads drawn from Bloomberg. (To compare for these on a common scale, we normalize different each variable by its average measures value in 2002, the first year of the sample.) SRISK computes capital shortfalls for financial firms based on 10 For instance, residential mortgage credit losses are low and stable prior to the crisis, due to the rising home price ‐ environment. As a result, our residential mortgage net charge off models exhibit little sensitivity to home price sample. growth unless the crisis period is included in the regression 12

15 market equity values and time series models of stock returns. Two SRISK measures are presented, on the GMES and MESSIM models maintained by the NYU Stern Volatility Lab. Measures based 11 although based on the same basic modelling approach, they differ in some details. are shown these measures rise sharply as the financial crisis unfolds in All and 2008, and the market ‐ 2007 based measures peak at higher normalized values than the CLASS capital gap. But notably, the rise in in CLASS capital gap leads the increase in SRISK the the period leading up to the crisis, particularly so in 2006 and early 2007. And most strikingly, CDS spreads of large U.S. banking organizations were extremely and actually falling in the period from 2004 until mid ‐ 2007, despite the risks building low period the system during this up (see also Eichengreen et al., 2012). in What these divergent trends? One plausible reason explains is the low risk premiums and high market valuations of U.S. banking firms prior to the financial crisis. Calomiris and Nissim (2012) document the average market ‐ to ‐ book ratio for that public banking firms exceeded 200% in the seven years prior to the crisis, to to around 100% in 2010 compared 2011. Low risk premia for bank to debt and equity, even if driven by speculative factors rather than fundamentals, will tend improve based measures of financial stability. We interpret the results in Figure 7 as market ‐ that careful analysis of bank accounting evidence of even the benefit without confidential data, information, can help supervisory provide useful early warning signal information about capital under stressful conditions, beyond information encapsulated in market prices. adequacy 3.2 Capital sensitivity to macroeconomic conditions: Cross ‐ sectional analysis conditions of projected net income and capital to macroeconomic The varies significantly sensitivity mix firms, across to differences in firms’ asset and income ‐ generating activities. To examine this due cross ‐ sectional variation in more detail, we compute for each firm the change in the Tier 1 common ratio equity over the course of the nine ‐ quarter crisis redux scenario (i.e., the difference between the end ‐ of ‐ scenario ratio under the crisis redux scenario and their last historical Tier 1 firm’s 11 We thank Robert Engle and Viral Acharya for providing historical time series for these two measures. The GMES model is based on the Dynamic Conditional Beta approach of Engle (2014), measured relative to the MSCI World pricing model measures simulation estimates are based on a while approach and capital asset Index, the MESSIM of with respect to the S&P 500 index. Regularly updated SRISK estimates are publicly available on the NYU beta http://vlab.stern.nyu.edu/welcome/risk/. Stern V ‐ Lab website: 13

16 common equity ratio). The more sensitive the firm’s net income and capital are to adverse conditions, the more negative this change in capital will be. We do this firm ‐ by ‐ firm macroeconomic each different points in time between 2002 and 2013:Q3 for of the 200 largest banking firms at at each point in time (a total of 200 firms x 47 quarters = 9400 observations). the illustrates correlations between the change in capital ratio and various firm Table 2 at different points in characteristics time, including: i) the starting Tier 1 common equity ratio of the firm, ii) a simple measure of asset liquidity, namely the sum of cash, interest bearing balances, securities federal funds expressed as a percentage of total assets, iii) a regulatory ‐ based and to asset risk, namely the ratio of risk ‐ weighted assets measure total assets, and iv) firm size, of measured the log of total assets. by In each case, we are interested in the overall cross ‐ sectional correlation over the sample period, as well as whether the correlation has evolved in recent years and due the introduction of supervisory stress testing to other changes in the regulatory and economic environment. We measure this by including an interaction term between the banking firm characteristic and a dummy equal to one from 2011:Q1 onwards. All regressions also variable identified include time fixed effects (i.e. a dummy for each quarter), so that the correlations are only on cross ‐ sectional variation across banking firms, rather than time ‐ series shifts in based bank and capital stress. Our main results are robust to the exclusion of characteristics these time fixed We cluster standard however. by entity. effects, errors primary finding from this is that in the recent period Our (since 2011), the projected change analysis in capital during the crisis redux scenario is correlated with the initial capital significantly negatively – in other words, the ratio ratio is projected to decline more steeply under stress for highly capital capitalized firms. This inverse relation is consistent with a “precautionary” view of bank capital structure (e.g. as discussed in Berger et al., 2008). Such a view argues that banking firms with more the risky income will endogenously choose to hold a larger capital volatile buffer, to reduce or likelihood of becoming undercapitalized. On the other hand, Berger and Bouwman (2013) argue that a risk ‐ shifting view or moral hazard view would yield the opposite prediction, that less ‐ well will be incentivized to hold capitalized riskier banks asset portfolios in equilibrium. This inverse is not observed prior to 2011 (in the earlier period the correlation is actually positive, relation as although not statistically significant). The difference in the strength of this relationship, 14

17 measured by the interaction term, is statistically significant at the 1 percent level in each specification. find relatively little correlation between the liquid We asset ratio and the projected capital decline the crisis redux scenario, although in column (7), firms with a high share of liquid during assets are found to Perhaps be sensitive to macroeconomic conditions. less counterintuitively, firms with a ratio of higher risk ‐ weighted assets to total assets actually experience a smaller projected decline in capital during the crisis redux scenario. This latter result suggests that the Basel I measure of risk weighted used over this sample assets best be a poor, or at noisy, measure of the period may of a banking firm’s assets to macroeconomic stress. sensitivity example, large diversified firms For with significant trading operations and securities portfolios hold a smaller fraction of assets in the form of loans, which attract a higher Basel I risk tend ‐ But such firms to be significantly weight. exposed to macroeconomic stress due to the volatility of trading income and other noninterest income. Finally, the projected capital decline is larger (i.e. more negative) for larger banking firms, 2011. particularly since Complementing table, Figure 8 shows how this relationship between the capitalization and initial change in the capital ratio over the stress scenario has evolved quarter ‐ by ‐ quarter since 2002. the before, to construct this figure, we use the CLASS projections As of capital (Tier 1 common equity, as before) under the crisis redux scenario for each of the 200 largest firms point each at in time between 2002:Q1 and 2013:Q3. We then regress the change in the ratio under the crisis redux sectional scenario on the initial capital ratio in each quarter (i.e., 47 separate cross ‐ The regressions). figure the time ‐ series evolution of the plots slope coefficient from that bivariate regression. Corroborating the evidence from Table 2, since 2011, firms with assets and income that are highly consistently exposed to the crisis redux macro scenario also have higher capital ratios. However, not this is true prior to 2011. During the 2008 ‐ 2010 financial crisis period, such “exposed” firms were less actually that well capitalized, likely reflecting the large losses experienced during the fact crisis depleted their capital ratios. Prior to 2008 the relationship was either positive or at best had negative. weakly 15

18 We highlight that caution should be exercised in applying a causal interpretation to these results, endogenously capitalization and other bank characteristics are chosen by the firm, and given that 2 likely to be correlated with a range of omitted variables. We also note that the overall R are of the in table 2 is quite low (ranging from 11.8% in column 2 to 19.3% for column regressions 7), implying that these broad firm characteristics account for only relatively small fraction of the variation in a the sensitivity of capital to macroeconomic shocks estimated by the CLASS model. Bearing in mind, however, the prima facie evidence that firms’ capital policies have these caveats become precautionary in nature in recent years more appears encouraging from a financial stability point of view. One possible explanation why capital policy has evolved, at least for the largest firms, of is implementation the annual supervisory stress tests by the Federal Reserve ‐‐ these tests are remain explicitly designed to ensure that all firms well ‐ capitalized even under a severe risk macroeconomic downturn. Other changes since the financial crisis, such as improved management, awareness of downside risks, or changes in supervisory practices, may also greater have affected firms policies, especially among capital planning with riskier portfolios. While banking beyond the scope of this paper, investigating these issues in more detail would be an interesting topic future research. for 4. Model Details We now turn to a more detailed description of the structure of the CLASS model regression equations, the data used to estimate the equations, the resulting specifications and parameter estimates based on historical data through 2013:Q3, and the auxiliary assumptions needed to complete CLASS model projections of net income and capital. Figure 9 presents a detailed the schematic CLASS model structure, including regression equations, calculation of the steps, and auxiliary assumptions. 4.1. equation structure Regression Each CLASS regression equation models a key income or expense ratio as a function of an autoregressive term and a parsimonious set (AR(1)) of macroeconomic variables. Some equations are and as time ‐ series models using historical data summed up across all BHCs estimated banks. 16

19 Other models are estimated using pooled quarterly data on individual firms, allowing us to control firm characteristics such as the composition of assets. for series specifications take the general The form: time ratio β α + ε ratio + macro + = β 2 t ‐ t 1 t 1 t ratio where an interest the financial ratio of and ratio is AR(1) term, macro is the set of is 1 t t t ‐ appropriate macroeconomic to that ratio. When statistically and economically significant, variables 12 also include a linear time the trend in the specification equations . For the models estimated using pooled individual BHC and bank data, the specification is: ratio α + β ratio + β = + β X + ε macro 3 t,i 2 t,i t ‐ t,i t , 1,i 1 where each observation is now indexed by firm i , and the equation includes X vector a , of firm ‐ t,i 13 types such as shares specific of different of loans in the loan portfolio characteristics, or the share of securities in the investment securities portfolio. risky Pooled regressions are estimated for the AFS returns equation, and for components of PPNR significantly affected by the composition of firm assets, such as net interest margin, compensation expense, and other non ‐ interest expense. clustered Standard errors are by time. will for The autoregressive nature of each equation implies that the projected ratio each firm converge from its most recent historical value towards a long ‐ run steady state value. These slowly The will significantly influenced by the assumed macroeconomic scenario. autoregressive paths be means that also CLASS model structure the projections are sensitive to the lagged value of the ratio to bank and BHC data, which are used for “seed” the model projections. The seed data is each 12 Time trends appear in three of the 22 CLASS models, and are intended to capture long ‐ term trends econometric in particular financial ratios over our sample period (for example the secular decline in net interest margin). In each calendar by 0.25 each quarter. When to zero in 1991:Q1 and increases case, the time trend is normalized historical generating projections, we hold the time trend constant at its most recent model value, rather than assuming the trend continues over the forecast horizon. 13 For example, the net interest margin (NIM) equation includes controls for the composition of the firm’s loan portfolio. is necessary because interest margins vary significantly This across firms (e.g. margins are higher for This facilities). card high concentration of credit card loans, due to the high interest rates on credit with firms a shares. implies that even the long ‐ run NIM projection will vary across firms, reflecting differences in these portfolio 17

20 particularly important for income and expense categories that are estimated to be highly (that is, with a large value of β autoregressive such ); categories, a low or in high ratio value in the 1 quarter used to seed the model will have persistent effects on the projected income path historical 14 stress test horizon. over the On occasion, the autoregressive structure of the CLASS regression create unrealistic shifts in projected income and equations capital in cases when an individual can bank experiences an idiosyncratically large income spike BHC that is unlikely to be repeated in or future quarters (e.g. realization of a large loss related to a acquisition). In such cases, we legacy shock correction to the model projections so that the in question does not have a persistent apply a on projected income. effect practice, we make such judgmental adjustments to the model In projections only rarely. 4.2 Data estimate To equations described above, the we combine two types of data measured at a quarterly frequency: report data on balance sheets, regulatory income and loan performance, and macroeconomic and financial market data used in the macroeconomic scenarios. Federal BHC and bank regulatory data are drawn from The Reserve Y ‐ 9C regulatory filings for BHCs and FFIEC banks. Condition Consolidated Reports of and Income (Call Report) filings for commercial The are based on quarterly data from 1991 to the present for all BHCs that file the FR Y ‐ regressions 15 the subset of commercial banks that do not have a parent that files a FR Y ‐ 9C, plus The data 9C. all U.S. ‐ headquartered, top ‐ tier BHCs and include independent commercial banks, as well as six large foreign ‐ owned BHCs subject to CCAR in 2014. Other BHCs and commercial banks whose parent domiciled outside the United States are excluded, as is are two BHCs that are not engaged in traditional commercial banking activities: DTCC and ICE Holdings. As noted above, the majority of the regression specifications are based on an aggregated time series the banking system, calculated by summing data across the individual banking firms. for institutions These are subject to breaks when new aggregate become banks or BHCs series or when 14 On the whole, this persistence is realistic, given the historical dynamics of bank income, and given that the . are estimated to maximize fit models to the historical data regression 15 This includes commercial banks that are self ‐ held and commercial banks that have holding companies that both filing. are too small to file a consolidated regulatory Y ‐ 9C 18

21 a BHC makes a significant acquisition from outside the banking industry. For example, the of Goldman Sachs and Morgan Stanley to bank holding companies significantly conversion in total assets (appearing in our data 2009:Q1); similarly, acquisitions of non ‐ increased industry financial firms, such as J.P. Morgan Chase’s acquisition of bank Washington and Bear Mutual Bank of America’s acquisition of Merrill Lynch, also create discontinuities. We do Stearns, not and part make adjustments for these sample breaks, in any because the pre ‐ conversion or pre ‐ acquisition data on the target firm needed to make such adjustments are not readily available in a format with the Call and Y ‐ 9C reports. However, since the regression variables are comparable ratios – and the newly converted or specified acquired institution enters both the numerator and as denominator of the ratio – the impact of these breaks is muted. The regression specifications based on a pooled sample firms rather than of aggregate industry data are estimated using a panel of the 200 largest banking institutions by assets in each quarter. The of remaining banks and BHCs are aggregated into a single observation, resulting in a total sample 201 entities. of equations include parsimonious The regression nine combinations macroeconomic and financial variables summarizing economic activity and financial market conditions. The final market each equation was based on a specification search of based on measures of overall specification 2 2 fit (R model and adjusted R ) as well as statistical significance of the macroeconomic variable, and rates accordance economic theory (e.g., that chargeoff with are positively correlated with poor economic conditions). The macroeconomic variables we use are a subset of those included in the scenarios provided by the Federal Reserve for the DFAST stress tests, and include the 10 ‐ year yield, the 3 ‐ month Treasury bill yield, the civilian unemployment rate, Treasury real gross bond (GDP), the CoreLogic U.S. home price product index, the BBB bond index yield, domestic commercial real estate prices and the U.S. Dow Jones Total Stock Market Index. Table 3 provides a included full of macroeconomic and financial market variables list in the CLASS model equations and describes the transformation of each variable used in the regressions (that is, whether the variable is expressed in levels, changes, percent changes, or some other form). estimates 4.3 Regression model 19

22 The CLASS model includes six regression equations for components of PPNR, fifteen equations for off (NCO) rates on different loan categories ‐ (e.g. first ‐ lien residential real estate, net charge loans, credit cards, C&I loans), and an equation for gains and losses on the AFS construction securities portfolio. Table 4 presents summary statistics the twenty ‐ two for ratios that are as part of projected the CLASS framework. Table 5 summarizes the set of macroeconomic variables included in each equation, and indicates which are statistically significant. Full equation specifications parameter estimates are presented in the appendix. and model specifications final used in The CLASS model represent the result of search the regression of specifications over different combinations of macroeconomic and controls; in some cases variables we also varied other modeling choices such as the weighting of each observation in the regression sample the functional form of the or macroeconomic variable. The Online Appendix presents a more detailed description of how the specification search was conducted, and presents estimates searches per for a number of the different specifications we tried (one table of specification equation; tables in total), as well as a graph of the in ‐ sample fit of each preferred 22 econometric model. In model cases, at least six almost different all specifications were estimated and considered for each equation. In choosing specifications, we put weight both on statistical fit and consistency with intuition, rather than relying on a economic purely mechanical approach to model specification such as LASSO. In part this is because of statistical concern that a purely our approach could lead to the the risk of overfitting available relatively limited time series history. 4.3.1 PPNR model contains six regression equations for components of PPNR, including net The interest CLASS expense), is, interest income minus interest (that trading income (which includes both income mark ‐ to ‐ market changes in value of trading positions and derivatives as well as fee and spread income income trading activities), non ‐ interest non ‐ trading on (such as deposit fees, income from fiduciary activities, and revenues from investment banking and insurance), and three components compensation of noninterest expense: expense, expenses related and to premises and fixed assets, 20

23 16 non ‐ interest expense. other Each of these components of PPNR is expressed as a ratio either of compensation (for non ‐ interest, non ‐ trading income, expense, fixed asset expense, and total assets non ‐ interest expense), trading assets (for trading revenue), or interest ‐ earning assets (for net other income). interest equation except PPNR for return Each trading assets is estimated by weighted least on squares using the pooled regression approach, weighting by the institution’s share of the relevant denominator asset balance (e.g. interest ‐ earning asset share in the case of net interest margin). Pooled include controls for the composition of firm assets and firm size: the regressions ratio of loans, estate loans, commercial real residential estate real commercial and industrial loans, credit card loans , trading assets , and securities to interest earning assets, and the firm’s assets scaled by the industry in assets quarter. Given these controls, the projected PPNR ratio for each BHC or bank converges to the long ‐ run unconditional conditional mean for firms with similar business focus and size, rather than the sample These controls are particularly important for the net interest margin model, since the mean. between and lending rates varies significantly across types of loans. For example, spread borrowing card balances historically credit have high net interest margins, compensating for the higher credit risk associated with these loans. is our final specifications, the net In interest margin positively related to short ‐ term Treasury yields as well as the slope of the yield curve, trading returns are sensitive to credit spreads (the change in the yield spread between BBB ‐ rated corporate bonds and 10 ‐ year Treasuries), and non ‐ trading noninterest is sensitive to stock returns. Compensation expense is positively correlated with income to stock while other noninterest expense is sensitive returns, credit spreads. As shown in the detailed results presented in Appendix A, most components of PPNR are highly autoregressive. ‐ 4.3.2. Net Charge Off Rates Loan 16 experimented with similar models for aggregate PPNR, however, explanatory power and sensitivity to We model. macroeconomic conditions are lower for the aggregate 21

24 The CLASS model includes 15 net charge ‐ off (NCO) models for major loan categories: first lien and lien residential mortgages, home equity lines of credit (HELOC), construction loans, junior mortgages, non ‐ farm non ‐ residential commercial credit cards, other consumer multifamily and commercial and industrial (C&I) loans, leases, loans to foreign loans, governments, loans to institutions, agriculture loans, other real estate loans, and all other loans. In each case, depository so dollar ‐ offs are scaled by the corresponding loan balance, net that the regression charge dependent variables is a loss rate. NCO rates on real estate loans are primarily associated with real estate downturns. From a price mortgage default represents a put option on the underlying real estate theoretical perspective, to collateralize the loan (e.g., Kau et al, 1992). used with this point, the empirical Consistent relationship between real estate price growth and real ‐ estate loan charge ‐ offs is highly non ‐ linear, with real estate price declines having a much larger effect on charge ‐ off rates than real estate price increases. this reason, the final equations include For an interaction between property price growth and a dummy variable for whether the change in the price index is less than zero. Quantitatively, this interaction term the is the key macroeconomic determinant of mortgage NCO rates in 17 models. most other loan types, the change in the unemployment rate was generally the macroeconomic For variable most correlated with loan losses, with increase in the unemployment rate causing an charge ‐ off rates to increase. Across the entire spectrum of loan categories, net charge ‐ off rates are highly with AR(1) autoregressive, coefficients between 0.5 ranging and 0.9. 4.3.3. Returns on Available ‐ for ‐ Sale (AFS) portfolios Realized gains and losses in a banking firm’s AFS securities portfolios occur only when the firm sells deemed those assets or the securities are Temporary to have experienced “Other Than Impairment” OTTI. Under current GAAP accounting, or OTTI is determined only status by credit 17 Residential mortgage charge ‐ in particular were low and relatively insensitive to macroeconomic conditions offs until the recent financial crisis. Although commercial real estate charge ‐ offs were high in the early 1990s, NCOs in indicators category were also low between this episode and the recent crisis. We found that business cycle this such as the change in the unemployment rate were generally statistically insignificant once we controlled for real specifications. estate price growth; consequently these variables were not included in the final 22

