FEDERAL RESERVE BANK OF SAN FRANCISCO WORKING PAPER SERIES De-leveraging or de-risking? How banks cope with loss Rhys M. Bidder, John R. Krainer, Adam H. Shapiro Federal Reserve Bank of San Francisco June 2017 Working Paper 2017-03 http://www.frbsf.org/economic-research/publications/working-papers/wp2017-03.pdf Suggested citation: Bidder, Rhys M., John R. Krainer, Adam H. Shapiro. 2017. “De-leveraging or de-risking? How banks cope with loss” Federal Reserve Bank of San Francisco Working Paper 2017-03. http://www.frbsf.org/economic-research/publications/working-papers/wp2017-03.pdf The views in this paper are solely the responsibility of the authors and should not be interpreted as reflecting the views of the Federal Reserve Bank of San Francisco or the Board of Governors of the Federal Reserve System. De-leveraging or de-risking? How banks cope with loss Rhys M. Bidder, John R. Krainer, Adam Hale Shapiro Federal Reserve Bank of San Francisco Abstract Using detailed bank balance sheet data we examine how banks respond to a net worth shock. We make use of variation in banks’ loan exposure to industries adversely affected by the oil price declines of 2014 and the implied variation in losses resulting from credit deterioration in those industries. In response to these losses, exposed banks reduced the risk of their balance sheets by shifting away from portfolio lending and towards assets with lower risk weights. Banks tightened credit on corporate lending and on mortgages that they would ultimately hold in their portfolio. However, they expanded credit for mortgages to be securitized and increased their holdings of agency mortgage backed securities. Our results imply that previous work suggesting that banks tighten credit in response to a shock (the traditional ‘bank lending channel’ view) provides only a partial story and is in some ways misleading. It appears that banks respond to a negative shock by de-risking rather than a uniform reduction in lending. In terms of the ultimate impact on borrowers, we find that the shock had only a minimal impact on the overall quantity of loans supplied to firms or households, reflecting substitution to other sources of financing. The findings have significant implications for regulatory policy and theories of financial frictions. I Disclaimer: The views expressed in this document and all errors and omissions should be regarded as those of the authors and not necessarily those of the Federal Reserve Bank of San Francisco, the Federal Reserve Board of Governors, or the Federal Reserve System. This paper has been screened to ensure that no confidential bank or firm-level data have been revealed. We thank Tesia Chuderewicz for excellent research assistance and Anita Todd for editorial assistance. We also thank seminar participants at Cardiff University, Warwick University, Bank of England, ECB, Federal Reserve Board of Governors, University College London, the Bay Area Finance Workshop, Fed System Micro Conference (especially, Mitchell Berlin), Chicago Fed, U.C. Davis, Econometric Society (Hong Kong). Email addresses: [email protected] (Rhys M. Bidder), [email protected] (John R. Krainer), [email protected] (Adam Hale Shapiro) 1. Introduction The recent financial crisis provided a striking example of how intertwined are the fortunes of banks and the wider economy. Even before the crisis, however, an extensive empirical literature had emerged documenting the importance of financial intermediaries and various theories had been proffered to explain why the capital structures of firms and banks appeared to influence the financing of projects, in contrast to the stark implications of the Modigliani-Miller theorem. The literature has mainly emphasized a tendency for credit supply to shrink following an adverse shock to a bank’s net worth or funding - often referred to as the ‘bank lending channel’.1 In this paper we show that this traditional view is only a partial characterization and part of a more nuanced set of responses. Our results show that, following a shock, there are important compositional changes in credit supply. Some types of credit do indeed contract. However, credit supply expands in other areas, such that the overall change is ambiguous. Furthermore, we show it is fruitful to expand analysis to look across the whole balance sheet - rather than simply examining lending. Although we focus much of our analysis on corporate and mortgage loans, we also identify coordinated effects outside the loan book on banks’ securities holdings — securities and lending portfolio decisions are inextricably connected. In response to the shock, banks ‘de-risk’ as opposed to de-leveraging in the traditional sense of reducing their scale of operations. They move out of corporate and some types of mortgage lending, into issuing mortgages that can be securitized and exchanged for the less heavily risk weighted (and more liquid) agency MBS. In what follows, reflecting the use of risk weighted assets in calculating regulatory capital ratios (risk adjusted measures of leverage), we will refer to this behavior as ‘de-leveraging by de-risking’ since risk weighted assets were not reduced so much through total assets but through reductions in the ‘average’ risk weight. 