25 factors, and need not incorporate changes in market prices due to interest rate risk, liquidity or of until the bonds are sold. Realized AFS gains and losses thus reflect a combination other factors, shocks, credit events, behavioral decisions price asset sales, and accounting judgment. asset about AFS returns are low Historically, and stable, but with occasional, large downward movements, 2008 and 2009. particularly during CLASS model’s approach to modeling realized gains and losses on AFS securities The recognizes the significant variation in the riskiness of these portfolios across firms and over time. Specifically, the model includes an interaction term between the share AFS securities that are “risky” (that is, of government and agency securities) and increases in the credit spread (BBB minus excluding U.S. 18 Treasuries). AFS returns are also found to be negatively correlated the change in Treasury with bond yields. 4.4 Auxiliary Assumptions 4.4.1 Balance Sheet Growth and Composition As above, the 22 regression equations produce discussed projections of accounting ratios – losses, revenues or expenses scaled by a loan, securities or asset balance. To translate these ratios into dollar values in order to calculate net income, the CLASS model requires projections of the balance sheet ‐ risk over the stress test horizon. Balance sheet projections are also needed to project weighted and to calculate capital ratios, since capital ratios have either risk ‐ weighted assets assets assets in the denominator. Because of total this mechanical relationship between capital ratios or asset balances, the results of CLASS and other stress testing models based on accounting data and are highly sensitive to the growth path of assets over the stress test as illustrated in the horizon, sensitivity exercise 5. presented in Section 18 Prior to 2001, BHCs and banks only reported the breakdown of risky securities into: securities issued by states and municipalities, foreign and domestic equity and debt securities. U.S. government agency and corporation the obligations reported without separately breaking out MBS. In were CLASS model, AFS securities backed by the and government or government agencies are “safe” assets that are unlikely to experience credit impairment U.S. thus incur OTTI. All other AFS securities are classified as “risky,” including municipal bonds, non ‐ agency mortgage ‐ backed securities and asset ‐ backed securities, and corporate debt. The aggregate fraction of AFS securities 2010. by half approximately to increased from less than 30 percent in 1994 assets risky of consisting 23

26 The model adopts a simple approach to balance sheet projections ‐‐ each BHC or bank’s total CLASS stress are assumed to grow at a fixed rate of per quarter (5% per year) over the test assets 1.25% growth rate was chosen to be This consistent with the long ‐ run nominal horizon. roughly historical assets in the U.S. banking industry. The same growth rate is assumed for all asset of growth implying that the composition of the balance sheet – that is, the proportion of total assets balances, trading different types of loans, securities, cash, by positions, and other assets – stays represented at its last historical value over the stress test. fixed The composition of liabilities is also assumed to stay while the book value of liabilities is calculated so that the balance sheet identity (assets fixed, and plus capital) holds equal at each point in time. If capital falls liabilities assets increase, the difference is made up with additional liabilities, with constant mix as of the starting quarter. Assuming that the growth rate of assets is the same for all institutions and for all scenarios is not “realistic” in the sense that banking industry assets historically tend to grow more slowly in stressed 19 environments than they do during expansions. economic However, assuming that banking both can industry assets continue to grow during economic stress be seen as rigorous from microprudential macroprudential perspectives. From a macroprudential perspective, it ensures and of banking industry capital strength are made in that the context of continued assessments 20 availability credit of firm from a microprudential , perspective, while ‐ level capital projections are made under the assumption that the firm continues to function as an active financial intermediary. Our assumption that balance sheet composition does not evolve with macroeconomic conditions is also not entirely consistent with historical experience. Incorporating scenario ‐ dependent shifts in composition would have two main effects in asset the CLASS framework. First, changes in the of relative share of risky and safe assets will affect projections a total net income via composition effect. example, a shift towards riskier loans types such as construction loans will increase the For overall loan rate, holding fixed the projected loss rate within each loan category. Second, asset loss shares used as control variables in several of are our regression models, particularly for components 19 Historical banking industry data illustrate that both the growth rate of bank assets and the composition of the balance sheet can vary significantly with economic conditions. For instance, Clark et al. (2007) document the cyclical variability in the share of retail ‐ related loans such as mortgages and credit cards. 20 approach. Greenlaw et al. (2012) argue in favor of this 24

27 of pre ‐ provision net revenue. Thus, movements in these shares would have effects on the projected variables in these equations (e.g., the net interest margin). dependent these Enriching model to explicitly incorporate the composition effects would be quite CLASS and is outside the scope of the present paper. However, complex, as first step, the Online a showing this paper presents econometric estimates based on historical data Appendix how the to categories industry assets in different asset of evolves with macroeconomic conditions. The share six categories we consider are cash and interest ‐ bearing balances, loans, trading assets, securities, federal and repos, funds asset Since the relationship between composition and macro and other. might vary variables with banking firm characteristics such as size, we estimate these regressions for the industry as a whole, and then separately for the largest 10 firms (resorted by total assets remainder each and for the quarter) of the industry. This preliminary analysis suggests that banking sector asset composition does indeed move with term macroeconomic conditions historically, particularly with the spread and with credit spreads the (the difference between BBB corporate bond yields and 10 ‐ year Treasury yields). An increase in term is associated with a contemporaneous shift from spread trading assets and to securities. loans increase in credit spreads is associated with a statistically significant shift out of trading assets An fed funds and repos into and securities portfolios and cash and interest ‐ bearing balances. At least in the latter case, our expectation is that incorporating these composition shifts would be likely to slightly reduce banks’ projected sensitivities to macroeconomic conditions somewhat (since in balances CLASS, projected losses on securities plus cash and interest bearing relatively are low and losses insensitive credit spreads compared to on trading assets). to is mixed between evidence that these relationships are different There small and large banks; as shown in the Online Appendix, we find a statistically significant difference in the macroeconomic sensitivities between these two size groups (at the 5 percent level) in five of the twelve for specifications. The evidence for heterogeneous sensitivities is strongest trading assets, perhaps not surprisingly given that small macroeconomic firms have few trading assets regardless of conditions. 25

28 Summing up, this initial analysis suggests that allowing for asset composition to shift with conditions could be a useful extension of the CLASS framework. We do however macroeconomic for see possible pitfalls in our “constant shares” assumption. First, allowing also relaxing adds significant complexity. Second, shifts to the extent that composition historical shifts in asset of periods shares stress represent flights to quality within bank portfolios that may not recur, during be appropriate from a macroprudential perspective it not allow for these channels when may generating projections. capital Allowance for Loan and 4.4.2 Losses (ALLL) Lease The CLASS model’s equations project total net charge ‐ offs each quarter over the stress test horizon. However, U.S. accounting rules, net charge ‐ offs do not directly affect net income. Instead, under increase recognize the provision expense incurred to accounting the allowance for loan losses rules reserve ALLL). This is not a straightforward exercise, (the however, since is estimated ALLL by the firm based on a set of accounting guidelines which leave scope for managerial discretion and judgment. an empirical matter, the choice of As provisioning rule has a quantitatively important 21 on net income and thus on the regulatory capital projections (see Section 5). effect The CLASS model assumes that the ALLL is in a range relative to projected future net bounded (under charge ‐ offs. If the ALLL is at least equal to the next four quarters of projected net charge ‐ offs 22 the scenario in question) macro but greater than 250% of that not level, then provision expense in quarter is set the equal to current ‐ quarter net charge ‐ offs. If the ALLL is below four quarters of future charge ‐ offs, then provision expense is set equal to an amount that would bring the ALLL to that provisions would exceed net charge ‐ offs level (so that quarter). However, if the ALLL is for greater than twice four quarters of future net charge ‐ offs, then provision expense is negative (an 23 down release), to bring the ALLL to that level. ALLL 21 A detailed discussion of how we compute ALLL and provision expense is presented in the Online Appendix. 22 on supervisory guidance suggesting that the ALLL should generally at minimum be sufficient to cover at Based of least of recent charge ‐ offs (Office of the Comptroller four the Currency et al., 2006; Federal Reserve quarters Board, 2013) 23 If necessary, we also adjust the ALLL at the start of the stress test horizon to ensure that the starting value of ALLL is inside this 100% to 250% range. To maintain the accounting identity that assets are equal to the sum of a avoid To equity. common to adjustment corresponding equal involves also this an equity, and liabilities 26

29 Other significant CLASS model assumptions 4.4.3 35% BHCs banks are assumed to pay tax at the and statutory rate. Tax losses may be carried Taxes: for regulatory capital purposes, subject to regulatory forward limits on qualifying deferred tax asset There are limits on the amount of DTA balances. that can be counted as regulatory capital, as (DTA) well on the recognition of DTA relative to future taxable income. The CLASS model includes a as calculation of and qualifying based on regulatory report data DTA the model’s projections of future income, although the calculation is necessarily taxable a simplification due to the complexity of the 24 and regulatory capital rules. accounting and Other Dividends Capital Distributions: As illustrated in Figure 9, changes in equity and regulatory capital over the stress horizon are determined two primary factors: after ‐ tax net income and capital actions such as dividend by both common and preferred shares, share repurchases, payments and new share issuance. The on assumes CLASS that BHCs and model banks do not new shares or make repurchases during the issue stress test horizon, and imposes a stylized rule for determining dividend payments, as illustrated in in the analysis presented sensitivity Section 5. The CLASS model uses a partial adjustment rule for dividends. In the long run, dividends converge to a payout ratio (a given fraction of net The industry payout ratio, computed as the sum income). common averaged of and preferred dividends as a fraction of industry after ‐ tax net income, approximately 50% of net after ‐ tax income prior to the financial crisis. Therefore, our baseline 40 ‐ is that assumption dividends converge to a long ‐ run payout of 45%, following a partial total ratio mechanism: adjustment in equity capital, we treat this adjustment as an addition to provision expense which we apply evenly discontinuity over the scenario horizon. See the Online Appendix for more details. 24 constraints on the available data, we implement some Given simple limits on allowable DTA. First, working with information from the FR Y ‐ 9C reports, we compute net DTA as the maximum of deferred tax assets minus deferred this value and disallowed DTA, then calculate allowable DTA as the difference between tax liabilities, or zero. We 10% which reported directly on the Y ‐ 9C. Any allowed DTA below is of Tier 1 capital is deemed to be dependent on future taxable income. Any excess over 10% of Tier 1 capital is deemed to be recoverable through loss carry ‐ backs. horizon, is held fixed over the stress test while any accumulated tax losses are applied to category latter This allowed DTA dependent on future taxable income at each point in the forecast. If at any point this balance reaches purposes. capital regulatory for forward carried to tax losses will not be able further be capital, 1 Tier of 10% 27

30 * Dividends (  Dividends + (1 ‐  ) [ Dividends ‐ Dividends ] , 0) max = ‐ t 1 t t t ‐ 1 * where Dividends = 45% x after ‐ tax net income and  is the speed ‐ of ‐ adjustment parameter. , t t are also restricted to be Dividends ‐ negative at each point in time. Given observed inertia in non dividends for banking firms (e.g. see Berger et al., 2008), we assume that dividends adjust slowly this target ratio. Our benchmark assumption is to towards set  = 0.90, meaning that ten percent of the gap between current and target dividends is closed each quarter (or 34% after one year). 5. Sensitivity Analysis and Specification Tests modelling illustrates the sensitivity of the CLASS projections to different This assumptions. It section also two external validity presents tests of the projections, comparing them to official DFAST supervisory stress projections, and to bank performance during the 2007 ‐ 09 financial crisis. Sensitivity Analysis 5.1 The CLASS model projections are sensitive to a variety of modeling assumptions needed to link to projections of loss, revenue and expense ratios the model’s ultimate projections of regulatory asset highlights capital. This section the sensitivity of the model’s projections to assumptions about growth, loss provisioning, and capital distributions. These sensitivity results are summarized in loan 10. Figure first panel of Figure 10 presents the results The for the asset growth rate assumption. Recall that the CLASS model assumes asset growth of 1.25% quarter (5% per year). In the figure we per compare our Tier 1 common equity ratio projections under this baseline assumption to projections per under other asset growth rates, ranging from 2.5% three quarter to ‐ 1.25% per quarter. As the figure shows, the path of the projected capital ratio is quantitatively very sensitive to which assumption chosen – after nine quarters, the Tier 1 is common equity ratio is around 13% under a ‐ 1.25% asset growth rate, but only 9% under a 2.5% growth rate. This variation is driven primarily by the mechanical fact that the Tier 1 common ratio is directly expressed as a ratio of risk ‐ weighted assets – high asset growth thus acts to reduce the Tier 1 common ratio, while asset shrinkage their increases it. Asset growth assumptions also affect the numerator of the capital ratio, through 28

31 effect on projected dollars of losses, revenues and expenses. For example a given projected ROA by definition imply a higher dollar value of net income when assets are higher. However, this will direct effect turns out to be less important than the impact of the asset growth numerator on the assumption risk ‐ weighted assets denominator of the capital ratio over 2 ‐ 3 year the 25 over which CLASS model projections are calculated. timeframe B of Figure 10 illustrates how the model projections are affected by the choice of Panel loan loss provisioning rule. We compare our benchmark assumption for provisions (that provision equal net charge ‐ offs as long as the ALLL stays in a “tunnel” between 100% and 250% of the next four quarters projected net charge ‐ offs) to a “four quarter rule” that sets ALLL equal to the next four of projected net charge ‐ offs under the scenario quarters in question, and to a rule that provision of always set equal to net charge ‐ offs. (See expense the Online Appendix for a is further discussion of the differences between these three approaches). Among these three approaches, the “provision expense NCO” rule produces the smallest decline in = the industry capital ratio, because it leaves ALLL constant at its last historical value, rather than revising ALLL upwards in line with the high projected future net chargeoffs as the adverse macroeconomic scenario plays out. buybacks the Finally, we vary the rule used for capital distributions, that is, sum of dividends, share and issuance (panel C of Figure 10). We consider three alternate capital distribution rules: (i) equity remain fixed at their last historical value, (ii) dividends the equal to are benchmark rule dividends by the CLASS model (i.e. dividends adjust gradually towards a payout ratio of 45% of net used to (iii) dividends are set equal income), zero over the entire scenario. Comparing the two and the under the crisis redux scenario, extreme industry Tier scenarios 75 ratio is about 1 common assumption higher under basis “zero dividend” points than under the “constant dividends” the assumption. The rule used by the CLASS model is in between these extremes, although closer to the 25 that As a numerical illustration, consider a firm initially has $100bn in assets and $10bn in capital, and thus has a capital of 10 percent. Assume for simplicity that the firm earns profit net of dividends equal to zero. For ratio this of growth asset cumulative to amounts quarters over nine asset compounded rate growth quarterly 2.5% a firm, 24.9% and resulting total assets of $124.9bn. In contrast, compounded ‐ 1.25% asset growth amounts to cumulative of ‐ 11.0%, and resulting total assets of $89.0bn. growth Since capital after nine quarters is still $10bn, “asset former the than case shrinkage” “asset the in nine higher significantly is quarters after ratio capital the latter. growth” case ‐‐ 11.2 percent of assets in the former compared to 8.0 percent of assets in the 29

32 “constant dividends” assumption, reflecting the model assumption of a slow adjustment speed for dividends. important this exercise illustrates, dividend behavior is As for the path of capital quantitatively a period of stress. This point is relevant to discussions of the 2007 to 2009 financial crisis, during a period when many commentators to argue banking firms were slow that cut dividends in response large losses to (e.g. see Acharya, Gujral and Shin, 2009). The dividend rule machinery within the CLASS model enables a simple evaluation of the quantitative impact of different behavioral rules for capital during a stressful macroeconomic event. distributions 5.2 CLASS and Comparing DFAST projections natural benchmark for the A model is the framework used in the Federal Reserve’s DFAST CLASS and CCAR stress tests. At a conceptual level, the analytical approach in both sets of stress test calculations the same: to project net is income and post ‐ stress regulatory capital ratios as they would occur, quarter ‐ by ‐ quarter, over the stress test scenario horizon, applying U.S. accounting and the regulatory capital rules. However, there are important differences in implementation that affect comparability the results, as summarized in Table 6. of first key difference is that the modeling approach used A more is much CLASS aggregated than in Federal Reserve’s official stress tests. For the most the part, the DFAST and CCAR stress test results loan, are derived from “bottom up” models based on granular risk characteristics of the securities, and trading portfolios, often at the individual borrower, loan or position level. These models use loan detailed data provided by the BHCs describing borrower characteristics, structure, or securities and factors likely to affect the default probability, other exposure at default, and loss given default of the positions. In contrast, the CLASS model uses a “top down” modeling approach based on the historical securities behavior of charge ‐ offs, gains and losses, trading performance, and other revenue expense variables. Although the CLASS models and use firm specific regulatory report data, this information is much less granular than the confidential BHC ‐ specific data used in the CCAR and tests. DFAST stress 30

33 In keeping with this very detailed supervisory approach, the DFAST and CCAR stress tests were on 18 individual large BHCs, and were expanded to a total of 30 BHCs with originally performed greater than $50 billion in 2014. In contrast, the CLASS model quickly generates results for assets the each of largest 200 commercial banking firms (BHCs and independent banks) and the sum of the st which are aggregated into a remaining single 201 institutions proxy BHC. are also differences in some of the There detailed modeling elements that affect both the nature of the loss projections and magnitude of the resulting post ‐ stress capital ratios.  and counterparty losses: the DFAST and CCAR stress tests include an instantaneous Trading global shock on trading and counterparty positions market at the largest BHCs, which is assumed to occur in the first quarter of the stress test horizon. The CLASS model does not is include trading shock specifically, though the trading revenue model this geared to produce the kind of large trading losses that were experienced during the recent financial crisis under a repeat of similar macroeconomic conditions. Even so, the additional global trading market shock included in the DFAST and CCAR stress tests is likely to generate larger and losses at the largest BHCs than the CLASS model. counterparty Balance sheets:  CLASS model includes stylized assumptions about balance sheet growth the do vary across BHCs or across macroeconomic scenarios. In contrast, the CCAR and that not tests include balance sheet DFAST growth paths that vary across both these stress dimensions. As illustrated in Section 5, differences in balance sheet growth can have ‐ significant on the resulting projections of post impacts stress capital ratios, largely due to the impact on projected RWA, the denominator of those ratios. Dividend  and capital distribution assumptions: The CLASS model makes stylized assumptions about stock dividends – linking these to earnings and an assumed long ‐ run payout common and repurchases. This means that the dividends ratio in the CLASS model are sensitive to – individual BHC performance and will change with the macroeconomic scenario; generally, dividends be higher in will good economic environments than in the stressed ones. The stylized DFAST stress test results also make assumptions about dividends and other repurchases distributions; dividends are assumed to be fixed at recent historical levels while 31