1 This channel is based on the observation that because of externally imposed minimal capital ratios or constraints on bank leverage that emerge endogenously from information frictions between banks and their investors, banks’ net worth will influence their ability to supply credit (Holmstrom and Tirole (1997), Chen (2001) Van den Heuvel (2008), Meh and Moran (2010), Gertler and Karadi (2011), Iacoviello (2015), Bocola (2016)). Associated with this argument is the idea that losses of liquidity by banks, regardless of whether they initially arise from damage to their net worth, cannot frictionlessly be offset with alternative sources of finance, again necessitating an adjustment to their lending or investment behavior (Bernanke and Blinder (1992), Kashyap and Stein (1993), Stein (1998)). 2 We exploit a particular natural experiment that induced a shock to net worth that varied across banks. The shock we analyze stems from the precipitous decline in the price of oil in mid-2014, which led many firms in the oil and gas (O&G) industry to delay or cease payment on their loans. Leading up to the oil price decline, banks held varying amounts of loans to firms in the O&G sector so that, implicitly, the shock induced variation across banks in the impact of the price decline. Beginning with corporate lending, we find evidence consistent with the traditional ‘bank lending channel’, on both the intensive and extensive margins. Lending from banks that were particularly afflicted by the oil shock declined significantly, and the probability of their relationship with a firm ending increased substantially. These results are consistent with patterns elsewhere on the balance sheet and also chime with bank-level data from lending surveys, indicating that much of this decline was due to tightened standards inducing firms to substitute to other sources of financing. A tightening of credit supply is also, to some extent, observed in banks’ mortgage lending. Although in this case the story is more subtle and suggests that the pull back in lending may more fruitfully be thought of in terms of a more general de-risking of the balance sheet. Banks who were more exposed to the oil shock tightened credit for mortgages that they would be holding on their balance sheet, similar to the commercial loan results. However, simultaneously, these banks also appeared to be loosening credit supply for those mortgages (primarily GSE-eligible) that would predominantly be securitized and thus shifted off their balance sheet in exchange for mortgage-backed securities (MBS). The lower risk weighting of the MBS, relative to the original underlying loans, are such that the overall risk profile of the balance sheet is reduced - which we show is consistent with aggregate data from public filings: affected banks’ MBS holdings increase dramatically (relative to unexposed banks), while their retained residential loan balances do the opposite. Turning to the ultimate impact of the banks’ decisions on borrowers, we find little evidence of a ‘credit channel’.2 Even though exposed banks seem to have tightened credit, the 2 The presence of a bank lending channel does not in itself imply that the state of banks’ balance sheets, ceteris paribus, will influence the broader economy. For this, it must be the case that fluctuations in bank lending cannot be offset by borrowers through substitution to alternative sources of external finance-from other banks or the capital markets. This latter step is often referred to as the credit channel. If there are easily accessible alternatives to bank finance in the public debt or equity markets, then the bank channel, even if it exists, may have a limited impact on firm activity (Whited (1992), Greenspan (1999), Sufi (2009), 3 firms that were borrowing from these banks prior to the shock appear to have been able to substitute for the vast majority of any lost financing via switching to alternative, less exposed banks in our sample, or by drawing on existing credit lines. This appears to have particularly been the case for firms with multiple bank relationships. Analogously, we show that in counties with a high indirect exposure to the oil shock (via exposed banks), there is apparently no difference in the overall level of portfolio mortgage loans (the type on which affected banks were individually pulling back) relative to counties without exposure. Understanding these subtleties of bank portfolio re-optimization is of enormous importance for various regulatory purposes (particularly stress testing and capital requirements), debates over bailouts and also the role of government sponsored enterprizes (GSEs). The policy implications of our results are clear in some respects - equity injections or ‘bailouts’ are essentially a mirror of the sort of negative net worth shock we consider here. However, in addition to this, it is notable that even though our shock originated in the corporate loan book, important adjustments are made through the mortgage book and related securities (MBS) holdings. To the extent that GSEs underpin this low cost technology for adjustment after a negative shock (thereby affecting the consequences of lending decisions), they may have broader moral hazard effects on bank behavior, beyond those already discussed in the literature. The data used in this study are from the FR Y-14 (Y14) regulatory filings newly required under post-crisis reforms. They contain loan-level information identifying both bank and borrower, describe various characteristics of the loan and provide financial information about the borrower. We go beyond traditional analyses by augmenting our commercial loan data with banks’ residential lending and connecting them to the banks’ securities (MBS) holdings. Our granular loan level data (featuring firm borrower identifiers) allows us to track the same firm across multiple banks. This means, as in Khwaja and Mian (2008), that we can exploit variation in lenders’ exposure to oil while controlling for credit demand and other effects that might confound inference on credit supply movements. The remainder of the paper is organized as follows: Section 2 relates our study to the literature. Section 3 describes the Y14 data, the source of the net-worth shock, and the econometric specifications. Section 4 shows the results of our analysis. Section 5 provides Lemmon and Roberts (2010)). 4 details on the robustness of our results. Section 6 discusses policy implications. Section 7 concludes. 2. Relation to literature While the cross-balance sheet aspect of our analysis takes us beyond (and in some ways puts us in tension with) the standard ‘bank lending channel’ literature, our work does include analyses of bank lending. As such, for our corporate and mortgage loan results, we draw on tools used in existing work. In particular, to isolate bank credit-supply effects from our shock we use the fixed-effects identification approach taken by Khwaja and Mian (2008) (KM), but making novel use of granular data on the balance sheets of the largest U.S. bank holding companies. This approach requires loan-level data which allows us to analyze the same firm borrowing simultaneously from multiple banks that vary in their balance sheet’s exposure to an exogenous shock. Firm fixed effects capture demand shifts and allow us to separately identify the shock’s impact on credit supply. This powerful technique relies on the (demanding) requirement that one has access to a credit-register type dataset.3 Overall, in the literature there appears to be consistent evidence of the bank lending channel - whether in response to explicit disruptions in liquidity or via damage to bank net worth that makes funding more difficult. With regard to the credit channel, the evidence is less clear and more nuanced. A common theme is that if the credit channel is present, it particularly applies to small firms and firms with relatively few alternative financing options. 3 Early tests used data at relatively high levels of aggregation—economy-wide or at the bank level (Bernanke and Lown (1991), Kashyap, Stein, and Wilcox (1993), Hancock and Wilcox (1993), Peek and Rosengren (1995), Stein and Kashyap (2000), Ashcraft (2005), Ashcraft (2006), Calem, Covas, and Wu (2013)). However, as noted by Khwaja and Mian (2008), such studies suffer from the difficulty of separating credit demand and supply effects of shocks to firms and banks. Ideally, one would wish to examine the effect of an isolated shock to bank balance sheets but these shocks often will simultaneously influence firms’ demand for credit, making identification difficult. Even if one can identify an exogenous shock that might plausibly be thought only to affect credit supply (see, for example, Peek and Rosengren (1997), Paravisini (2008), Ivashina and Scharfstein (2010), Chava and Purnanandam (2011), Greenstone, Mas, and Nguyen (2014)), there remains the concern that correlated credit demand and supply shifts might confound estimates of the lending channel. Notable papers that use the KM technique to avoid these problems are: Schnabl (2012), Chodorow-Reich (2014), Jimenez, Mian, Peydro, and Saurina Salas (2014), Iyer, Peydro, da Rocha-Lopes, and Schoar (2014), Acharya, Eisert, Eufinger, and Hirsch (2014), Cingano, Manaresi, and Sette (2016), Bottero, Lenzu, and Mezzanotti (2016), De Jonghe, Dewachter, Mulier, Ongena, and Schepens (2016) and Heider, Saidi, and Schepens (2017) 5 Many of these papers use data from economies that are known to be bank dependent and/or have less developed capital markets than the United States.4 Thus their relevance to the U.S. case, with