34 are set to zero. Thus, distributions of capital to shareholders do not vary across or within 26 in the DFAST stress tests. scenarios macroeconomic Capital Rules:  CCAR and DFAST Regulatory stress tests incorporate RWA projections that The the phase ‐ in of any new capital regulations over the stress test horizon. In contrast, capture model RWA projections implicitly the carry forward the regulatory capital rules in CLASS are the time of the last historical observation, since RWAs place assumed to grow at assets. proportionately with Bearing these differences in mind, Table 7 compares and DFAST projections as of 2013:Q3 for CLASS the 30 firms subject to the DFAST. This is done under the severely adverse macroeconomic scenario specified the Federal Reserve as part of the DFAST and CCAR 2014 stress tests. DFAST projections by from the public are taken reported in Board of Governors of the Federal Reserve information The first three columns of results (2014). examine asset ‐ weighted aggregate projections for System the 30 firms for the change in the capital ratio (Tier 1 common / RWA) and for key components of income. final two columns of results The are the results of a cross ‐ sectional regression comparing the CLASS and DFAST results across firms. highlight two key features of the We comparison. First, CLASS and DFAST projections are is significantly positively correlated for key components of income and loss. This association strongest for PPNR – the regression comparing and projections has a CLASS slope coefficient DFAST 2 unity (0.845) and an close to R of 0.869. The association is also strong for provision expense quite 2 2 (R total assets percentage a as of and for pre ‐ tax return on assets (R = = 0.338). For each of 0.658) two the association between the these sets of projections is positive and statistically categories, the significant 1% level, even with only 30 firms. at The association for the change in the capital 2 the scenario is also ratio positive over although no longer statistically significant (R This = 0.091). less strong relationship, relative to net income, in part reflects the different assumptions for asset DFAST, and other distributions underlying CLASS and dividends as well as the fact that growth, DFAST incorporates some factors during the scenario which do not flow through net income but do 26 Capital distributions in the DFAST stress tests are equal to actual capital distributions in the first quarter of the place stress test horizon (since these distributions have already taken by the time the stress test calculations are horizon. test stress quarter ‐ nine the of quarters level for the remaining 8 a constant at set are and made) being 32

35 affect regulatory capital over the DFAST projection horizon, such as fair value losses on securities firms subject to advanced approaches under Basel II/III. (These are not reflected in the portfolios for projections, which are based on model Basel I accounting only). CLASS CLASS projections are less conservative than DFAST for Second, both PPNR and provision expense, project a significantly smaller decline in industry capitalization than DFAST. This and difference also reflects in methodology and data availability, as well differences as the fact that CLASS does not model some loss and components of loss projected in DFAST – such as the short run shock” “trading value firms with large trading portfolios, and fair unrealized losses on available ‐ for ‐ sale applied to portfolios. It also reflects other modelling differences, securities as the fact that DFAST holds such firm dividends constant under the scenario, while CLASS assumes that payouts adjust slowly to changes in net income according to a partial adjustment mechanism. We interpret the positive correlations between and DFAST projections as encouraging CLASS evidence that CLASS provides a reasonable proxy as to how more detailed stress tests might have performed prior to the financial crisis or if applied to a broader range of firms. not CLASS should necessarily viewed as a good tool for measuring the absolute level of any undercapitalization in be relative system, given its more optimistic projections the to DFAST. However, our banking that interpretation is likely to be useful in is evaluating how capital adequacy under stress CLASS has evolved over time, or how it varies across firms. 5.3 Comparing CLASS to the 2007 ‐ 09 Crisis Experience In similar vein, we compare CLASS projections as of 2007:Q2 to ex ‐ post realized firm performance over during the financial crisis period. We conduct this comparison nine quarters for net income components, and over six for quarters firms capital ratios’ (i.e. the change in the capital ratio from 2007:Q2 2008:Q4). We stop at the end of 2008 for the capital comparison because it is the point to at which industry capitalization was minimized – banking sector ratios increased sharply in capital 2009 as firms recapitalized by issuing equity and cutting dividends, in significant part due to the the 2009 To compute CLASS projections, we seed SCAP. CLASS model with 2007:Q2 banking data and project forward using the actual realized path of macroeconomic and financial market then conditions from 2007:Q3 onwards. We compare the resulting CLASS realized projections to 33

36 accounting data. This is done for each of the 200 largest banking firms in 2007:Q2 that are still in the data six quarters later (164 entities in total). active this comparison are presented Results Table 8, which follows a of similar format to Table 7. As in table shows, the the model projections are quite similar to realized performance for PPNR, net chargeoffs net income. CLASS somewhat under ‐ predicts total industry realized losses ‐‐ and ROA projected by CLASS is 0.13%, compared to a annualized realized value of ‐ 0.05% cumulative industry (note: these values are much lower than annualized both ROA in the period prior to the crisis, which generally ranged between 1% and 1.5%). CLASS also projects a smaller decline in the and industry of Tier 1 common equity to RWA, in part due to the difference in projected ROA, ratio in part due to the fact that net distributions (dividends capital share repurchases and net of issuance) declined in net terms more slowly than the partial adjustment rule embedded in the CLASS model. are cross ‐ sectionally, CLASS projections Looking significantly positively correlated with actual and outcomes during the financial crisis for several key financial ratios: PPNR as a percentage of total assets, for the net chargeoff rate on loans, return on assets, and for the change in the capital 27 ratio. compute total Interestingly, the correlations are also stronger when we them weighted by other assets than on an unweighted basis – in words, CLASS projections are more correlated rather actual realized performance among larger firms. Our interpretation of this finding is that CLASS with picking well in up reasonably performs in risk across different types of assets (e.g., differences loans versus Treasury securities), but is less useful construction in identifying differences in risk within particular asset class, given the lack of a detailed risk information such as geography, credit scores or loan ‐ to ‐ valuation ratios on individual loans. This is likely make CLASS to more effective for firms engaged in a range of activities, rather than smaller firms which may be relatively more types concentrated in particular to of lending or lending in a particular geographic region. Similar of our between CLASS and DFAST, CLASS projections comparison capital are less correlated with 27 As a robustness CLASS test, we repeated the results shown in table 8 using a “point ‐ in ‐ time” version of the model estimated using data only up to 2007:Q2, rather than the full sample. Even under this version of the model, and its ROA major components (e.g., across firms CLASS income projections are significantly positively correlated 2 and 0.094 0.082, for the asset ‐ weighted ROA in the “full sample” and “point ‐ in ‐ time” versions is R the 0.079, and is 0.025 and respectively, respectively in the unweighted case). 34

37 actual realizations for the capital ratio than for the components of net income – reflecting that policy during this period did not always closely correspond to the partial adjustment firms’ capital in the used framework. assumption CLASS we view these results as encouraging evidence that CLASS, while a simplified framework Again, that from many features of bank risk, performs reasonably well as a tool for projecting the abstracts net income evolution and capital under stressful macroeconomic conditions. Future of improve improvements the CLASS framework could further to the fit between model predictions and actual realizations under stress. 6. Summary and Conclusions model is a top ‐ down capital stress testing framework designed to The insights into CLASS provide the stability and capital resiliency of the banking system against stressed economic and U.S. financial market conditions. The CLASS model is based on simple econometric models and publicly available data, rather than the more regulatory detailed confidential data that underpins the DFAST and CCAR supervisory stress tests. One advantage of this approach is that model projections can be to generated quickly, making the CLASS framework amenable conducting a range of “what if” as analyses. For example, by adjusting key assumptions in the model – such those the governing used asset growth or the amount and timing of capital distributions – the model can rate be of to assess how the banking industry capital might under different circumstances, as well as change provide some insight into how these assumptions might affect the more detailed, firm ‐ specific stress benchmark test results generated by supervisors and banks. The model is also useful as a Covas, framework which other top ‐ down models (e.g., Rump and Zakrajsek, 2013; Kapinos and Mitnik, 2015) et can be compared against. For example, Covas al. adapt many features of the CLASS framework, use a but of approach, rather than OLS, in modelling quantile effect regression the macroeconomic conditions on banking system income and capital. The CLASS model projections suggest that the vulnerability of the U.S. banking industry to under ‐ significantly capitalization stressed economic conditions has declined in since the financial crisis of 2007 to 2009. This is consistent with the increases result in regulatory capital ratios that have show occurred since this period. What is perhaps less obvious is that CLASS model projections 35

38 increasing capital vulnerability starting as far back as 2004, well before either regulatory capital in market indicators suggested a capital shortfall the industry. Although our baseline ratios or are based on the CLASS model estimated on data incorporating the financial crisis projections this rising vulnerability is still period observed if itself, we the same models estimate based only on available at the time. data CLASS the future, In we plan to further refine the model to account for the risk sensitivities of individual banks and BHCs. For instance, loan loss rate projections might be better tailored to individual by including institutions performing ‐ information about non ‐ loans to firm specific lagged net charge ‐ off rates in “seeding” the projections. The supplement models for projecting PPNR could be made more granular by disaggregating the various PPNR sub ‐ components further (e.g., separately projecting interest income on loans and interest expenses on deposits). Another avenue for approaches future development is to explore different to projecting the balance sheet, some of which would allow individual balance sheet components to grow at different rates in loans different scenarios (e.g., to capture shifts between and securities over the business cycle). additional It would also be of interest to conduct the out sample testing of ‐ CLASS framework. of ‐ and Welch (2012) present cautionary evidence that for “top ‐ down” models of the type Guerrieri modestly this paper, macroeconomic variables only in improve out ‐ of ‐ sample forecasting estimated in That said, we find section 5.3 of this paper that power. CLASS projections are positively bank performance during the financial correlated crisis with period, even when the model is only using pre ‐ crisis data. It would be interesting, although outside the scope of the estimated present paper, to also test whether CLASS projections are also correlated with bank financial distress the or failure during crisis. said, a difficult challenge for testing CLASS out ‐ of ‐ sample is That significant sample period contains only one period of that macroeconomic and banking system our distress. We have no data (yet) to tell us whether models estimated using a sample period including ‐ the 2007 09 crisis perform well in banking sector performance during the next crisis or projecting severe recession. This is a general problem for stress testing models – the goal of such models is to observed. project losses in the tails of the distribution, which by definition are rarely 36

39 Several other avenues for model development also promising. One would be to streamline seem the model in a way that would allow us to run many scenarios very quickly and thus to take a the statistical approach to determining the underlying vulnerabilities of banking system (e.g., to explore the characteristics of distribution, generate scenarios that capital declines in the tail of the see what these scenarios have in common). to Another would be to integrate liquidity stress into framework the model, for example using a is like Eisenbach et al. (2014). In short, the CLASS model framework living a that is expected time. to evolve and develop over 37

40 References Irvind Gujral and Hyun Song Shin. 2009. “Dividends and Bank Capital in the Financial Acharya, Viral, 2007 ‐ 09.” Working of New York University. Crisis paper, Acharya, Viral, Robert Engle and Dianne Pierret. 2013. “Testing Macroprudential Stress Tests: The Risk of Regulatory Risk Weights.” Journal of Monetary Economics , 65, 36 ‐ 53. Acharya, Viral, Robert Engle and Matthew Richardson. 2012. “Capital Shortfall: A New Approach to Ranking Regulating Systemic Risks.” American Economic Review , 102(3), 59 ‐ 64. and Berger, and Christa Bouwman. 2013. “How Does Allen Bank Capital Affect Bank Performance During Financial Crises?” Journal of Financial Economics , 109(1), 146 ‐ 176. Do Berger, Allen, Robert DeYoung, Mark Flannery, David Lee and Ozde Oztekin. 2008. “How Large Banking Organizations Manage Their Capital Ratios?” Journal of Financial Services Research , ‐ 149. 34, 123 Board Governors of the Federal Reserve System. 2009a. “Supervisory of Capital Assessment and Implementation.” Program: April Design 24, 2009. http://www.federalreserve.gov/n ewsevents/press/bcreg/20090424a.htm _________. 2009b. “Supervisory Capital Assessment Program: Overview of Results.” May 7, 2009. http://www.federalreserve.gov/n ewsevents/press/bcreg/20090507a.htm “Capital Plan Rule.” Federal Register _________. Volume 2011. 76, Number 231, December 1, 2011, pp. 74631 ‐ 74648. Review _________. 2012. “Comprehensive Capital Analysis and 2012: Methodology and Results for Stress Scenario Projections.” March 13, 3012. http://www.federalreserve .gov/bankinforeg/stress tests ‐ capital ‐ planning.htm ‐ “Dodd ‐ Frank Act Stress Test 2013: Supervisory _________. Stress Test 2013a. Methodology and Results. March 2013. http://www.federalreserve .gov/bankinforeg/stress tests ‐ capital ‐ ‐ planning.htm ________. 2013b. “Comprehensive Capital Analysis and Review 2013: Assessment Framework and ‐ Results. 2013. http://www.federalreserve .gov/bankinforeg/stress March tests ‐ capital ‐ planning.htm ________. 2014. “Dodd ‐ Frank Act Stress Test 2014: Supervisory Stress Test Methodology and Results.” March 2014. http://www.federalreserve.gov/new sevents/press/bcreg/bcreg20140320a1.pdf 38

41 Bookstaber, Richard, Jill Cetina, Greg Feldberg, Mark Flood, and Paul Glasserman. 2013. “Stress Office Promote Financial Stability: Assessing Progress and Looking to the Future.” of Tests to Working Paper #0010, Research July 18, 2013. Financial Charles and Doron Nissim. 2012. “Crisis ‐ related Shifts Calomiris, the Market Valuation of Banking in Activities.” 23(3), Journal Financial Intermediation , of 400 ‐ 435. Timothy, Astrid Dick, Beverly Clark, Hirtle, Kevin J. Stiroh and Robard Williams. 2007. “The Role of Retail Banking in the U.S. Banking Industry: Risk, Return and Industry Structure.” Federal Reserve of New York Economic Policy Review . 13:3. Bank European Committee Banking on the “Aggregate Supervision. of 2010. 2010 EU wide Outcome Stress Test Exercise Coordinated by CEBS in Cooperation with the ECB.” 23 July 2010. wide http://www.eba.europa.eu/risk ‐ and ‐ data/eu ‐ analysis ‐ stress ‐ testing ‐ Covas, Francisco, Ben Rump and Egon Zakrajsek. 2013. “Stress ‐ Testing U.S. Bank Holding International Companies: A Dynamic Panel Quantile Regression Approach.” Journal of Forecasting , 30(3), 691 ‐ 713. Subprime Eichengreen, Barry, Ashoka Mody, Milan Nedeljkovic and Lucio Sarno. 2012. “How the Crisis Global: Evidence from Bank Credit Default Swap Spreads.” Went Journal of Money and Finance , 31(5), pp. 1299 ‐ 1318. International Todd Keister, James McAndrews and Eisenbach, Tanju Yorulmazer. 2014. “Stability of Thomas, Funding Models: An Analytical Framework.” Federal Reserve Bank of New York Economic Policy , 20:1, 29 ‐ 47. Review 2014. Engle, Conditional Beta.” Robert. Working paper, NYU Stern School “Dynamic of Business. European Banking Authority. 2011. “2011 EU ‐ Wide Stress Test Aggregate Report.” July 2011. testing http://www.eba.europa.eu/risk ‐ analysis ‐ and ‐ data/eu ‐ wide ‐ stress ‐ Frame, Scott, Kristopher Gerardi and Paul Willen. 2015. “The Failure of Supervisory Stress W. Testing: Mae, Freddie Mac, and OFHEO.” FRB Fannie Atlanta Working Paper No. 2015 ‐ 3. Greenlaw, David, Anil K. Kashyap, Kermit Schoenholtz, and Hyun Song Shin. 2012. “Stressed Out: Macroprudential for Stress Testing.” Chicago Booth Working Principles Paper 12 ‐ 08, No. January 2012. Guerrieri, Luca and Michelle Welch. 2012. “Can Macro Variables Used in Stress Testing Forecast the Performance of Banks?” Federal Reserve Board Finance and Economics Discussion Paper ‐ 49. 2012 39

42 Kapadia, Sujit, Mathias Drehmann, John Elliott, and Gabriel Sterne. 2012. “Liquidity Risk, Cash Flow Constraints, and Systemic Feedbacks.” Bank of England Working Paper #456. June 21, 2012. Kapinos, Pavel S. and Oscar A. Mitnik. 2015. “A Top ‐ Down Approach to Stress ‐ Testing Banks.” FDIC Center Financial Research Paper for No. 2015 ‐ 02. Kau, James, B., Donald C. Keenan, Walter J. Muller III and James F. Epperson. 1992. “A Generalized and Valuation Model for Fixed ‐ Rate Residential Mortgages.” Journal of Money, Credit 3, Banking Vol. 24, No. , pp. 279 ‐ 299. Office of the Comptroller of the Currency, Board of Governors of the Federal Reserve System, Federal Deposit Insurance Corporation, National Credit Union Administration, and Office of “Interagency Thrift Supervision. 2006. Policy Statement on the Allowance for Loan and Lease Losses.” December 2006. Wong, Eric and Cho ‐ Hoi Hui. 2009. “A Liquidity Risk Stress ‐ Testing Framework with Interactions Between Market and Credit Risks.” 06/2009. Hong Kong Monetary Authority Working Paper 40

43 Estimated econometric models Appendix: Appendix Table 1: Components and PPNR Securities Specifications Net Interest Noninterest Fixed Asset Other Return on Compensation Return on Nontrading Margin Nonint. Nonint. Exp. Nonint. Exp. Trading AFS Income Securities Ratio Exp. Ratio Assets Ratio Ratio Macroeconomic variables Annualized Real GDP growth (%) 0.000552 (0.000665) Term Spread (10 year minus 3 0.0426*** months, pct. pt) (0.0139) 0.0220** 3 Month Treasury Yield (%) (0.0106) Quarterly change in 10 year Treasury -0.580*** yield (pct. pt) (0.161) 0.00345*** Stock Market returns (quarterly, %) 0.00407* (0.00245) (0.000886) Quarterly change in BBB bond -0.671 0.179* spread (pct. pt) (0.452) (0.0939) Quarterly change in BBB Spread if -2.559*** change is positive (else zero) (0.588) -0.0310*** Quarterly change in BBB Spread if change is positive x Risky AFS Ratio (0.0030) Time-series controls Lagged dependent variable 0.793*** 0.284 0.894*** 0.853*** 0.816*** 0.128 0.904*** 0175) (0.0221) (0.0340) (0.0951) (0.0390) (0.0143) (0.181) (0. Time trend (annual, 1991:Q1 = 0) -0.00528* -0.00186*** (0.00317) (0.000348) Balance sheet ratios (as % of interest earning assets) Residential Real Estate Loans 0.00476*** -0.00155 -0.000722 -0.000321 -0.00227 (0.00141) (0.00185) (0.000918) (0.000207) (0.00211) Commercial Real Estate Loans 0.00648*** -0.00364** -0.00109* -0.000328 -0.000938 (0.00162) (0.00174) (0.000647) (0.000201) (0.00136) Commercial and Industrial Loans 0.00685*** -0.000877 -0.000229 -0.000470* -0.00171 (0.00134) (0.00189) (0.00147) (0.000252) (0.00202) Credit Card Loans 0.0184*** 0.00990*** -0.00115 -0.000554*** 0.0153*** (0.00369) (0.00245) (0.00105) (0.000170) (0.00337) -0.00626*** -0.00252** -0.00129*** -0.000807 Trading Assets -0.00146 (0.00110) (0.000459) (0.00201) (0.00161) (0.00223) 0.00393*** 0.00309 -0.00172* -0.000853*** 0.00886*** Securities Ratio (0.00118) (0.00201) (0.00103) (0.000244) (0.00241) Other Asset Share (firm assets as % of 0.00743*** -0.00581*** -2.07e-05 -7.56e-05 -0.00369** (0.00141) (0.00181) (0.000840) (0.000159) (0.00175) industry assets) Constant Term 0.234* 0.233* 1.989*** 0.261*** 0.130*** 0.148 0.272*** (0.124) (0.127) (0.602) (0. 0964) (0.0253) (0.111) (0.0535) Observations 17,565 17,565 67 17,565 17,565 17,565 12,875 2 R 0.885 0.876 0.449 0.828 0.835 0.772 0.0352

44 and Industrial Co mmercial 0.117* * * 0.558* * * 90 Other Loans 0.607 0.152* * * (0.0463) (0.0407) (0.131) 0.798* ** (0.0338) 0.133* ** 0.164* ** 0.820 90 (0.0680) (0.0488) 0.0870 0.574* * * 90 Foreign Governments 0.360 0.145 (0.156) (0.243) (0.167) Nonfarm Nonresidential 0.823** * 0.0395* 0.797 -0.00928* ** (0.00343) (0.0990) (0.0229) 90 0.0297* * 0.597* * * 90 Agriculture 0.440 0.0834* * * (0.0212) (0.0120) (0.136) 0.776* * * 0.0480* * 0.765 (0.00467) -0.0114** (0.0218) (0.105) 90 All other loans Commercial real estate 0.0510 0.351* * 90 Ins titutions Depo s ito ry 0.133* * * 0.158 (0.0446) (0.0411) (0.137) 0.801* ** 0.113* Construction Multifamily (0.0222) -0.0473* * 0.878 (0.0657) (0.0887) 90 0.573* * * 90 0.191* * * Estate Other Real 0.555 -0.00365 (0.00534) (0.0150) -0.00933* (0.0681) (0.149) HELOC 0.893* ** 0.0528* -0.00492 (0.0501) (0.0287) (0.00330) (0.00831) 0.955 -0.0284** * 90 0.102* * * 0.635* * * 90 0.264* * * 0.616 Le a s e s (0.0934) (0.0782) (0.0219) Junior Lien Re s id e n t ia l 0.867** * 0.176* 0.911 -0.0671* * * (0.0847) (0.0994) (0.0212) (0.0109) -0.0153 90 0.150* * * 0.701* * * 0.0191* * 90 0.264* * * 0.825 Co n s u mer Other (0.00862) (0.0934) (0.0993) (0.0306) Residential real estate Consumer loans Firs t Lien Residential 0.884* * * 0.0231 (0.0776) (0.0168) (0.00200) -0.0192** (0.00756) 0.917 -0.00147 90 0.359* * * 0.856* * * 90 0.721* * * 0.899 Cred it Card (0.221) (0.0795) (0.0477) 2 2 R Ob s erv atio n s Lag g ed d ep end ent v ariab le Ho me p rice g ro wth (%, y ear-ov er-y ear) Annualized change in Unemployment (%) Con s tan t Appendix Table 2: NCO Specifications Panel A. Real Es tate and Commercial Loans Commercial Property Price Growth if Home price growth if growth is negative (els e zero) negative (else zero) Commercial Property Price Growth if negative over-year) (els e zero) Commercial property price growth (%, year- Panel B. Consumer and all other loans Ob s erv atio n s R Co n s tan t A n n u alized ch an g e in Un emp lo y men t (%) Time tren d (an n u al) Lag g ed d ep en d en t v ariab le

45 1: Figure Model Structure CLASS Assumptions about growth in asset, Macroeconomic scenario liability balances Forecast BHC & bank Substitute into regression models Forecasts for net income and capital key revenue Predict key revenue, loss ratios (e.g. (Tier-1 common / RWA). ratios, NCO rates Sum up across firms to NIM, NCO rates etc.) for each firm as [firm by firm] compute system function of lagged values + macro data estimates Current regulatory data for each firm Other auxiliary assumptions (e.g. provisioning rule, (e.g. current NCO rates, revenues, dividends, goodwill etc.) expenses etc.)

46 performance 2: CLASS projections of PPNR and loan Figure Net Interest Margin on trading assets Return net interest income, % interest ‐ earning assets, annualized trading income, % trading assets, annualized 4.5 10 4 5 3.5 0 3 -5 -10 2.5 1990q1 2015q1 2005q1 2000q1 1995q1 1990q1 2010q1 1995q1 2000q1 2005q1 2010q1 2015q1 historical crisis redux historical base crisis redux base ratio expense interest ‐ Non ratio income interest ‐ non trading ‐ Non Noninterest annualized % expense, total assets of ‐ interest income, % total assets, non ‐ trading non 4.4 3 4.2 2.8 4 2.6 3.8 2.4 3.6 2.2 3.4 2 3.2 1.8 3 1.6 2000q1 1995q1 1990q1 1995q1 2010q1 2005q1 2010q1 2015q1 2015q1 2005q1 2000q1 1990q1 historical crisis redux base historical crisis redux base ‐ provision net revenue ratio Net charge ‐ off rate Pre total total of % NCOs, annualized assets, loans, % PPNR, annualized 3 3.5 2.5 3 2 2.5 2 1.5 1.5 1 1 .5 .5 0 2015q1 2010q1 2005q1 2000q1 1995q1 1990q1 date 2015q1 1990q1 1995q1 2000q1 2005q1 2010q1 historical crisis redux base crisis redux historical base 44

47 of Figure 3: Return on assets ( Annualized after ‐ tax net income, % assets) total 1.5 1 .5 0 -.5 -1 -1.5 -2 2015q1 2005q1 2010q1 1990q1 1995q1 2000q1 crisis redux base historical

48 4: Capital projections: Tier 1 Figure common equity (percent of RWA) firms A. Industry aggregate B. Distribution of capital across 13 100 12 80 11 60 10 9 40 % industry assets 8 20 7 6 0 2000q1 2005q1 2010q1 2015q1 15 5 10 base crisis redux base crisis redux historical 46

49 “gap” Figure 5: Evolution of industry capital 500 400 300 200 100 Capital Gap ($bn), 5% and 8% Thresholds 0 2012q1 2010q1 2008q1 2002q1 2004q1 2006q1 8% threshold 5% threshold 47

50 6: Point ‐ in ‐ time and full Figure sample industry capital “gap” 800 600 400 Capital Gap ($bn), 8% Threshold 200 0 2004q1 2002q1 2008q1 2006q1 2010q1 2012q1 Full sample Point-in-time of measures capital Comparing 7: vulnerability Figure average value in 2002. Each measure is normalized by its 800 600 400 % of average 2002 value 200 0 2002q1 2010q1 2008q1 2004q1 2006q1 2012q1 CLASS Capital Gap (8% threshold) SRISK: GMES SRISK: MESSIM Mean CDS Spread 48

51 Figure 8: Time series evolution of correlation between capital ratio and change in capital ratio under stress scenario point on the line represents the Each estimate from a cross ‐ sectional regression of point redux capital starting capital ratio against the projected change in the ratio under the crisis value indicates that firms with capital negative ratios that decline sharply under the A scenario. stress scenario also have higher starting capital ratios. A positive value indicates the reverse. .3 .2 .1 0 Slope Coefficient -.1 -.2 2014q1 2005q1 2002q1 2008q1 2011q1 49

52 Capital and Income Net of Computation Schematic, 9: Figure

53 assumptions Figure 10: Sensitivity to Asset A. rate growth 12 11 10 9 8 Tier 1 Common Ratio 7 6 2015q1 2010q1 2005q1 2000q1 Historical 0.00% 1.25% -1.25% 2.50% B. Provisioning assumption 12 11 10 9 8 Tier 1 Common Ratio 7 6 2005q1 2000q1 2015q1 2010q1 Historical Provision expense = NCO Benchmark CLASS rule ALLL = four quarter rule C. Payout rule 12 11 10 9 8 Tier 1 Common Ratio 7 6 2005q1 2015q1 2010q1 2000q1 Zero Dividends Historical Benchmark CLASS rule, (d=0.9) No Change in Dividends 51

54 macroeconomic Table 1: Summary of scenarios variables Selected scenarios: Macro Historical Redux Crisis Baseline 3 3Q Middle First Q3 2013 Q Last 3Q First 3Q Last 3 3Q Middle Q 12.40 rate (end) Unemployment 7.00 6.70 6.30 7.80 9.70 7.30 2.89 2.59 1.86 ann) GDP (%, growth (1.54) (2.87) 0.47 2.89 Equity prices (% 19.44 (0.70) 4.00 4.08 (12.39) (31.82) 19.39 ch) Home price growth (% ch, ann) 10.90 2.52 2.64 3.07 (15.40) (21.73) (11.74) are based on the supervisory scenarios data posted and Note: The historical data baseline scenario reported here ‐ ‐ capital ‐ tests .gov/bankinforeg/stress http://www.federalreserve (see 2013 by 1 November on Reserve Federal the any reflect not do They planning.htm). revisions. data subsequent 52

55 in 2: Determinants of change Table capital during stress scenario Dependent minus (projected scenario variable: Change in Tier 1 common equity ratio during crisis redux historical) (2) (4) (3) (1) (5) (6) (7) Characteristics value) Firm (last historical Tier 0.028 Ratio Common 0.033 0.039 1 0.025 (0.034) (0.035) (0.034) (0.032) ‐ 0.152*** ‐ Ratio ‐ 1 Tier 0.160*** 0.182*** ‐ 0.127** Common (0.051) (0.048) (0.043) x Post 2011Q1 (0.050) 0.041** Liquidity Ratio 0.006 0.004 0.004 (0.014) (0.015) (0.015) (0.016) Liquidity 0.029** Ratio x ‐ 0.026*** ‐ 0.016 ‐ 0.003 Post (0.014) 2011Q1 (0.010) (0.010) (0.012) RWA / total assets 1.742 4.749*** (1.135) (1.081) 2.536** / total RWA assets x 1.711** Post 2011Q1 (0.772) (1.284) 0.110** ln(total assets) ‐ 0.025 ‐ 0.010 (0.050) (0.081) (0.075) x assets) 0.133** ‐ ln(total ‐ 0.216*** ‐ 0.214*** (0.059) Post 2011Q1 (0.064) (0.082) 9,398 9,398 9,398 9,398 9,400 9,400 Observations 9,400 2 R 0.193 0.122 0.118 0.123 0.14 4 0.118 0.131 Yes Time FEs Yes Yes Yes Yes Yes Yes the during ratio equity Note: Pooled regression is based on each historical quarter's projections of the change in the Tier 1 common crisis redux scenario and firm characteristics as of that historical quarter. Variables are winsorized at their 1% and 99% values, to share. asset by limit the influence of outliers. weighted Clustered on Entity. Observations are 53

56 SD Historical 2.95 11.76 Mean Historical 9.85 3.13 7.55 1.98 2.51 2.49 5.49 1.92 8.50 7.61 2.70 2.09 1.16 0.70 -0.06 0.38 2.20 1.72 0.92 0.10 0.00 0.42 0.00 3.00 2.06 -1.20 0.05 1.20 Value 2013:Q3 ௧ ௧ିଵ ௧ ሿൈ4 ௧ିଵ ௧ିଵ ௧ ൈ100 ൈ100 ൈ400 ൈ100 ሿ ሿ ሿ ሻ ሿ ሻ ሻ ሻ ௧ିସ ௧ିଵ ௧ିଵ ௧ିସ ݈ܻ݀݁݅ݕݎݑݏܽ݁ݎܶ.݋݉ 3 െ ݈ܻ݀݁݅ݕݎݑݏܽ݁ݎܶ.ݎݕ െ10 ݀ܽ݁ݎ݌ݏܤܤܤ െ ௧ ௧ ௧ ݈ܻ݀݁݅ݕݎݑݏܽ݁ݎܶ.ݎݕ 10 െ ݐ݊݁݉ݕ݋݈݌ܷ݉݁݊ %െ ௧ ௧ ܫܲܲܥሺ݈݊ሻെ ܲܦܩሺ݈݊ሻെ ܶܭܯሺ݈݊ሻെ ܫܲܪሺ݈݊ሻെ ௧ ௧ ௧ ௧ ݈ܻ݀݁݅ݕݎݑݏܽ݁ݎ݄ܶݐ݊݋݉3 ܫܲܪሺ݈݊ ܲܦܩሺ݈݊ ሾ ܶܭܯሺ݈݊ ݀ܽ݁ݎ݌ݏܤܤܤ ܫܲܲܥሺ݈݊ ሾ ሾ ሾ 54 ݈ܻ݀݁݅݀݊݋ܤܤܤܤ ݈ܻ݀݁݅ݕݎݑݏܽ݁ݎܶ.ݎݕ10 ݐ݊݁݉ݕ݋݈݌ܷ݉݁݊%ሾ ݈ܻ݀݁݅ݕݎݑݏܽ݁ݎܶ.ݎݕ10 Year Treasury Yield in Percent Variable Definition Annual Commercial Property Price Index (log change) Annualized Change in the Civilian Unemployment Rate (%) Annualized Real GDP growth (%) Annual House Price Index (log change) Quarterly change in BBB bond spread (pct. pt) Quarterly change in 10 year Treasury yield (pct. pt) Table 3: Summary Statistics of Macroeconomic Variables Term Spread (10 year minus 3 months, pct. pt) Quarterly growth in Stock market returns (%, log change) Spread of BBB Bond Index to 10 3 Month Treasury Yield (%)

57 SD Historical Mean Historical 0.27 0.17 0.47 2.61 3.57 0.51 1.54 1.70 0.13 0.30 0.45 0.09 2.44 2.92 2.59 1.94 2.28 0.39 1.45 1.55 0.20 Value 2013:Q3 ൈ400 ൈ400 ൈ400 ൈ400 ൈ400 ൈ400 ൈ400 ݏݐ݁ݏݏ ܣ ݏ݁݅ݐ݅ݎݑܿ݁ܵܵܨ ܣ ݏݐ݁ݏݏ 55 ܣ ݏݐ݁ݏݏ ݏݐ݁ݏݏ ݏݐ݁ݏݏ ݏݐ݁ݏݏ ܣ ܣ ܣ ܣ ݁ݏ݊݁݌ݔܧݐ݁ݏݏ ܣ ݈ܽݐ݋ܶ ݈ܽݐ݋ܶ ݈ܽݐ݋ܶ ݈ܽݐ݋ܶ ݃݊݅݀ܽݎܶ ݁݉݋ܿ݊ܫ݃݊݅݀ܽݎܶ ݁݉݋ܿ݊ܫݐݏ݁ݎ݁ݐ݊ܫݐ݁ܰ ݏ݁݅ݐ݅ݎݑ݈ܿ݁ܵ݁ܽܵݎ݋ܨ݈ܾ݈݁ܽ݅ܽݒ ݀݁ݔ݅ܨ ܣ ݁ݏ݊݁݌ݔܧ݊݋݅ݐܽݏ݊݁݌݉݋ܥ ݃݊݅݊ݎܽܧݐݏ݁ݎ݁ݐ݊ܫ ݈ܽݐ݋ܶ ݊݋ݏ݊݅ܽܩݐ݁ܰ݀݁ݖ݈ܴ݅ܽ݁ ݁݉݋ܿ݊ܫ݃݊݅݀ܽݎܶെ݁݉݋ܿ݊ܫݐݏ݁ݎ݁ݐ݊݅݊݋ܰ urns (Annualized, in Percentage Points) ݁ݏ݊݁݌ݔܧݐݏ݁ݎ݁ݐ݊݅݊݋ܰݎ݄݁ݐܱ൅.ݎ݅ܽ݌݉ܫ݈݈݅ݓ݀݋݋ܩ൅.ݎ݅ܽ݌݉ܫ݊݋݅ݐܽݖ݅ݐݎ݋݉ ܣ Variable Definition Net Interest Margin Noninterest Nontrading Income Ratio Table 4: Accounting Ratios Modeled in CLASS Panel A: Components of PPNR and AFS ret Fixed Asset Noninterest Expense Ratio Other Noninterest Expense Ratio Compensation Noninterest Expense Ratio Return on Trading Assets Return on AFS Securities

58 (NCO) Rates in Percentage Points Panel B: Annualized Net Charge Off Historical 2013:Q3 Historical Variable Definition Value Mean SD ݏ݊ܽ݋ܮܧܴܴ݊݁݅ܮݐݏݎ݅ܨ݊݋ݏܱܥܰ ൈ400 0.36 0.42 0.56 First Lien Residential Real Estate ݏ݊ܽ݋ܮܧܴܴ݊݁݅ܮݐݏݎ݅ܨ ܬ ݏ݊ܽ݋ܮܧܴܴ݊݁݅ܮݎ݋݅݊ݑ ݊݋ݏܱܥܰ ൈ400 2.38 1.66 2.22 Junior Lien Residential Real Estate ݏ݊ܽ݋ܮܧܴܴ݊݁݅ܮݎ݋݅݊ݑ ܬ ݏ݊ܽ݋ܮܧܴܴܥܱܮܧܪ݊݋ݏܱܥܰ ൈ400 HELOC Residential Real Estate 0.93 0.72 0.96 ݏ݊ܽ݋ܮܧܴܴܥܱܮܧܪ ݏ݊ܽ݋ܮܧܴܥ݊݋݅ݐܿݑݎݐݏ݊݋ܥ݊݋ ݏܱܥܰ Construction Commercial Real ൈ400 0.27 1.42 2.01 ݏ݊ܽ݋ܮܧܴܥ݊݋݅ݐܿݑݎݐݏ݊݋ܥ Estate ݏ݊ܽ݋ܮܧܴܥݕ݈݂݅݉ܽ݅ݐ݈ݑܯ݊݋ ݏܱܥܰ Multifamily Commercial Real ൈ400 0.09 0.42 0.57 ݏ݊ܽ݋ܮܧܴܥݕ݈݂݅݉ܽ݅ݐ݈ݑܯ Estate ݏ݊ܽ݋ܮܧܴܥܴܰܨܰ݊݋ݏܱܥܰ Non-Farm Non-Residential ൈ400 0.21 0.41 0.50 ݏ݊ܽ݋ܮܧܴܥܴܰܨܰ Commercial Real Estate ݏ݊ܽ݋ܮ݀ݎܽܥݐ݅݀݁ݎܥ݊݋ݏܱܥܰ ൈ400 3.27 5.15 1.77 Credit Card ݏ݊ܽ݋ܮ݀ݎܽܥݐ݅݀݁ݎܥ ݏ݊ܽ݋ܮݎ݁݉ݑݏ݊݋ܥݎ݄݁ݐܱ݊݋ ݏܱܥܰ ൈ400 Other Consumer 1.02 1.62 0.86 ݏ݊ܽ݋ܮݎ݁݉ݑݏ݊݋ܥݎ݄݁ݐܱ ݏ݊ܽ݋ܮܫ&ܥ݊݋ݏܱܥܰ ൈ400 Commercial and Industrial (C&I) 0.27 0.87 0.67 ݏ݊ܽ݋ܮܫ&ܥ ݏ݁ݏܽ݁ܮ݊݋ݏܱܥܰ ൈ400 Leases 0.15 0.50 0.39 ݏ݁ݏܽ݁ܮ ݏ݊ܽ݋ܮ݁ݐܽݐݏܧ݈ܴܽ݁ݎ݄݁ݐܱ݊݋ ݏܱܥܰ ൈ400 0.48 0.43 0.55 Other Real Estate ݏ݊ܽ݋ܮ݁ݐܽݐݏܧ݈ܴܽ݁ݎ݄݁ݐܱ ݏݐ′ݒ݋ܩ݊݃݅݁ݎ݋ܨ݋ݐݏ݊ܽ݋ܮ݊݋ ݏܱܥܰ ൈ400 Loans to Foreign Governments 0.04 0.61 3.73 ݏݐ′ݒ݋ܩ݊݃݅݁ݎ݋ܨ݋ݐݏ݊ܽ݋ܮ ܣ ݏ݊ܽ݋ܮ݁ݎݑݐ݈ݑܿ݅ݎ݃ ݊݋ݏܱܥܰ ൈ400 0.06 0.21 0.19 Agriculture ܣ ݏ݊ܽ݋ܮ݁ݎݑݐ݈ݑܿ݅ݎ݃ .ݐݏ݊ܫݕݎ݋ݐ݅ݏ݋݌݁ܦ݋ݐݏ݊ܽ݋ܮ݊݋ ݏܱܥܰ ൈ400 Loans to Depository Institutions -0.04 0.21 0.47 ݏ݊݋݅ݐݑݐ݅ݐݏ݊ܫݕݎ݋ݐ݅ݏ݋݌݁ܦ݋ݐ ݏ݊ܽ݋ܮ ݏ݊ܽ݋ܮݎ݄݁ݐܱ݊݋ݏܱܥܰ ൈ400 0.20 0.36 0.40 Other ݏ݊ܽ݋ܮݎ݄݁ݐܱ 56