its deep capital markets, might be questioned. Alternatively, the relevance of the work may not be universal due to the nature of the funding being examined - trade finance in the case of Amiti and Weinstein (2011) and exclusively syndicated loans in the case of Acharya, Eisert, Eufinger, and Hirsch (2014), Chodorow-Reich (2014) and Chodorow-Reich and Falato (2017). While syndicated lending is an important source of bank financing, one might worry that the information it reveals is incomplete.5 In contrast to Chodorow-Reich (2014), we analyze a broader class of loans (not simply syndicated) sampled from a time period in which the broader financial system was not under stress. While it is certainly of enormous interest to know how the bank lending and credit channels were operating during the financial crisis, and around the time of the Lehman Brothers failure in particular (as in Chodorow-Reich (2014)), it is possible that such an environment may limit the external validity of some of the results. A similar concern could be raised also for Acharya, Eisert, Eufinger, and Hirsch (2014), who analyze the period of the European sovereign debt crisis. In our case, we will analyze the effects of shocks to banks at times when there are no worries about the collapse of the financial system. This is clearly of great importance when assessing the credit channel as some of the effects found in extant work may have reflected a severe but unusual lack of alternative sources of financing. Thus we regard our work as complementary to the literature based on financial crises, in speaking to the importance of the broad credit channel in more normal times (see also Jimenez, Ongena, Peydro, and Saurina (2014)). The precise alignment between our results based on our balance sheet data and those from the lending surveys also provides important confirmation of the reliability of survey data for tracking bank behavior, as in Ciccarelli, Maddaloni, and Peydro (2015), for example. This enhances confidence in these series as (qualitative) observations of banks’ credit supply schedules. 4 Jimenez et al. use Spanish data, Iyer et al. use Portuguese, Cingano et al. and Bottero et al. use Italian, Schnabl uses Peruvian, Khwaja and Mian use Pakistani, and Acharya et al. and de Jonghe et al. use European. 5 Our dataset is not a superset of that used by Chodorow-Reich, although the Y14 does provide information that allows one to identify if loans are syndicated. 6 Our results on banks’ changing of their risk (weight) profile in response to the shock provide an important extension to the traditional empirical literature testing for the importance of net worth, as predicted by financial frictions models. In a world where regulatory capital ratios are frequently defined with respect to risk weighted, rather than total, assets, understanding how banks’ respond along this margin is vital. Indeed, it points to interesting extensions of models in the class of Holmstrom and Tirole (1997) to accommodate such issues and perhaps also address issues of regulatory risk-weight arbitrage (Das and Sy (2012), Duchin and Sosyura (2014), Efing (2015)). To the extent that the agency MBS are also highly liquid assets one can also partly regard our de-risking as encompassing an attempt to mitigate liquidity risk, particularly since there is a powerful nexus between solvency and liquidity (see Pierret (2015) and Bank for International Settlements (2015), for example). Models that endogenize bank portfolio responses to shocks, particularly in the context of stress testing are currently in high demand (Borio, Drehmann, and Tsatsaronis (2012), Halaj (2013) and Halaj (2016)). Our results, though derived from a ‘normal shock in normal times’, should provide some scope to help calibrate such models and reveal the subtle details of bank adjustments. With regard to the interrelation of various regulatory ratios (capital and liquidity, risk weighted and unweighted) in post crisis reform, the distinctions we draw between credit tightening and de-risking as forms of adjustment also allude to important subtleties to consider (see Cohen (2013) and Cecchetti and Kashyap (2016) for empirical and theoretical discussions of these issues). 3. Empirical approach In this section we describe the novel data set that we will be using. Next, we explain how we exploit a shock to bank net worth and then specify our econometric approach to ascertain if this shock induced a credit supply shift (bank lending channel) and whether this influenced borrower outcomes (credit channel). 