59 Firm balance sheet controls X X X X X Time trend (annual) Controls negative (else zero) Commercial Property Price Growth if year-over-year) Commercial property price growth (%, negative (else zero) Home price growth if growth is Home price growth (%, year-over-year) (%) Annualized change in Unemployment change is positive x Risky AFS Ratio Quarterly change in BBB Spread if change is positive (else zero) Quarterly change in BBB Spread if Expl anatory Vari ables (pct. pt) Quarterly change in BBB bond spread 57 Macroeconomic Variables Stock Market returns (quarterly, %) yield (pct. pt) Quarterly change in 10 year Treasury 3 Month Treasury Yield (%) pct. pt) Term Spread (10 year minus 3 months, Annualized Real GDP growth (%) Specifications Included and signficant (10% level) Included but insignificant Firm controls included Model Net Interes t Margin Other Real Estate Compensation Noninterest Expense Ratio Return on Trading Assets Other Noninterest Expense Ratio Noninterest Nontrading Income Ratio Return on A FS Securities First Lien Residential Junior Lien Residential HELOC Construction M u lt ifamily Nonfarm Nonresidential Cred it Card Other Consumer Commercial and Industrial Leases Fixed Asset Noninterest Expense Ratio Depos itory Ins titutions Agriculture Foreign Governments Other Loans 5: PPNR Specs Net Charge Off Specs X Legend Modeled Variables Table

60 Table Comparison and Differences Between Stress Test Frameworks 6: CLASS Model DFAST/CCAR Top down models based ‐ on Approach Modeling focused models the up ‐ on Bottom net aggregated outcomes (e.g., individual of risk characteristics offs) for broad income charge ‐ loans, securities, and trading categories securities and loan and positions securities portfolios Detailed available balance sheet information supervisory Publicly Data individual statement income and from regulatory BHCs, often at the data from Call and Y ‐ 9C or report level of individual loans filings securities largest BHCs and Coverage The 200 30 BHCs with assets exceeding banks, plus the rest independent $50 billion (starting in 2014). the industry. Results of Results reported in the aggregate reported at the aggregate industry and at the individual BHC level level revenue global Trading and Counterparty Trading modeled based instantaneous Separate the on market shock on the trading and macroeconomic scenario counterparty the 6 of positions BHCs largest Stylized assumptions that result in Dividends For assumptions DFAST, stylized run dividends converging to a long ‐ at that hold dividends fixed average payout ratio relative to recent historical levels and income net no repurchases assume Sheet Growth Stylized assumption for all Balance Varies across institutions and scenarios scenarios institutions in all across Weighted Assets Changes proportionately with the Risk Changes with the macroeconomic implicitly balance sheet, ‐ carrying scenario, incorporating the phase regulatory capital regulatory new forward prevailing in of any capital rules rules Capital Model Captures key elements, but Regulatory More detailed and precise capital regulatory involves approximations of certain calculations of calculations complex 58

61 and 7: Comparison between CLASS DFAST projections Table under the DFAST severely adverse macroeconomic scenario based on data as of 2013:Q3 for the 30 Projections to the 2014 CCAR and DFAST supervisory stress tests. subject firms firms across DFAST: vs CLASS α = (CLASS ε + ) DFAST . β + 2 Slope Difference DFAST CLASS Category coefficient ( β )R Income PPNR/Assets 0.869 0.845*** 0.39 1.57 1.97 a Expense/Assets 2.88 Provision 0.658 0.729*** 0.89 ‐ 1.99 ‐ 0.02 ‐ 0.24 Other/Assets 0.044 0.26 0.008 ‐ Net Income Before Tax / Assets ‐ 0.05 ‐ 1.57 1.52 0.533*** 0.338 ‐ Change in T1C / RWA 1.77 ‐ 3.63 1.87 0.145 0.091 2 from calculated is Loans R Provision a) Expense Total / Expense Provision of projections DFAST Note: "Other" assets. total of 0.86% is loss trading projected losses. counterparty and trading of inclusive reported is PPNR held and sale for loans of value fair in change projected the as well as (AFS/HTM), securities on losses/gains realized includes losses. impairment goodwill and option, value fair the under measured investment for held loans CLASS projections to performance during the financial crisis Comparing 8: Table CLASS projections are over 2007:Q3 and compared to the actual evolution of capital the six quarters between to from quarters nine the 2009:Q4. over income net of evolution actual the 2007:Q3 and 2008:Q4, Actual vs predicted: across firms (1) predicted (actual + β . = + ε ) α Industry values Weighted Unweighted Slope Slope 2 2 )R Model Actual Difference coefficient ( β ( )R coefficient β and loan Income (9 quarter cumulative, annualized): performance PPNR / total assets 1.54 1.47 0.07 0.552*** 0.223 0.194*** 0.068 1.99 Net rate 1.93 chargeoff ‐ 0.05 1.284*** 0.674 0.609*** 0.120 Return on assets 0.13 ‐ 0.05 0.18 0.558*** 0.094 0.229** 0.025 (6 Change in T1C / RWA qtr) ‐ 1.12 ‐ 1.77 0.65 0.593*** 0.079 0.288*** 0.086 (1) 98%) and Based on winsorized OLS (winsorized at 2% 59

62 Online Appendix for: Assessing Financial Stability: The Capital and Loss Assessment under Stress Scenarios (CLASS) Model By Beverly Hirtle, Anna Kovner, James Vickery and Meru Bhanot Contents 1. Details of methodology for specification of CLASS regression models page 2 2. Relationship between asset composition and macroeconomic conditions page 27 page 30 3. Loan Losses, Reserves and Provisions: Additional Discussion 1

63 1. Details of methodology for specification of CLASS regression models This section describes in more detail our approach for choosing the specifications of each income and loss equation used in the CLASS model. We conducted a search across at least six different specifications for each equation. Decisions about whic h specifications to consider and finally select were guided by the following four principles: Statistical and economic significance. We chose macroeconomic variables and bank- (i) level controls that were statistically and economically significantly related to the income component ratio historically. (ii) We customized our Consistency with economic theory and prior empirical research. specification search to include specifications that were consistent with our priors based both theory and empirical research suggest on existing academic research. For example that mortgage loss rates rise in a convex fashion as real estate price growth declines, reflecting the fact that mortgage default is effectively a put option on the value of the underlying real estate collateral. We capture this convexity by including a variable that measures real estate price growth inter acted with a dummy for price growth being 1 This variable was not negative in the equations governing mortgage loss rates. considered in other models. In contrast, we focused on specifications including different combinations of financial market variables in income categories most closely related to market conditions (e.g., trading income and non-trading noninterest income). Sensitivity to macroeconomic conditions. In three of 22 categories, no macroeconomic (iii) variable was found to be statistically significantly related to the income or loss rate being modelled over the sample period. Given our priors that income these categories would still exhibit some sensitivity to the macroeconomy, we included a single macroeconomic variable with the expected sign in these specifications; to err on the side of assuming that bank condition is sensitive to macroeconomic conditions. (iv) Parsimony. Other things equal, we selected simple specifications, both for ease of understanding and also to avoid over-fitting models in-sample. 1 We find that this nonlinearity is important empirica lly, as shown in the tables below. We find little or no relationship between real estate price growth and mortga ge NCOs in the region where price growth is positive. There is however a quantitatively important and statistica lly significant inverse relationship between price growth and loan loss rates in the region where price growth is negative. 2

64 In each income and expense category, we conducted a specification search including at least six different model specifications, in order to identify a preferred specification based on these principles. For each income component, the specifications considered included: 2  AR(1) – A simple autoregressive model with one lag and no macroeconomic variables. – The AR(1) specification plus a set of plausible macroeconomic  Basic Specification variables. For each net chargeoff regression this set included at least three common macro variables: stock market returns, GDP growth, and the change in the unemployment rate. For each PPNR regression the set included these three plus the term spread. In addition to this common set, we then included other variables as appropriate for the equation in question (e.g., real estate price growth for the NCO rate equations for different types of commercial and residential mortgages). This specification also included controls for bank portfolio ear time trend is included in our preferred characteristics for the PPNR specifications. A lin specification whenever statistically significant and/or visually obvious from a graphical 3 inspection of the data. Preferred Variables – Includes all explanatory variables from Basic Specification equation  which are statistically significant and have the expected sign.  Non-linearities – Includes Preferred Variables and allows for the average level or the dependent variable to be different in the worst 10% of realizations of the macroeconomic variable as well as the slope of the relationship with Preferred Variables to differ in the worst 10% of realizations of that variable. For non-linear transformations of the change in the bond spread and real estate prices we allow the slope to differ when the change is greater than zero. We then also considered other specifications judged appropriate for the equation in question (e.g., we experimented with weighted least squares in some cases, weighting by the recency of the data to 4 , or with other transformations of the allow for structural changes in the macro relationships 2 We have also experimented with specifications with more lags; these did not generally result in significant additional explanatory power. it requires only one quarter of historical An AR(1) model is also attractive because data to “seed” the model when producing recursive model projections. This is helpful because of the difficulty of adjusting for mergers (especially out-of-indust ry mergers) over the historical period. 3 Note that when computing model projections, we assume t continue over the forecast that this time-trend does no horizon, instead it is held constant at its most recent historical value. 4 Two different methodologies are used, which we refer to as linear and exponential time-weighting. The weight assigned to each observation under the exponential approach is equal to 1 / ( current quarter – ( quarter of - first quarter in observation - 1)). The weight assigned under the linear approach is equal to ( quarter of observation 3

65 macroeconomic variables (e.g., using a two-year change in real estate prices rather than a one-year change). As also discussed in the main text, while most models are estimated using time-series models based on the aggregate banking industry, we estimate so me models using a pooled regression approach. This allows us to control for firm characteristics, which may reflect differences in bank business models that may be related to the dependent variable of interest. The use of these controls allows net interest margin and non-interest expense categories to vary directly with banks’ business focus, and effectively means that banks converge to the long run average of banks with a similar business focus and size. For example, banks that focus on credit card loans (with historically higher net interest margins) will continue to have higher levels of projected net interest margin, even in the long run. In these regressions we estimate the same basic specifications and cluster standard errors by quarter to account for correlated residuals arising from multiple observations in each quarter. In these pooled regression specifications we weight by the size of each institution in th at category (the relevant assets 5 of the institution in that quarter divided by the relevant total banking industry assets in that quarter). Results from our model specification searches are presented in tables on the following pages (one . These specification searches were conducted using table for each model; 22 tables of results in total) a slightly shorter sample period than the equations presented in the main text (data from 1991 up to 2012:Q1 rather than up to 2013:Q3). As can be seen by comparing the estimated equations in the main paper with those below, however, this difference in sample period causes only small changes in estimated relationships (reflecting that it was a period when macroeconomic conditions were not particularly unusual). series ) / ( total number of quarters ). In earlier versions of CLASS, all ti me-weighted specifications used (what we now call) the exponential method. 5 compensation expense by total asset share and AFS by AFS NIM is thus weighted by interest earning asset share, balance share. 4

66 Specification Searches Net Interest Margin 4.5 4 3.5 3 2.5 1 1 1 1 1 1 1 1 q q q q q 1q 5q 3q Net Interest Income Over Interest Earning Assets 9 99 97 9 01 03 9 0 9 0 9 9 9 2 2007q1 2005 1 2009q1 2011q1 2 1 1 1 19 Fitted Values Historical Data are annualized, unadjusted, and for all domestic Y9C and call report BHC and bank filers. Net Interest Margin eferred Variables AR-1 with Ratios AR-1 Basic Specs Preferred Specifications Pr with Non-Linearity 0.7981*** Lagged Net Interest Margin 0.8487*** 0.7222*** 0.7232*** 0.7236*** 0.7237*** (0.0428) (0.0712) (0.0711) (0.0458) (0.0706) (0.0710) 0.0147 0.0043 0.0477** Term Spread (10Y-3M) 0.0092 (0.0165) (0.0279) (0.0272) (0.0285) Stock Market Quarterly Log Change 0.0001 (0.0015) Real GDP Annualized Log Change 0.0066 (0.0036) Annualized Change in Unemployment -0.0104 (0.0081) 3M Treasury Yield 0.0258* 0.0041 -0.0044 -0.0026 (0.0123) (0.0216) (0.0197) (0.0219) Term Spread X <10% Dummy 0.3001 (0.2048) Term Spread <10% Dummy -0.1116 (0.0593) 0.0039*** 0.0040*** 0.0041*** 0.0041*** Residential RE Loan Ratio 0.0045** (0.0010) (0.0010) (0.0010) (0.0010) (0.0015) Commercial RE Loan Ratio 0.0061*** 0.0055** 0.0063*** 0.0062*** 0.0062*** (0.0017) (0.0013) (0.0013) (0.0014) (0.0013) 0.0088*** 0.0086*** 0.0062*** C&I Loan Ratio 0.0087*** 0.0085*** (0.0015) (0.0022) (0.0022) (0.0022) (0.0022) 0.0223*** Credit Card Loan Ratio 0.0173*** 0.0224*** 0.0225*** 0.0224*** (0.0057) (0.0057) (0.0057) (0.0057) (0.0039) -0.0141** -0.0140** -0.0142** -0.0141** -0.0069*** Trading Assets Ratio (0.0046) (0.0046) (0.0045) (0.0046) (0.0019) 0.0005 0.0004 0.0036** 0.0004 0.0004 Securities Ratio (0.0012) (0.0010) (0.0009) (0.0010) (0.0010) 0.0086*** Asset Share 0.0086*** 0.0070*** 0.0085*** 0.0085*** (0.0015) (0.0024) (0.0025) (0.0024) (0.0024) -0.0042 -0.0168* -0.0159*** -0.0152 -0.0143* Time Trend -0.0099*** (0.0035) (0.0039) (0.0029) (0.0077) (0.0069) (0.0080) 1.2982 1.3712*** 1.3935* 1.2883* 0.3554 Constant 1.0205*** (0.2508) (0.2976) (0.5991) (0.6605) (0.3747) (0.6375) Observations 16350 16350 16350 16350 16350 16350 No No No No No With IEA Share Weighting Yes 0.88 Adjusted R-squared 0.71 0.71 0.71 0.71 0.69 Sample is for all BHCs (Y9-C) plus all other commercial ba nks (Call) from 1991q1-2012q1. All institution flow variables are ann ualized. Standard errors (clustered by quarter) in parentheses. * significant at 10%; ** significant at 5%; *** significant at 1% 5

67 Noninterest Non-Trading Income Ratio 3 2.5 2 1.5 03q1 91q1 Noninterest Non-Trading Income Over Total Assets 2009q1 2007q1 2005q1 2001q1 20 19 1999q1 1997q1 1995q1 1993q1 2011q1 Historical Fitted Values Data are annualized, unadjusted, and for all domestic Y9C and call report BHC and bank filers. Preferred Preferred AR-1 Basic Specs with Non-Linearity Specifications Specifications (2011) 0.9311*** 0.9309*** Lagged NINT Ratio 0.8986*** 0.9309*** 0.9305*** (0.0118) (0.0118) (0.0161) (0.0118) (0.0118) -3M) -0.0056 Term Spread (10Y (0.0116) 0.0051* 0.0064* 0.0062* Stock Market Quarterly Log Change 0.0059* (0.0028) (0.0025) (0.0025) (0.0025) Real GDP Annualized Log Change 0.0049 (0.0074) 0.0295 Annualized Change in Unemployment (0.0177) Stock Change X <10% Dummy 0.0216 (0.0145) Stock Change <10% Dummy 0.3856 (0.2107) -0.0012 Residential RE Loan Ratio (0.0020) -0.0043* Commercial RE Loan Ratio (0.0020) C&I Loan Ratio -0.0016 (0.0022) Credit Card Loan Ratio 0.0104*** (0.0027) Trading Assets Ratio -0.0018 (0.0024) 0.0031 Securities Ratio (0.0021) Asset Share -0.0062*** (0.0018) Constant 0.1543*** 0.2650 0.1401*** 0.1389*** 0.1429*** (0.0293) (0.0401) (0.1428) (0.0304) (0.0303) 16350 Observations 16350 16350 16350 16350 Adjusted R-squared 0.87 0.87 0.87 0.87 0.88 Sample is for all BHCs (Y9-C) plus all other commercial ba nks (Call) from 1991q1-2012q1. All institution flow variables are ann ualized. Robust standard errors in parentheses. * significant at 10%; ** significant at 5%; *** significant at 1% 6

68 Trading Ratio 5 0 -5 Trading Income Over Trading Assets -10 1 2002q1 1994q 2012q1 2010q1 2008q1 1996q1 2004q1 1998q1 2000q1 2006q1 Fitted Values Historical Data are annualized, unadjusted, and for all domestic Y9C and call report BHC and bank filers. Trading Ratio with Level of Preferred Preferred With Stock AR-1 Basic Specs Specifications Bond Spread Variables Market Growth 0.3088* Lagged Trading Ratio 0.2906* 0.377 0.2257 0.4391 0.2349 (0.1265) (0.1811) (0.2259) (0.1737) (0.1380) (0.2377) -3M) -0.2379 Term Spread (10Y (0.1775) 0.0682 0.0091 Stock Market Quarterly Log (0.0342) (0.0220) Change 0.0362 Real GDP Annualized Log Change (0.1238) Annualized Change in -0.4115* Unemployment (0.1913) -0.0925 0.5612 Bond Spread (BBB – 10Y) 1.2388** (0.3479) (0.4520) (0.3890) -0.0553 -2.6590*** Quarterly Change in Bond -0.1329 -0.9131 (0.6849) (0.7776) (0.4567) (0.8571) Spread Change in Bond Spread X >0 -4.7328** -2.9786*** -4.5751** Dummy (1.4892) (0.6138) (1.4240) 2.1556*** 1.9404** Constant 1.2346 0.2925 1.2888 1.093 (0.7394) (0.9410) (0.7081) (0.6037) (0.7432) (0.7184) Observations 61 61 61 61 61 61 0.47 0.48 0.24 0.41 0.48 0.18 Adjusted R-squared ualized. Robust standard Sample is for all BHCs (Y9-C) plus all other commercial ba nks (Call) from 1991q1-2012q1. All institution flow variables are ann errors in parentheses. * significant at 10%; ** significant at 5%; *** significant at 1% 7

69 Compensation Noninterest Expense Ratio 2 1.8 1.6 1.4 Compensation NIE Over Total Assets 1.2 1 1 q1 q1 q1 q1 q1 q1 q1 q1 q1 7 9q 3 5 7 1q 5 3 1 1 9 00 00 00 00 99 00 01 2 2 2 199 199 2 2 199 1 2 199 Historical Fitted Values Data are annualized, unadjusted, and for all domestic Y9C and call report BHC and bank filers. Compensation Noninterest Expense Ratio Preferred AR-1 riables with Non-Linearity Preferred Va Basic Specs Specifications Preferred Specifications (2011) 0.9428*** 0.8865*** 0.9429*** 0.9430*** 0.9458*** 0.8868*** Lagged Compensation Ratio (0.0360) (0.0360) (0.0360) (0.0347) (0.0196) (0.0197) 0.0022 Term Spread (10Y-3M) (0.0030) Stock Market Quarterly Log Change 0.0010* 0.0011 0.0034*** 0.0012 0.0034*** (0.0009) (0.0009) (0.0005) (0.0004) (0.0007) -0.0007 Real GDP Annualized Log Change (0.0024) Annualized Change in Unemployment -0.0039 (0.0045) 0.0082*** Stock Change X <10% Dummy (0.0014) Stock Change <10% Dummy 0.1442*** (0.0242) -0.0012 Residential RE Loan Ratio -0.0003 -0.0003 -0.0003 -0.0012 (0.0011) (0.0011) (0.0004) (0.0004) (0.0004) -0.0019* Commercial RE Loan Ratio -0.0008 -0.0008 -0.0008 -0.0019* (0.0006) (0.0006) (0.0006) (0.0008) (0.0008) 0.0001 -0.0013 -0.0012 C&I Loan Ratio 0.0001 0.0001 (0.0017) (0.0018) (0.0005) (0.0005) (0.0005) -0.0009 -0.0009 -0.0020 -0.0020 -0.0009 Credit Card Loan Ratio (0.0012) (0.0012) (0.0007) (0.0007) (0.0007) 0.0000 -0.0035** 0.0000 Trading Assets Ratio 0.0000 -0.0035** (0.0015) (0.0015) (0.0015) (0.0012) (0.0012) Securities Ratio -0.0007 -0.0025* -0.0006 -0.0007 -0.0025* (0.0006) (0.0006) (0.0006) (0.0012) (0.0013) Asset Share -0.0004 (0.0009) 0.3374** 0.3429** 0.1343 0.1344 Constant 0.0912 0.1376 (0.0984) (0.0984) (0.0994) (0.0564) (0.1181) (0.1121) Observations 16350 16350 16350 16350 16350 16350 With Asset Share Weighting No No No No Yes Yes 0.82 0.82 Adjusted R-squared 0.91 0.91 0.91 0.91 Sample is for all BHCs (Y9-C) plus all other commercial ba nks (Call) from 1991q1-2012q1. All institution flow variables are ann ualized. Standard errors (clustered by quarter) in * significant at 10%; ** significant at 5%; *** significant at 1% parentheses. 8

70 Fixed Assets Noninterest Expense Ratio .6 .5 .4 Fixed Assets NIE Over Total Assets .3 1 1 1 1 1 q1 q q q q1 q q1 q q1 1 3 05 993 007 009 991 200 2011 200 1999q1 20 2 1997 1995q1 2 1 1 Fitted Values Historical Data are annualized, unadjusted, and for all domestic Y9C and call report BHC and bank filers. Fixed Asset Noninterest Expense Preferred Specifications Preferred Specifications riables with Non-Linearity Preferred Va Basic Specs AR-1 (2011) 0.9155*** 0.9155*** Lagged Fixed Asset Ratio 0.9155*** 0.8862*** 0.9208*** 0.8865*** (0.0213) (0.0213) (0.0213) (0.0201) (0.0174) (0.0175) Term Spread (10Y-3M) 0.0001 (0.0009) Stock Market Quarterly Log Change -0.0001 (0.0001) -0.0002 -0.0001 -0.0003 0.0005 Real GDP Annualized Log Change 0.0005 (0.0004) (0.0003) (0.0005) (0.0007) (0.0007) 0.0001 Annualized Change in Unemployment (0.0007) Change in GDP X <10% Dummy 0.0005 (0.0008) Change in GDP <10% Dummy 0.0040 (0.0040) -0.0003 -0.0002* -0.0002* Residential RE Loan Ratio -0.0003 -0.0002* (0.0002) (0.0002) (0.0001) (0.0001) (0.0001) -0.0003 Commercial RE Loan Ratio -0.0002 -0.0002 -0.0002 -0.0003 (0.0002) (0.0002) (0.0001) (0.0001) (0.0001) -0.0004* -0.0002 C&I Loan Ratio -0.0004* -0.0002 -0.0002 (0.0002) (0.0002) (0.0002) (0.0002) (0.0002) -0.0004** Credit Card Loan Ratio -0.0004** -0.0004** -0.0004** -0.0004** (0.0001) (0.0001) (0.0001) (0.0001) (0.0001) -0.0011** -0.0006** -0.0011** -0.0006** Trading Assets Ratio -0.0006** (0.0004) (0.0004) (0.0002) (0.0002) (0.0002) Securities Ratio -0.0003 -0.0007*** -0.0003 -0.0007*** -0.0003 (0.0002) (0.0002) (0.0001) (0.0001) (0.0001) -0.0014*** Time Trend -0.0007** -0.0007** -0.0007** -0.0014*** (0.0002) (0.0002) (0.0002) (0.0004) (0.0003) Asset Share -0.0001 (0.0002) 0.1463*** 0.0883*** 0.1479*** Constant 0.0353*** 0.0890*** 0.0887*** (0.0242) (0.0233) (0.0227) (0.0224) (0.0090) (0.0227) Observations 16350 16350 16350 16350 16350 16350 With Asset Share Weighting No No No Yes Yes No 0.86 0.86 Adjusted R-squared 0.86 0.86 0.86 0.86 nks (Call) from 1991q1-2012q1. All institution flow variables are ann Sample is for all BHCs (Y9-C) plus all other commercial ba ualized. Standard errors (clustered by quarter) in * significant at 10%; ** significant at 5%; *** significant at 1% parentheses. 9

71 All Other Noninterest Expense Ratio 2.2 2 1.8 1.6 1.4 1.2 1 1 1 1 q1 q q q1 q1 q q 9 All Other Noninterest Expense Over Total Assets 03 01 005 991 2 1 199 1997q1 1995q1 20 2007 1993 2009q1 2011q1 20 Fitted Values Historical Data are annualized, unadjusted, and for all domestic Y9C and call report BHC and bank filers. All Other Noninterest Expense Preferred Preferred Specifications Specifications riables with Non-Linearity Preferred Va Basic Specs AR-1 (2011) 0.8125*** 0.8124*** 0.8621*** 0.8121*** 0.8318*** Lagged Fixed Asset Ratio 0.8298*** (0.0661) (0.0500) (0.0660) (0.0661) (0.0351) (0.0348) Term Spread (10Y-3M) 0.0246 0.0122 (0.0133) (0.0126) Stock Market Quarterly Log Change 0.0063*** 0.0051** (0.0018) (0.0018) 0.0124 0.0214** Real GDP Annualized Log Change (0.0080) (0.0086) Annualized Change in Unemployment 0.0439 0.0391 (0.0277) (0.0247) Quarterly Change in Bond Spread 0.1954* 0.1960** 0.1932* 0.0990 0.0761 (0.0904) (0.0911) (0.0931) (0.1284) (0.0734) Change in Bond Spread X <10% Dummy 0.3759* (0.1565) Change in Bond Spread <10% Dummy -0.2226*** (0.0621) Residential RE Loan Ratio -0.0038* -0.0013 -0.0040* -0.0013 -0.0038* (0.0023) (0.0023) (0.0015) (0.0016) (0.0015) -0.0019 -0.0007 Commercial RE Loan Ratio -0.0019 -0.0022 -0.0009 (0.0014) (0.0015) (0.0032) (0.0032) (0.0032) -0.0007 C&I Loan Ratio -0.0027 -0.0028 -0.0027 -0.0018 (0.0026) (0.0026) (0.0025) (0.0022) (0.0021) Credit Card Loan Ratio 0.0144*** 0.0140** 0.0147*** 0.0142** 0.0142** (0.0037) (0.0037) (0.0043) (0.0044) (0.0044) Trading Assets Ratio -0.0009 -0.0005 -0.0013 -0.0006 -0.0009 (0.0019) (0.0019) (0.0029) (0.0029) (0.0029) 0.0086** Securities Ratio 0.0002 0.0003 0.0086** 0.0003 (0.0014) (0.0014) (0.0014) (0.0026) (0.0026) Asset Share -0.0039* (0.0018) 0.1189 0.2868 0.3991 Constant 0.2267** 0.2874 0.0710 (0.1134) (0.1174) (0.2183) (0.2177) (0.0711) (0.2169) 16350 Observations 16350 16350 16350 16350 16350 Yes No No No No Yes With Asset Share Weighting 0.79 0.79 Adjusted R-squared 0.72 0.72 0.72 0.72 Sample is for all BHCs (Y9-C) plus all other commercial ba nks (Call) from 1991q1-2012q1. All institution flow variables are ann ualized. Standard errors (clustered by quarter) in parentheses. * significant at 10%; ** significant at 5%; *** significant at 1% 10

72 Realized Gains/Losses on AFS Securities 1 0 -1 -2 -3 Realized Gains/Losses on AFS Securities 1 1991q1 2011q1 2009q1 2007q1 1995q1 2003q1 2001q1 1993q1 1999q 1997q1 2005q1 Fitted Value Historical Data are annualized, unadjusted, and for all domestic Y9C and call report BHC and bank filers. Gains and Losses on AFS Securities with Non- Unsafe Ratio Preferred Preferred AR-1 Basic Specs Specifications Linearity Variables Interactions 0.1307 0.1341 Lagged AFS Return 0.1346 0.1299 0.1353 0.1419 (0.1002) (0.1037) (0.1015) (0.1055) (0.1054) (0.1117) -0.0111 Term Spread (10Y - 3M) (0.0456) Stock Market Quarterly Log Change -0.0069 (0.0079) 0.0585 Real GDP Annualized Log Change (0.0400) -0.0307 Annualized Change in Unemployment (0.0710) -0.1237 Bond Spread (BBB - 10Y) -0.0142 -0.1560 (0.0981) (0.1109) (0.1433) Quarterly Change in Bond Spread -0.4305 -0.4094 (0.2747) (0.2882) Change in the 10 Year Treasury Yield -0.7114*** (0.1819) -0.0031** Unsafe Ratio X Bond Spread (0.0011) Unsafe Ratio X Change in Bond Spread -0.0130 -0.0141 (0.0074) (0.0084) -0.0315*** Unsafe Ratio X Change in Bond Spread (0.0032) X >0 Dummy 0.4924** 0.1225 Constant 0.1759 0.2811*** 0.4205* 0.3903*** (0.2933) (0.0886) (0.1648) (0.1805) (0.0777) (0.0638) 11609 11609 11609 11609 Observations 11609 11609 0.02 Adjusted R-squared 0.04 0.02 0.03 0.02 0.02 Sample is for all BHCs (Y9-C) plus all other commercial ba ualized. Standard errors (clustered by quarter) in nks (Call) from 1991q1-2012q1. All institution flow variables are ann parentheses. * significant at 10%; ** significant at 5%; *** significant at 1% 11

73 Net Charge Off Rate on First Lien RRE Loans 2 1.5 1 .5 0 Net Charge Off Rate on First Lien RRE Loans 91q1 03q1 2001q1 2011q1 1993q1 1995q1 2009q1 2007q1 2005q1 19 20 1997q1 1999q1 Historical Fitted Value Data are annualized, unadjusted, and for all domestic Y9C and call report BHC and bank filers. Net Charge Off Rate on First Liens 2 Year Case 1 Year Case Unemployment Unemployment AR-1 Basic Specs Shiller Shiller w/ 2 Year CS w/ 1 Year CS Variables Variables 0.8357*** Lagged NCO Rate on First Liens 0.9718*** 0.9139*** 0.9298*** 0.9229*** 0.8265*** (0.0586) (0.0614) (0.0634) (0.0618) (0.0663) (0.0620) Stock Market Quarterly Log Change 0.0014 (0.0020) 0.0126 Real GDP Annualized Log Change (0.0080) 0.0556* Annualized Change in Unemployment 0.0307 0.0074 (0.0173) (0.0245) (0.0244) -0.0017 House Price Index 4Q Log Change -0.0051* -0.0008 (0.0018) (0.0019) (0.0024) 4Q ∆ in HPI X <0 Dummy -0.0142* -0.0069 (0.0086) (0.0069) House Price Index 8Q Log Change -0.0003 -0.0004 (0.0009) (0.0009) -0.0136** -0.0123* ∆ in HPI X <0 Dummy 8Q (0.0054) (0.0050) 0.0272 0.027 0.016 Constant 0.0206 0.0284 0.0114 (0.0129) (0.0147) (0.0141) (0.0152) (0.0135) (0.0349) 84 Observations 84 84 84 84 84 Adjusted R-squared 0.83 0.94 0.94 0.95 0.94 0.95 ualized. Robust standard Sample is for all BHCs (Y9-C) plus all other commercial ba nks (Call) from 1991q1-2012q1. All institution flow variables are ann errors in parentheses. * significant at 10%; ** significant at 5%; *** significant at 1% 12

74 Net Charge Off Rate on Junior Lien RRE Loans 8 6 4 2 0 Net Charge Off Rate on Junior Lien RRE Loans 2007q1 2001q1 1997q1 1991q1 2011q1 2009q1 1999q1 2005q1 2003q1 1993q1 1995q1 Fitted Value Historical Data are annualized, unadjusted, and for all domestic Y9C and call report BHC and bank filers. Net Charge Off Rate on Junior Liens 2 Year Case 1 Year Case Unemployment Unemployment AR-1 Basic Specs Shiller Shiller w/ 2 Year CS w/ 1 Year CS Variables Variables Lagged NCO Rate on First Liens 0.9173*** 0.9163*** 0.8278*** 0.9902*** 0.8289*** 0.9194*** (0.0661) (0.0620) (0.0668) (0.0685) (0.0643) (0.0663) Stock Market Quarterly Log Change 0.0046 (0.0047) 0.0328 Real GDP Annualized Log Change (0.0224) Annualized Change in Unemployment 0.0710 -0.0032 0.0462 (0.0462) (0.0693) (0.0617) -0.0059 House Price Index 4Q Log Change -0.0034 -0.0046 (0.0052) (0.0051) (0.0048) 4Q ∆ in HPI X <0 Dummy -0.0650** -0.0608** -0.0497 (0.0226) (0.0271) (0.0202) House Price Index 8Q Log Change -0.0035 -0.0035 (0.0029) (0.0028) -0.0491*** -0.0496*** in HPI X <0 Dummy ∆ 8Q (0.0131) (0.0117) 0.0730 0.1256 Constant 0.0643 0.1251 0.0553 -0.0532 (0.0481) (0.0470) (0.0648) (0.0665) (0.0518) (0.0889) 84 84 84 84 Observations 84 84 0.95 0.97 0.96 0.97 0.96 0.97 Adjusted R-squared ualized. Robust standard nks (Call) from 1991q1-2012q1. All institution flow variables are ann Sample is for all BHCs (Y9-C) plus all other commercial ba errors in parentheses. * significant at 10%; ** significant at 5%; *** significant at 1% 13

75 Net Charge Off Rate on HELOC RRE Loans 4 3 2 1 0 Net Charge Off Rate on HELOC RRE Loans 2007q1 2001q1 1997q1 1991q1 2011q1 2009q1 1999q1 2005q1 2003q1 1993q1 1995q1 Fitted Value Historical Data are annualized, unadjusted, and for all domestic Y9C and call report BHC and bank filers. Net Charge Off Rate on HELOCs 2 Year Case 1 Year Case Unemployment Unemployment AR-1 Basic Specs Shiller Shiller w/ 2 Year CS w/ 1 Year CS Variables Variables Lagged NCO Rate on First Liens 0.9250*** 0.9265*** 0.8385*** 0.9964*** 0.8392*** 0.9297*** (0.0277) (0.0338) (0.0373) (0.0341) (0.0301) (0.0305) Stock Market Quarterly Log Change 0.0050* (0.0019) 0.0042 Real GDP Annualized Log Change (0.0060) Annualized Change in Unemployment 0.0298 -0.0009 0.0230 (0.0169) (0.0187) (0.0183) -0.0004 House Price Index 4Q Log Change 0.0011 0.0003 (0.0016) (0.0017) (0.0017) 4Q ∆ in HPI X <0 Dummy -0.0315*** -0.0312*** -0.0257*** (0.0064) (0.0075) (0.0075) House Price Index 8Q Log Change -0.0011 -0.0011 (0.0009) (0.0009) -0.0243*** -0.0245*** in HPI X <0 Dummy ∆ 8Q (0.0056) (0.0051) 0.0175 0.0483** Constant 0.0253 0.0481** 0.0087 -0.0145 (0.0180) (0.0162) (0.0181) (0.0177) (0.0143) (0.0245) 84 84 84 84 Observations 84 84 0.97 0.99 0.98 0.99 0.99 0.99 Adjusted R-squared ualized. Robust standard nks (Call) from 1991q1-2012q1. All institution flow variables are ann Sample is for all BHCs (Y9-C) plus all other commercial ba errors in parentheses. * significant at 10%; ** significant at 5%; *** significant at 1% 14

76 Net Charge Off Rate on Construction CRE Loans 8 6 4 2 0 Net Charge Off Rate on Construction CRE Loans 2007q1 1991q1 1993q1 2011q1 2009q1 1999q1 2005q1 2003q1 2001q1 1997q1 1995q1 Historical Fitted Value Data are annualized, unadjusted, and for all domestic Y9C and call report BHC and bank filers. Net Charge Off Rate on Construction CRE Loans Unemployment w/ Exponential Preferred w/ Linear Time AR-1 Basic Specs and 4Q Index Time Weighting Specifications Weighting Change 0.7677*** 0.7895*** Lagged NCO Rate on Construction 0.9279*** 0.8189*** 0.7993*** 0.8144*** (0.0918) (0.0742) (0.0941) (0.0882) (0.0875) (0.0954) -0.0046 Stock Market Quarterly Log Change (0.0096) 0.0216 Real GDP Annualized Log Change (0.0521) Annualized Change in Unemployment 0.1531 0.1344 (0.1645) (0.1731) CRE Price Index 4Q Log Change 0.0103 0.0095 (0.0135) (0.0134) -0.0724** -0.0668* Change in CRE Index X <0 Dummy -0.0712** -0.0656** -0.0676** (0.0260) (0.0264) (0.0227) (0.0243) (0.0235) Constant 0.0941 0.1452 -0.0207 0.0940 0.0390 0.1499* (0.2076) (0.0603) (0.0726) (0.1127) (0.2078) (0.0549) 84 Observations 84 84 84 84 84 0.82 Adjusted R-squared 0.86 0.89 0.89 0.90 0.89 nks (Call) from 1991q1-2012q1. All institution flow variables are ann ualized. Robust standard errors in parentheses. Sample is for all BHCs (Y9-C) plus all other commercial ba * significant at 10%; ** significant at 5%; *** significant at 1% 15