3.1. The Y14 Data We make use of data collected from the quarterly FR Y-14Q and monthly FR Y-14M filings required of bank holding companies (BHCs) with more than $50 billion in assets. The Y14 has been filed since 2012 and contains detailed data on the banks’ balance sheet exposures, capital components and categories of pre-provision net revenue. The primary 7 purpose of the filing is to help assess the capital adequacy of banks, in support of supervisory stress testing programs that the Federal Reserve is required to run under the Dodd-Frank Act. The commercial loan data are drawn from the quarterly corporate loan schedule in the FR Y-14Q. This schedule provides a ‘credit register’ — loan-level information identifying borrower and lender and many characteristics of the loans themselves. Since this paper is one of the first to employ the Y14 data, we provide some perspective on how much of the market for corporate lending we are able to capture.6 Table 1 shows that in 2014:Q2, just prior to oil prices’ precipitous decline, the commercial loans included in the Y14 totaled $1.17 trillion, or nearly 70 percent of the $1.69 trillion in commercial and industrial loans extended by the entire BHC population that files a Y-9C report. For perspective, the total amount of debt owed by U.S. nonfinancial corporations at that time was $2.39 trillion. Compared to other bank loan data sets used in the literature (e.g., syndicated loans data from the Shared National Credits Program (SNC) or DealScan), the Y14 also captures loans that are typically smaller in size and to smaller-scale borrowers. Thus, the dataset contains loans that are more likely to be closer to the theoretical objects in models of financial frictions, in that they are from smaller firms and thus more likely to be subject to credit constraints. The FR Y-14M filings contain loan-level monthly data on banks’ mortgage positions, including characteristics of the borrower, the mortgage and the ‘investor type’. The latter classification allows us to determine whether a given mortgage loan in the dataset is actually held on the balance sheet by the bank as part of its own asset portfolio (what we will refer to as a ‘portfolio loan’), or whether it is instead a loan that is owned by some other party perhaps after having been securitized (typically by a GSE, although some loans are also privately securitized). In this case the bank would only be reporting the loan in the Y14 due to its role in servicing the mortgage, for example. Our sample of commercial loans consists of loans originated by 28 large BHCs and our sample of residential loans consists of 26 banks. In table 2 we illustrate bank characteristics 6 Other papers that have used these novel regulatory data are Abdymomunov (2014) (operational losses), Black, Krainer, and Nichols (2016), Johnson and Sarama (2015), Epouhe and Hall (2016), Calem and Sarama (2016) (real estate) and, making limited use of the corporate loans data to assess the use of private information by syndicate leaders, Balasubramanyan, Berger, Bouwman, and Koepke (2016). In contemporary ongoing work Berrospide and Edge (2016) use loan data in examining the effects of capital regulations. 8 for each sample, based on their financial statement data reported in the FR Y-9C (Y9) prior to the decline in oil prices. These variables will be used as bank controls in our commercial and residential loan analysis below. 3.2. The oil price decline of 2014 Our analysis exploits a natural experiment. Specifically, we examine variation in bank behavior induced by the dramatic oil price decline of 2014. As illustrated in figure 1, the drop in oil prices was both sudden and significant. Furthermore, oil futures prices clearly indicate that, although a moderate pullback in prices might have been expected, the speed and extent of the decline was a surprise. Indeed, much commentary in the early part of 2014 was bullish on oil, such as Hamilton (2014). Performance of loans to firms in the oil and gas (O&G) industry depend critically on the behavior of oil prices. As shown in figure 1, the rate of loans identified as either past due, charged-off, or a non-accrual (what we refer to as ‘problem loans’) spiked in the O&G industry following the oil price decline. We define firms as being in the O&G industry if the bank reported NAICS codes 211, 213111, and 213112 (oil and gas extraction, drilling oil and gas wells, and support activities for oil and gas operations).7 Figure 2 shows that these NAICS codes are obvious outliers in terms of being ‘net makers’ of oil-related products, based on the ‘make’ and ‘use’ tables provided by the Bureau of Economic Analysis (BEA).8 Returning to figure 1, we see that the fraction of O&G loans that were in problem status climbed from 0.6 percent in 2014:Q2 to 10.4 percent in 2016:Q3. The rates in other sectors, by contrast, do not exhibit such an increase. Taking all sectors other than O&G we observe essentially no trend. Restricting the sample to net users of O&G products (where net users are defined as firms in industries with net trade shares less than −0.1 percent - depicted on the left hand side in figure 2) we again observe no obvious pattern. This pattern among net O&G users emphasizes that there was apparently little scope for passive hedging of the 7 Some banks report SIC or GICS codes, but the vast majority report NAICS codes. We map the SIC and GICS to the NAICS industries. See the data appendix for more details. 