77 Net Charge Off Rate on Multi-Family CRE Loans 2.5 2 1.5 1 .5 0 Net Charge Off Rate on Multi-Family CRE Loans 03q1 91q1 2007q1 19 20 2011q1 2001q1 2005q1 1999q1 1997q1 1995q1 1993q1 2009q1 Fitted Value Historical Data are annualized, unadjusted, and for all domestic Y9C and call report BHC and bank filers. Net Charge Off Rate on Multi-Family CRE Loans Unemployment Preferred w/ Exponential w/ Linear Time AR-1 Basic Specs and 4Q Index Time Weighting Specifications Weighting Change Lagged NCO Rate on Construction 0.7077*** 0.7881*** 0.8557*** 0.7694*** 0.7422*** 0.7807*** (0.0912) (0.1230) (0.1025) (0.1143) (0.1044) (0.0978) Stock Market Quarterly Log Change -0.0017 (0.0032) Real GDP Annualized Log Change 0.0149 (0.0194) Annualized Change in Unemployment 0.0345 0.0173 (0.0603) (0.0480) CRE Price Index 4Q Log Change 0.0007 0.0002 (0.0059) (0.0060) Change in CRE Index X <0 Dummy -0.0142 -0.0156** -0.0157** -0.0153** -0.0137 (0.0075) (0.0072) (0.0049) (0.0054) (0.0051) 0.0739 Constant 0.0533* 0.0032 0.0467* 0.0452 0.0492* (0.0242) (0.0833) (0.0202) (0.0522) (0.0468) (0.0218) Observations 84 84 84 84 84 84 0.74 Adjusted R-squared 0.68 0.77 0.76 0.79 0.76 errors in parentheses. ualized. Robust standard nks (Call) from 1991q1-2012q1. All institution flow variables are ann Sample is for all BHCs (Y9-C) plus all other commercial ba * significant at 10%; ** significant at 5%; *** significant at 1% 16

78 Net Charge Off Rate on Non-Farm Non-Res CRE Loans 2 1.5 1 .5 0 91q1 03q1 1999q1 2009q1 1993q1 1995q1 1997q1 2007q1 2005q1 19 20 2001q1 2011q1 Net Charge Off Rate on Non-Farm Non-Res CRE Loa Fitted Value Historical Data are annualized, unadjusted, and for all domestic Y9C and call report BHC and bank filers. Net Charge Off Rate on Non-Farm Non-Res CRE Loans Unemployment w/ Exponential Preferred w/ Linear Time AR-1 Basic Specs and 4Q Index Specifications Time Weighting Weighting Change 0.7935*** Lagged NCO Rate on Construction 0.8794*** 0.8453*** 0.8169*** 0.8314*** 0.8501*** (0.1340) (0.1004) (0.0928) (0.1203) (0.0922) (0.1009) Stock Market Quarterly Log Change -0.0004 (0.0033) 0.0190 Real GDP Annualized Log Change (0.0196) 0.0484 Annualized Change in Unemployment 0.0200 (0.0388) (0.0561) CRE Price Index 4Q Log Change -0.0003 -0.0009 (0.0039) (0.0037) -0.0087 Change in CRE Index X <0 Dummy -0.0126*** -0.0117* -0.0089 -0.0101** (0.0056) (0.0037) (0.0051) (0.0037) (0.0047) Constant 0.0469 0.0375 -0.0136 0.0309 0.0425 0.0253 (0.0862) (0.0240) (0.0162) (0.0403) (0.0462) (0.0227) Observations 84 84 84 84 84 84 Adjusted R-squared 0.77 0.79 0.73 0.79 0.79 0.85 nks (Call) from 1991q1-2012q1. All institution flow variables are ann Sample is for all BHCs (Y9-C) plus all other commercial ba errors in parentheses. ualized. Robust standard * significant at 10%; ** significant at 5%; *** significant at 1% 17

79 Net Charge Off Rate on C&I Loans 3 2 1 Net Charge Off Rate on C&I Loans 0 2009q1 2001q1 1993q1 1999q1 1995q1 2011q1 1991q1 2007q1 2005q1 2003q1 1997q1 Historical Fitted Value Data are annualized, unadjusted, and for all domestic Y9C and call report BHC and bank filers. Net Charge Off Rate on C&I Loans w/ Exponential w/ Linear Time with Non- Preferred AR-1 Basic Specs Time Linearity Specifications Weighting Weighting 0.7934*** 0.8219*** Lagged NCO Rate on C&I Loans 0.8773*** 0.7494*** 0.8185*** 0.7937*** (0.0769) (0.0727) (0.0663) (0.0705) (0.0636) (0.0650) 0.0042 Stock Market Quarterly Log Change (0.0043) Real GDP Annualized Log Change 0.0328* (0.0150) 0.1333*** 0.1042 Annualized Change in Unemployment 0.1404*** 0.1343** 0.1253*** (0.0330) (0.0544) (0.0245) (0.0341) (0.0476) 0.1247 Bond Spread (BBB-10Y) (0.0683) -0.0395 Unemployment X <10% Dummy (0.0948) Unemployment <10% Dummy 0.3429 (0.2323) 0.1434** 0.1523** -0.0866 0.1717*** Constant 0.1054* 0.1465* (0.0529) (0.0495) (0.0499) (0.0674) (0.1096) (0.0507) Observations 84 84 84 84 84 84 Adjusted R-squared 0.88 0.82 0.82 0.87 0.83 0.76 Sample is for all BHCs (Y9-C) plus all other commercial ba nks (Call) from 1991q1-2012q1. All institution flow variables are ann errors in parentheses. ualized. Robust standard * significant at 10%; ** significant at 5%; *** significant at 1% 18

80 Net Charge Off Rate on Credit Card Loans 12 10 8 6 4 2 Net Charge Off Rate on Credit Card Loans 09q1 97q1 2001q1 20 1999q1 2011q1 2003q1 2005q1 2007q1 19 1995q1 1993q1 1991q1 Historical Fitted Value Data are annualized, unadjusted, and for all domestic Y9C and call report BHC and bank filers. Net Charge Off Rate on Credit Card Loans with Non- w/ Linear Time w/ Exponential Preferred AR-1 Basic Specs Specifications Weighting Time Weighting Linearity 0.8607*** 0.8635*** 0.8616*** 0.8479*** 0.9196*** 0.8611*** Lagged NCO Rate on CC Loans (0.0596) (0.0547) (0.0499) (0.0506) (0.0509) (0.0540) 0.0109 Stock Market Quarterly Log Change (0.0068) -0.0462 Real GDP Annualized Log Change (0.0249) 0.3468*** 0.1955* 0.2711** 0.2876*** Annualized Change in Unemployment 0.3283*** (0.0713) (0.0925) (0.0783) (0.0852) (0.0661) 0.3158* Unemployment X <10% Dummy (0.1581) Unemployment <10% Dummy -0.3854 (0.4484) 0.6607** 0.6967* 0.8009** Constant 0.4194 0.6965** 0.8067* (0.2337) (0.2754) (0.3601) (0.2284) (0.2431) (0.2487) Observations 84 84 84 84 84 84 Adjusted R-squared 0.84 0.90 0.89 0.90 0.89 0.90 errors in parentheses. nks (Call) from 1991q1-2012q1. All institution flow variables are ann ualized. Robust standard Sample is for all BHCs (Y9-C) plus all other commercial ba * significant at 10%; ** significant at 5%; *** significant at 1% 19

81 Net Charge Off Rate on Other Consumer Loans 4 3 2 1 0 Net Charge Off Rate on Other Consumer Loans 2007q1 2011q1 1991q1 1993q1 2009q1 2001q1 2005q1 2003q1 1999q1 1997q1 1995q1 Historical Fitted Value Data are annualized, unadjusted, and for all domestic Y9C and call report BHC and bank filers. Other Consumer Loans Net Charge Off Rate on w/ Exponential w/ Linear Time with Non- Preferred AR-1 Basic Specs Time Weighting Specifications Linearity Weighting 0.7541*** Lagged NCO Rate on Other Consumer Loans 0.7464*** 0.5490*** 0.5439*** 0.5538*** 0.6180*** (0.1175) (0.1174) (0.1113) (0.1078) (0.1193) (0.0845) 0.0023 Stock Market Quarterly Log Change (0.0062) 0.0044 Real GDP Annualized Log Change (0.0170) 0.1581*** 0.0912* Annualized Change in Unemployment 0.1526*** 0.1536*** 0.1633*** (0.0246) (0.0431) (0.0263) (0.0350) (0.0328) 0.0101 Unemployment X <10% Dummy (0.0716) Unemployment <10% Dummy 0.3130 (0.1998) 0.0042 Time Trend 0.0219 0.0359** 0.0349** 0.0249 0.0319** (0.0125) (0.0123) (0.0109) (0.0114) (0.0142) (0.0121) -0.4331 -0.6635* 0.2043 Constant -0.5092 -0.8003* -0.7492* (0.3365) (0.3893) (0.3096) (0.4781) (0.3828) (0.3550) Observations 84 84 84 84 84 84 0.84 Adjusted R-squared 0.77 0.81 0.81 0.81 0.80 Sample is for all BHCs (Y9-C) plus all other commercial ba nks (Call) from 1991q1-2012q1. All institution flow variables are ann ualized. Robust standard errors in parentheses. * significant at 10%; ** significant at 5%; *** significant at 1% 20

82 Net Charge Off Rate on Leases 2 1.5 1 .5 Net Charge Off Rate on Leases 0 03q1 91q1 2011q1 2005q1 2007q1 2009q1 19 20 2001q1 1999q1 1997q1 1995q1 1993q1 Fitted Value Historical Data are annualized, unadjusted, and for all domestic Y9C and call report BHC and bank filers. Net Charge Off Rate on Leases w/ Exponential with Non- w/ Linear Time Preferred AR-1 Basic Specs Time Specifications Linearity Weighting Weighting 0.6414*** 0.6657*** Lagged NCO Rate on Leases 0.7403*** 0.6511*** 0.6790*** 0.6394*** (0.0787) (0.0955) (0.0754) (0.0813) (0.0950) (0.1021) Stock Market Quarterly Log Change 0.0049 (0.0033) 0.0002 Real GDP Annualized Log Change (0.0125) 0.0984*** 0.1211** 0.1044*** 0.1010*** Annualized Change in Unemployment 0.1113*** (0.0421) (0.0231) (0.0212) (0.0213) (0.0254) -0.0967 Unemployment X <10% Dummy (0.0627) 0.2391 Unemployment <10% Dummy (0.1637) Constant 0.1267** 0.1655*** 0.1610** 0.1600** 0.1740*** 0.1521** (0.0511) (0.0418) (0.0545) (0.0516) (0.0421) (0.0393) 84 84 84 84 Observations 84 84 0.75 Adjusted R-squared 0.54 0.63 0.62 0.62 0.67 nks (Call) from 1991q1-2012q1. All institution flow variables are ann ualized. Robust standard errors in parentheses. Sample is for all BHCs (Y9-C) plus all other commercial ba * significant at 10%; ** significant at 5%; *** significant at 1% 21

83 Net Charge Off Rate on Other Real Estate Loans 3 2 1 0 Net Charge Off Rate on Other Real Estate Loans 2001q1 2011q1 1991q1 1993q1 2009q1 2007q1 2005q1 2003q1 1995q1 1997q1 1999q1 Historical Fitted Value Data are annualized, unadjusted, and for all domestic Y9C and call report BHC and bank filers. Net Charge Off Rate on Other Real Estate Loans w/ Exponential w/ Linear Time with Non- Preferred AR-1 Basic Specs Time Weighting Specifications Linearity Weighting 0.5561** 0.5414*** 0.3619** Lagged NCO Rate on Other RE Loans 0.6988*** 0.2696 0.5544** (0.1331) (0.1654) (0.1426) (0.1675) (0.1291) (0.1476) Stock Market Quarterly Log Change 0.0006 (0.0043) 0.0455** Real GDP Annualized Log Change (0.0162) Annualized Change in Unemployment 0.0837 0.0173 (0.0563) (0.0608) -0.0067 -0.0110* CRE Price Index 4Q Log Change -0.0104* -0.0119* -0.0061 (0.0038) (0.0039) (0.0048) (0.0054) (0.0047) -0.0149 -0.0064 -0.0082 Change in CRE Index X <0 Dummy -0.0071 -0.0156 (0.0196) (0.0176) (0.0197) (0.0134) (0.0147) 0.1949** Constant 0.1206** 0.0712 0.1985** 0.1923** 0.2009** (0.0619) (0.0442) (0.0626) (0.0650) (0.0639) (0.0628) Observations 84 84 84 84 84 84 Adjusted R-squared 0.56 0.50 0.49 0.55 0.55 0.52 Sample is for all BHCs (Y9-C) plus all other commercial ba ualized. Robust standard errors in parentheses. nks (Call) from 1991q1-2012q1. All institution flow variables are ann * significant at 10%; ** significant at 5%; *** significant at 1% 22

84 Net Charge Off Rate on Loans to Depository Institutions 3 2 1 0 -1 1 Net Charge Off Rate on Loans to Dep. Inst'ns 2001q1 1991q1 2003q1 2005q1 2007q1 2009q1 2011q1 1993q1 1995q1 1997q1 1999q Fitted Value Historical Data are annualized, unadjusted, and for all domestic Y9C and call report BHC and bank filers. Net Charge Off Rate on Loans to Depository Institutions w/ Exponential with Non- w/ Linear Time Preferred AR-1 Basic Specs Time Weighting Specifications Linearity Weighting Lagged NCO Rate on Depository Loans 0.3403* 0.3283* 0.2934* 0.3156* 0.2634 0.3108* (0.1506) (0.1544) (0.1367) (0.1235) (0.1519) (0.1539) Stock Market Quarterly Log Change -0.0069 (0.0084) 0.0321 Real GDP Annualized Log Change (0.0362) 0.0521 0.0875 Annualized Change in Unemployment 0.0529 0.0526 0.0812 (0.0745) (0.0428) (0.0321) (0.0437) (0.0551) -0.1066 Unemployment X <10% Dummy (0.1294) 0.1878 Unemployment <10% Dummy (0.3128) Constant 0.1434** 0.1338*** 0.0707 0.1451** 0.1356** 0.1551* (0.0791) (0.0492) (0.0376) (0.0488) (0.0604) (0.0435) Observations 84 84 84 84 84 84 Adjusted R-squared 0.11 0.11 0.11 0.09 0.10 0.18 ualized. Robust standard Sample is for all BHCs (Y9-C) plus all other commercial ba errors in parentheses. nks (Call) from 1991q1-2012q1. All institution flow variables are ann * significant at 10%; ** significant at 5%; *** significant at 1% 23

85 Net Charge Off Rate on Agriculture Loans 1 .8 .6 .4 .2 0 Net Charge Off Rate on Agriculture Loans 91q1 2003q1 2005q1 2007q1 2009q1 1999q1 2001q1 19 1997q1 1995q1 1993q1 2011q1 Fitted Value Historical Data are annualized, unadjusted, and for all domestic Y9C and call report BHC and bank filers. Net Charge Off Rate on Agriculture Loans w/ Exponential with Non- w/ Linear Time Preferred AR-1 Basic Specs Time Specifications Linearity Weighting Weighting 0.5887*** 0.5773*** Lagged NCO Rate on Agriculture Loans 0.6274*** 0.5886*** 0.6082*** 0.5810*** (0.1458) (0.1442) (0.1436) (0.1439) (0.1551) (0.1664) Stock Market Quarterly Log Change -0.0002 (0.0027) -0.0024 Real GDP Annualized Log Change (0.0069) 0.0262* 0.0310 0.0242** 0.0288* Annualized Change in Unemployment 0.0253 (0.0201) (0.0116) (0.0088) (0.0125) (0.0166) -0.0321 Unemployment X <10% Dummy (0.0609) Unemployment <10% Dummy 0.1037 (0.2486) Constant 0.0799*** 0.0926** 0.0860*** 0.1121** 0.0935** 0.0864*** (0.0226) (0.0332) (0.0280) (0.0227) (0.0330) (0.0245) Observations 84 84 84 84 84 84 0.40 0.41 0.41 0.41 0.40 0.39 Adjusted R-squared errors in parentheses. ualized. Robust standard nks (Call) from 1991q1-2012q1. All institution flow variables are ann Sample is for all BHCs (Y9-C) plus all other commercial ba * significant at 10%; ** significant at 5%; *** significant at 1% 24

86 Net Charge Off Rate on Loans to Foreign Governments 30 20 10 0 -10 1 1 Net Charge Off Rate on Loans to For. Gov'ts 2001q1 2005q1 2007q1 2009q1 2011q 2003q1 1993q1 1995q1 1997q1 1999q 1991q1 Fitted Value Historical Data are annualized, unadjusted, and for all domestic Y9C and call report BHC and bank filers. Net Charge Off Rate on Loans to Foreign Governments w/ Exponential with Non- w/ Linear Time Preferred AR-1 Basic Specs Time Specifications Weighting Linearity Weighting 0.5053** 0.5709*** 0.5764*** 0.5741*** Lagged NCO Rate on For. Gov’t Loans 0.5572*** 0.2295** (0.1664) (0.1650) (0.0855) (0.1495) (0.1544) (0.1651) Stock Market Quarterly Log Change -0.0215 (0.0261) Real GDP Annualized Log Change 0.2871 (0.1603) 0.0485 0.6045 0.4835 Annualized Change in Unemployment -0.0002 0.1417 (0.0410) (0.0771) (0.5621) (0.1889) (0.2980) -0.7393 Unemployment X <10% Dummy (0.5773) 0.0741 Unemployment <10% Dummy (0.3816) -0.5632 0.1491 Constant 0.1579 -0.0081 -0.1791 0.3233 (0.2617) (0.3425) (0.3387) (0.1999) (0.2657) (0.1135) Observations 84 84 84 84 84 84 0.35 0.34 0.04 0.26 0.35 0.35 Adjusted R-squared nks (Call) from 1991q1-2012q1. All institution flow variables are ann errors in parentheses. Sample is for all BHCs (Y9-C) plus all other commercial ba ualized. Robust standard * significant at 10%; ** significant at 5%; *** significant at 1% 25

87 Net Charge Off Rate on Other Loans 2 1.5 1 .5 0 Net Charge Off Rate on Other Loans 91q1 03q1 19 2005q1 2007q1 2009q1 2011q1 2001q1 1993q1 20 1999q1 1997q1 1995q1 Fitted Value Historical Data are annualized, unadjusted, and for all domestic Y9C and call report BHC and bank filers. Net Charge Off Rate on All Other Loans w/ Exponential w/ Linear Time with Non- Preferred AR-1 Basic Specs Time Linearity Weighting Specifications Weighting 0.5568*** 0.6132*** 0.5562*** Lagged NCO Rate on Depository Loans 0.7042*** 0.5810*** 0.5989*** (0.1117) (0.1268) (0.1155) (0.1292) (0.1166) (0.1283) -0.0042 Stock Market Quarterly Log Change (0.0060) -0.0037 Real GDP Annualized Log Change (0.0259) 0.1087* 0.0805 Annualized Change in Unemployment 0.0924* 0.1179** 0.1014* (0.0455) (0.0456) (0.0373) (0.0413) (0.0405) 0.0049 Unemployment X <10% Dummy (0.1219) Unemployment <10% Dummy 0.1910 (0.3476) 0.1427** 0.1361** 0.1531** 0.1615* Constant 0.1056* 0.1731* (0.0860) (0.0483) (0.0514) (0.0474) (0.0746) (0.0420) 84 84 84 84 Observations 84 84 0.64 Adjusted R-squared 0.49 0.59 0.59 0.59 0.66 nks (Call) from 1991q1-2012q1. All institution flow variables are ann ualized. Robust standard errors in parentheses. Sample is for all BHCs (Y9-C) plus all other commercial ba * significant at 10%; ** significant at 5%; *** significant at 1% 26