8 BEA produces year 2007 use and make table for 369 industries, defined by three, four, and five digit NAICS codes. They do not include a separate category for NAICS code 213112. For each industry, we constructed the net trade with industries 324 (petroleum refineries), 211 (oil and gas extraction), 213111 (oil and gas drilling) as a share of that industry’s total output. Specifically, the net trade share for industry i is equal to i i M AKE324,211,231111 −U SE324,211,231111 . OU T P U T i 9 shock within the loan book. Loans are debt contracts and given the already low rates of default expected on the typical loan, the upside from having lent to industries whose costs may have declined with the oil price do not offset the downside from exposure to O&G. Another concern might be the explicit hedging of borrowers’ credit risk using derivatives (such as CDS). However, Minton, Stulz, and Williamson (2009) suggest that derivatives positions were largely held by banks for their dealer activities, rather than to protect against the credit risk of their loan book. In addition, we consulted the FR-Y9C and calculated that all but one of our banks had purchased CDS protection of notional value of around 1 percent or lower of their total loan book and the number was approximately 3 percent for the remaining bank. Thus, hedging of credit exposures via derivatives appears to be limited, as theories of adverse and selection and moral hazard might lead us to expect in contexts in which banks have superior information on the underlying loans relative to the broader market (see Duffee and Zhou (2001) for discussions of banks’ informational advantage). Based on these observations, our analysis will use the banks’ loan book exposures to O&G as our treatment variable. Variation in this exposure implies that some banks experienced larger shocks to their loan books than others due to the oil price shock. We calculate O&G exposure as the share of each bank’s total committed exposure (including term loans as well as both drawn and undrawn lines of credit) reported on the Y14 that is accounted for by lending to firms from the O&G industry over the period 2012:Q3-2013:Q4. We start our sample in 2012:Q3 because the data set appears still to be in flux before that point in terms of the banks that are present, undermining comparability. In addition, since we will be examining the effects of a natural experiment, data from earlier periods would be of questionable relevance. Additionally, we leave a buffer before the period of the price decline to guard against picking up simple reversion to the mean and other concerns of endogeneity. The average exposure was 5.9 percent and, importantly for our analysis, there was considerable variation in this treatment variable across banks, implying a standard deviation of 4.9 percent. Roughly speaking, a one standard deviation increase in pre-shock O&G exposure represents a 48 basis point increase in the overall problem-loan rate for the ‘average’ bank over the course of the post period.9 9 The fraction of O&G loans in problem status increased by 9.8 percent points between 2014:Q2 and 2016:Q3. A one standard deviation increase in exposure (0.049) would cause a 0.049 ∗ 9.8 = 0.48 percentage point increase in the overall rate. 10 It is clear, at least from the market’s perspective, that O&G exposure was important. In figure 3, we group banks into quartile bins based on their pre-shock O&G exposure. Banks that were more exposed to the O&G industry, defined here as those in the top quartile, underperformed in terms of their stock prices and market capitalization following the oil price decline, where the timing of the separation in performance is notable. This is indicative of the repeated guidance given to investors at the time, allowing the information to be encoded in the stock price. The scale of the stock price response does appear striking yet further confirmation of the importance of the impact of the oil price decline on banks comes from media and survey evidence. The price declines of 2014 drew comment in many quarters and the implications for banks’ performance were broadly discussed in the markets and the press (see Alloway (2014) and Jenkins (2014)). Fairly soon after the decline in prices, concerns were raised about its implications for banks. Furthermore, discussion of the implications of the decline continued (and continue) beyond the period of the shock (see Noonan and McLannahan (2014), SNC (2016) and Vickery, Thomas, and Velasquez (2016)). From the Senior Loan Officers’ Survey in January 2015 (to be discussed further below) there is explicit evidence of credit tightening due to the oil price decline: Some survey respondents specifically noted their concerns about the oil and gas sector resulting from the sharp decline in the price of oil as a reason