88 t 90 (12) 0.951 -0.211 (0.234) (0.292) (0.413) (0.125) 0.959** 0.113** -0.588** (0.0467) 0.708*** + e t variables r each asset ng industry, Reverse Repos 90 (11) 0.955 0.104 (0.283) (0.447) (0.347) (0.113) 1.147** -0.0116 -0.0305 -0.0760 -0.722** (0.0123) (0.0930) (0.0593) (0.0448) 0.732*** Federal Funds Sold and 0.0937** (0.00822) 0.000763 ans, (iii) securities d for the remainder assets. Fo 90 (10) 0.976 (0.222) (0.134) 0.104** 0.581** -0.285** (0.0439) (0.0388) (0.0159) 0.907*** 0.0350** + d. other controls t-1 Other Assets 90 (9) 0.977 -0.313 -0.204 (0.158) (0.233) (0.197) 0.112** -0.0266 0.0363* 0.00129 (0.0467) (0.0422) (0.0488) (0.0204) 0.634*** 0.907*** -0.00333 (0.00741) (0.00915) r the commercial banki repos, (vi) other 90 (8) 0.964 -0.328 (0.144) (0.362) (0.241) -0.348** (0.0128) (0.0712) (0.0421) (0.0312) 1.383*** 0.736*** 0.102*** -0.0227* -0.0945** as % of total assets Trading Assets 90 (7) 0.965 -0.314 (0.175) (0.374) (0.308) -0.0266 -0.0132 -0.104** (0.0986) (0.0639) (0.0480) (0.0270) 1.436*** 0.738*** 0.100*** -0.486*** (0.00977) (0.00823) -0.0241** and one that includes only macroeconomic s sold and reverse asset composition fo largest firms in each calendar quarter) an 90 (6) 0.967 0.329* (0.129) (0.731) (0.190) 0.0124 0.268** (0.0357) (0.0464) (0.0139) 2.286*** 0.876*** 0.184*** (0.00821) -0.0468*** gory for the industry. + c . asset category 90 (5) t 0.968 27 (0.130) (0.968) (0.215) 0.0427 0.0133 0.324** 0.473** 0.0136* Securities (Ex. Trading) (0.0658) 2.670*** 0.855*** (0.0502) (0.0576) (0.0165) 0.195*** (0.00812) (0.00751) -0.0569*** total assets: (i) cash and interest bearing balances, (ii) lo ts, (v) federal fund regression models of 90 (4) 0.980 (6.826) (0.527) (0.110) (0.135) -1.030* 20.28*** 0.668*** (0.0827) -0.220*** -0.372*** :Q3, of the form: Loans rcentage points of total assets. ion for that asset cate 90 (3) 0.477 0.982 -0.803 -0.193 (4.544) (0.499) 0.0104 (0.417) (0.133) -0.225** 0.00418 15.08*** 0.749*** (0.0735) (0.0165) (0.0999) (0.0173) (0.0591) -0.159*** economic variables timated time series stry, for the subset of largest firms (10 90 (2) 0.391 0.454 0.916 (0.746) (0.535) (0.100) (0.251) with a large set of macroeconomic variables, 0.510** 0.931*** (0.0259) -0.00559 om 1991:Q1 to 2013 to maturity, (iv) trading asse Balances 90 = a + b . macro (1) t 0.321 0.323 0.921 0.0785 (0.809) (0.563) (0.110) (0.249) 0.550** -0.0235 0.00582 0.912*** (0.0277) (0.0480) (0.0546) -0.00882 in the initial specificat (0.00675) (0.00694) which together sum up to 100% of -0.000204 Cash and Interest Bearing ent variables are measured in pe specifications, one % of total assets bond spread (pct. pt) BBB 2009Q1) ≥ Relationship between asset composition and macroeconomic conditions Time trend (annual) Constant Lagged Dependent Variable Dummy ( Term Spread (10 year minus 3 months, pct. pt) Stock Market returns (quarterly, %) Quarterly change in Annualized change in Unemployment (%) Home price growth (%, year-over-year) 2 R Observations Other variables Macroeconomic variables All Robust standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1 asset category as We analyze six asset categories, based on Y-9C and Call Report data fr The tables on the following pages present es 2. classified as available for sale or held that were statistically significant A. Asset share regressions, Industry We present estimates for the aggregate indu of the industry. Depend category we show two

89 90 90 (12) (12) 0.884 0.933 -0.212 (0.470) (0.333) (0.358) (0.679) (0.246) (0.130) (0.117) 0.177** 0.929** -0.0705 (0.0727) (0.0705) 1.859*** 0.690*** 0.666*** -0.961*** -0.648*** (0.00763) -0.000527 Reverse Repos Reverse Repos 90 90 (11) (11) 0.889 0.937 0.147 -0.482 -0.106 (0.112) (0.161) (0.549) (0.397) (0.322) (0.783) (0.296) (0.110) 0.0800 (0.117) 0.137** 2.031** -0.0222 -0.0169 0.00297 0.00823 0.00241 (0.0625) (0.0429) (0.0178) (0.0701) (0.0141) (0.0383) 0.850*** 0.709*** 0.722*** Federal Funds Sold and Federal Funds Sold and -0.00349 -0.00361 -0.00163 -1.199*** (0.00590) (0.00628) (0.00852) 90 90 (10) (10) 0.934 0.987 (0.201) (0.178) (0.349) (0.103) 0.136** 0.453** -0.0693 -0.432** (0.0616) (0.0181) (0.0321) (0.0423) (0.0336) 0.923*** 0.885*** 0.913*** 0.0402** 0.0730** (0.00911) 0.0259*** Other Assets Other Assets 90 90 (9) (9) 0.935 0.989 -0.114 -0.355 (0.297) (0.232) (0.197) (0.114) (0.345) 0.0357 0.0179 0.167** -0.214* -0.0404 0.00676 (0.0759) (0.0748) (0.0229) (0.0276) (0.0109) (0.0725) (0.0423) (0.0148) (0.0459) (0.0146) (0.0345) 0.956*** 0.891*** 0.882*** 0.535*** -0.00216 0.0632** -0.00417 -0.256*** -8.54e-05 (0.00326) (0.00457) 0.0410*** 90 90 (8) (8) 0.864 0.672 -0.111 0.192* (0.386) (0.101) (0.241) (0.827) (0.172) (0.146) 0.0266 -0.0816 -0.0258 0.0437* -0.623** -0.414** (0.0867) (0.0338) (0.0188) (0.0883) (0.0584) (0.0246) 2.735*** 0.823*** 0.515*** 0.246*** 0.0228** (0.00411) (0.00982) -0.000348 Trading Assets Trading Assets 90 90 (7) (7) 0.673 0.869 -0.107 -0.120 -0.123 0.187* (0.551) (0.186) (0.249) (0.108) (0.775) (0.184) (0.137) 0.0262 0.0424 -0.410** 0.00757 -0.0293* (0.0957) (0.0358) (0.0335) (0.0138) (0.0870) (0.0525) (0.0262) (0.0151) (0.0102) 2.814*** 0.826*** 0.505*** 0.260*** 0.0227** -0.939*** (0.00405) (0.00287) 0.000264 -0.0333** -0.000537 90 90 (6) (6) 0.965 0.840 0.455* 0.151* 0.0217 (0.303) (0.261) (0.124) (0.182) (0.584) (0.766) 0.0127 0.0883 1.348** -0.0370 0.00577 (0.0761) (0.0206) (0.0523) (0.0340) (0.0459) (0.0142) (0.0111) 0.591*** 0.175*** 0.863*** 0.865*** 3.257*** (0.00697) -0.0635*** 90 90 (5) (5) 0.968 0.844 0.113 0.172 28 0.0228 (0.352) (0.263) (0.126) (1.141) (0.683) (0.224) 0.0679 0.0221 0.0142 0.0466 0.525** 0.178** 1.800** Securities (Ex. Trading) Securities (Ex. Trading) 0.00743 (0.0955) (0.0855) (0.0583) (0.0230) 0.199*** 0.817*** (0.0136) 0.825*** 0.818*** (0.0103) (0.0672) (0.0495) (0.0523) (0.0137) 4.112*** (0.0178) 0.0214** (0.00633) -0.0803*** 90 90 (4) (4) 0.968 0.870 -0.188 -0.253 -0.798 (0.796) (0.163) (6.164) (5.972) (0.419) (0.132) (0.177) (0.104) 0.0206 15.45** 15.02** -0.414** -0.266** (0.0933) 0.762*** 0.723*** (0.0223) Loans Loans 90 90 (3) (3) 0.891 0.973 -0.619 -0.321 1.290* 0.0245 (0.802) (0.263) (0.128) (0.467) (5.240) (3.985) (0.181) (0.442) (0.722) (0.148) 0.0643 -0.780* -0.0653 -0.0119 0.0430* -0.896** 13.58*** 0.751*** 0.714*** (0.0841) (0.0697) (0.0829) (0.0228) (0.0108) (0.0162) 17.97*** (0.0272) (0.0232) -0.00707 -0.00881 -0.500*** -0.235*** 90 90 (2) (2) 0.331 0.881 0.895 0.152 -0.200 (0.532) (0.769) (0.649) (0.907) (0.103) (0.120) (0.199) (0.368) 0.0138 0.523*** 1.011*** (0.0282) (0.0289) 3.477*** 2.374*** 0.993*** -0.100*** Balances Balances 90 90 (1) (1) 0.911 0.897 -0.363 -0.112 0.357* (0.802) (0.501) (0.559) (0.761) (0.112) (0.108) (0.296) (0.214) 0.0341 0.0219 -0.0372 0.00614 0.00499 0.334*** 1.021*** 0.211*** 1.098*** 2.708*** 4.356*** (0.0295) (0.0226) (0.0530) (0.0562) (0.0963) (0.0490) -0.00803 -0.00917 -0.116*** (0.00696) (0.00856) (0.00825) (0.00694) -0.0268*** Cash and Interest Bearing Cash and Interest Bearing er than 10 largest firms) bond spread (pct. pt) bond spread (pct. pt) BBB BBB 2009Q1) 2009Q1) ≥ ≥ Constant Constant Time trend (annual) Time trend (annual) Lagged Dependent Variable Lagged Dependent Variable Dummy ( Dummy ( Term Spread (10 year minus 3 months, pct. pt) Term Spread (10 year minus 3 months, pct. pt) Stock Market returns (quarterly, %) Stock Market returns (quarterly, %) Quarterly change in Quarterly change in Home price growth (%, year-over-year) Annualized change in Unemployment (%) Home price growth (%, year-over-year) Annualized change in Unemployment (%) 2 2 R Observations Observations R Other variables Other variables Macroeconomic variables Macroeconomic variables Top 10 Bottom Robust standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1 Robust standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1 C. Remainder of industry (oth B. Large firms only (aggregate of 10 largest banking firms in each calendar quarter)

90 D. Hypothesis tests for asset share regressions esents f-tests of the joint significance of the The first column of results in the table below pr macroeconomic variables in table A (the asse t share regressions for the entire industry). The second column of results presents f-test of the equality of the short-run coefficients on the of the industry (table macroeconomic variables for large firms (table B a bove) versus the remainder C above). F-Statistics [p-value] Joint Significance Specification Dependent of Macro Variables Variable No. ≠ Big Small 1 0.166 0.068* Cash and Interest Bearing Balances 2 0.045** 0.115 0.010** 0.047** 3 Loans 4 0.007*** 0.350 5 0.000*** 0.902 Securities (Ex. Trading) 6 0.000*** 0.627 7 0.014** 0.001*** Trading Assets 8 0.022** 0.009*** 9 0.007*** 0.808 Other Assets 10 0.020** 0.366 11 0.088* 0.017** Federal Funds Sold and Reverse Repos 12 0.014** 0.010*** 29

91 3. Loan Losses, Reserves and Provisions: Additional Discussion Rather than simply expensing credit losses when they are finally realized (e.g. when the property osure auction), BHCs and banks reserve in advance securing a delinquent mortgage is sold at a forecl future probable credit losses on their loan portfolio, in accordance with supervisory rules and against Generally Accepted Accounting Principles. Accounting for loan and lease losses involves three closely related measures: Net charge-offs (NCOs): NCOs are the credit losses realized by the firm in the current  accounting period, net of any recoveries (that is, net of any payments received on loans previously viewed as uncollectible). The reserve held by the firm against “estimated Allowance for loan and lease losses (ALLL):  credit losses”, that is, losses that have not yet occurred but are “likely to be realized” in the 6 The ALLL is recorded as a contra-asset on the firm’s balance sheet. future. Provision expense for loan and lease losses: The expense incurred in the current accounting  period in order to set aside additional reserves against future loan losses. Note that ALLL is a stock , while NCOs and provision expense are flows . The ALLL represents the existing stock of reserves. The realization of NCOs reduces the ALLL over time, while provision hus, there is a mathematical identity between these expenses incurred by the firm increase the ALLL. T three accounting variables for a given firm: + Provision for loan and lease losses = ALLL – net chargeoffs ALLL t t t t-1 NCOs do not directly affect net income, but have an important indirect effect on the income statement, since (as seen in the above equation) higher NCOs mu st be offset by a higher provision expense in order to keep the level of loan loss reserves at a given target level. Determining the “appropriate” level of loan loss reserv es is inherently subjective, since it relies on an assessment of future probable credit losses. As a starting point, supervisory standards state that firms 6 As stated in the Federal Reserve Bank Holding Company Supervision Manual (Federal Reserve Board of Governors, 2013), “the term estimated credit losses means an estimate of the current amount of loans that it is probable the institution will be unable to collect given facts and circumst ances since the evaluation date. Thus, estimated credit losses represent net charge-offs that are likely to be realized for a loan or group of loans”. Each BHC and bank is required to vel of loan loss reserves. See also Statement of Financial maintain and apply a consistent process for assessing the le Accounting Standards No. 5 (Financial Accounting Standards Board, 1975), the accounting standard which deals with loss contingencies. 30

92 should generally set aside reserves for each category of loans at least equal to the annualized (12-month) 7 Relative to this benchmark, however, the historical NCO rate for loans with similar risk characteristics. appropriate level of reserves should also take into account environmental factors expected to cause losses to differ from historical experience, such as shifts in economic conditions and lending standards. Firms are also expected to hold additional reserves for loan types with effective lives greater than 12 months and extended workout periods, such as certain types of commercial loans. Provision Expense and NCOs (% of total loans) 4 3 2 1 0 1990q1 1995q1 2000q1 2005q1 2015q1 2010q1 Net Chargeoffs Provision Expense The above graph plots the historical behavior of annu alized net chargeoffs and provision expense, scaled by total loans. As the figure shows, these two variables move together closely up to 2006, but then diverge sharply during the Great Recession. During 2007-08, provision expense increased much more quickly than realized NCOs, reflecting BHC expectations of high future NCOs to come. Conversely, as the economy recovered, provision expense fell significantly in advance of NCOs. quarterly net chargeoffs Next we plot the level of the ALLL over time, alongside the sum of realized annualized NCO rate over the entire sample period average over the following 12 months, as well as the (Box Graph 2). ALLL as a percentage of total loans trends downwards significantly between 1990 and 7 See section 2065.3 of “Bank Holding Company Supervision Manual.” Division of Banking Supervision and Regulation. Available online at: http://www.federalreserve.gov/boarddocs/supmanual/supervision_bhc.htm 31

93 2006, before rising sharply during the recession. ALLL remains elevated today compared to pre-crisis levels. Note that actual ALLL almost always lies above the level predicted by a naïve application of the “12-month” rule (either based on average NCOs over the entire sample period, or realized future NCOs tent with supervisory guidance, which recommends over the subsequent 12 months). This appears consis that ALLL should in general be at least equal to the annualized NCO rate (i.e. the 12-month rule is an approximate lower bound). ALLL and realized net chargeoffs (% of total loans) 4 3 2 1 0 2015q1 2010q1 2005q1 2000q1 1995q1 1990q1 Next 12-months NCOs Average 12-months NCOs ALLL Assumptions for Reserves and Provision Expense in the CLASS Model Any accounting-based stress testing model, including CLASS, must include an assumption about the loan loss provisioning rule used by firms. Note that for a given path of NCOs, we make an assumption ALLL or provision expense (but not both, since they are identically linked). about either vior. Our code allows us to toggle between them We consider three different rules for provisioning beha when calculating the capital projections. Provisions = NCO. 1. Under this approach, banks simply expense whatever net charge-offs were incurred in the current quarter. This is equivalent to assuming that loan loss reserves remain constant 32

94 over the stress test horizon. As can be seen above, this rule is a reasonable approximation to bank behavior over the period 1991-2006, but has been less so since the financial crisis. ALLL = next four quarters of projected NCOs under the macro scenario. This rule is motivated 2. by the supervisory recommendations that reserves be generally at least equal to the historical annualized NCO rate, and that the level of reserves should be sensitive to macroeconomic conditions. Under this approach, we first forecast total NCOs in each quarter over the stress test horizon. We then calculate loan loss reserves as being equal to the next four quarters of net charge- = NCO + [LLR - offs. We then calculate provision expense in each quarter as: Provision expense t t t ]. LLR t-1 3. “Tunnel”: Provisions = NCO, as long as LLR are within a range of projected NCOs. Under this hybrid approach, which is the approach generally employed in CLASS in practice (and used for the projections in this paper), provisions are set equal to current NCOs (as in rule 1), as long as provisions do not deviate “too far” from being equal to the sum of future NCOs. In particular we use a range of 100%-250% of projected 12-month NCOs.  ALLL = NCOs: NCOs (and compute provisions accordingly)  ALLL < 4 4 ALLL > 250% *  NCOs:  NCOs (computing provisions accordingly) ALLL = 2.5 * 4 4  NCOs < LLR < 250% *  NCOs: Provision expense = NCOs 4 4 The key disadvantage of approach 1, is that setting provision expense equal to current charge-offs is not forward looking. For example, heading into a r ecession, loan loss reserves should increase in expectation of higher NCOs in future quarters, even if those charge-offs have not yet occurred. In contrast, approaches 2 and 3 are forward looking. An important limitation of approach 2 (setting ALLL e qual to the next four quarters of NCOs), is that this rule understates historical ALLL, especially during non-recession periods, as shown in the graph above. Approach 3, the tunnel approach, provides one way to deal with this issue, since it allows reserves to be as much as 250% of the sum of the next four quarters of projected NCOs. One issue that arises under approaches 2 and 3 is that projected ALLL is calculated using a behavioral rule that is likely to differ from the most recent hist orical value of ALLL. For example, as can be seen under a baseline scenario, the four quarter rule (approach 2) will generally imply a lower projected ALLL than the most recent historical value of ALLL. And under a stress scenario, the starting projected 33

95 ALLL may be higher than its most recent historical value. Thus, a “true-up” is required to shift ALLL up or down to its value as computed under the provisioning rule, in turn implying a one-off spike (upward or downward) in provision expense. To avoid such a spike, we smooth the shock to provision expense implied by the true-up over the duration of the scenario. For example, imagine we are applying the four-quarter rule to a firm with ALLL = $100m as of the most recent historical date. Under the macro scenario being projected, the ALLL as calculated by the tunnel rule at this date is $145m. As a result, a “true-up” of $45m is applied to the ALLL to bring it up to its model-implied value, consequently implying an additional $45m in provision expense. This $45m is spread evenly over the scenario (e.g., under a nine quarter scenario, it would be applied in $5m increments in each quarter of the scenario). The default provisioning rule used by the CLASS model is the “tunnel” rule described above. This simple rule could easily be made more sophisticated in future versions of the CLASS model, either by using a formula that is dependent on the composition of the firm’s loan portfolio, or by estimating a formal econometric model of ALLL or provision expense. References Financial Accounting Standards Board (1975), “Financial Accounting Standard 5: Accounting for Contingencies.” Federal Reserve Board of Governors (2013), “Bank Holding Company Supervision Manual.” Division of Banking Supervision and Regulation. Available online at: http://www.federalreserve.gov/boarddocs/supmanual/supervision_bhc.htm 34

Related documents