that they had tightened their lending policies. Furthermore, there is some suggestion that our exposure variable correlates with exposure to the O&G industry through other avenues. While tracking every aspect of exposure across the balance sheet is beyond the scope of this paper, it appears that the more exposed banks by our measure also typically made more mortgage and non-O&G commercial lending in oilproducing regions of the country. We assess this by using 2013 BEA data on the fraction of non-farm employee compensation attributable to the O&G industry by county.10 Using the granular Y14 data which includes the location of mortgage borrowers and the mailing address of commercial borrowers we see there is a positive correlation between our exposure variable and the banks’ residential lending and non-O&G commercial lending by region. While we 10 BEA Table CA6N: compensation of employees by industry. 11 found that default rates of residential and non-O&G borrowers did not rise in these regions, one might imagine impacts on banks’ fee income and, more broadly, the demand for loans. These correlations emphasize that our exposure variable plausibly indexes our banks based on their broader O&G exposure.11 It is important to point out that the committed exposures reported by the banks include both on- and off-balance-sheet amounts. In particular, undrawn or partially drawn credit lines are reflected in the Y14 as loans with utilized exposures lower than the committed exposure. The off-balance-sheet commitments can sum to particularly large amounts. Indeed, committed O&G loans make up a non-trivial fraction of both total on-balance-sheet loans and total assets, averaging 4.6 and 2.3 percent over the pre-period, respectively.12 In addition, if one alternatively expresses exposure as a fraction of equity capital, the mean and standard deviation across banks are 19.1 and 21.0 percent, respectively, emphasizing that multiple banks had significant net worth riding on the performance of the O&G sector and, by implication, oil prices. The prominence of regulatory capital requirements and the sensitivity of markets to off-balance-sheet exposures shown by the prominent role they played in the financial crisis give reason to believe that our exposure variable is an effective index of bank vulnerability to the oil price decline (Cornett, McNutt, Strahan, and Tehranian (2011), Gilchrist and Zakrajšek (2012), Acharya, Schnabl, and Suarez (2013) and Berrospide and Meisenzahl (2015)). In section 6.2.1 we make the connection to net worth more explicit where we use the bank’s pre-shock O&G exposure as a first stage instrument for the bank’s change in market capitalization. In the narrowest sense, the shock implies an adverse change in the payoffs from loans to O&G firms resulting in a mechanical effect on net worth. More broadly, however, the shock could affect the market’s perception of bank profitability going forward, reflecting reputational loss and endogenous responses over concern from direct losses. This broader net worth channel could have an effect on the ability of the banks to fund their 11 There is some limited evidence that banks with higher O&G exposure held riskier O&G loans, implying our measure could be underestimating the true risk exposure to the O&G industry. The correlation between pre-shock O&G exposure with the post-period problem-loan rate for O&G loans is 0.15, but statistically insignificant (p-value = 0.25). 12 While our exposure variable is committed O&G lending as a share of total committed commercial loans, we also show in a robustness section that our results are robust to using committed O&G/total on-balancesheet loans and committed O&G/total assets. 12 existing amount and structure of assets. Finally, we note that our shock came at a time when the financial markets were generally not suffering from intense stress. While the oil-shock was non-trivial, it was also not of the order that led to any real existential concern for the banks that were particularly exposed. Thus, we regard our analysis as providing an insight into how banks behave in response to a ‘normal shock’ in ‘normal times’. As such it is an important complement to studies that make use of ‘crisis shocks’ that roiled broader financial markets and may even have raised concerns about institutions’ survival (for example, U.S. banks’ exposure to Lehman Brothers during the financial crisis in the case of Chodorow-Reich (2014), or European banks’ exposure to sovereign debt during the Eurozone sovereign debt crises in the case of Acharya, Eisert, Eufinger, and Hirsch (2014)). 3.3. Econometric specifications The main specification used in our empirical analysis follows Khwaja and Mian (2008). Specifically, we employ a fixed effects regression of the following form: ∆Yij = βj + β1 Zi + β2 Xi + εij , (1) where Yij is a loan between bank i and borrower j from the pre- to the post-shock period. βj is a borrower fixed effect, Zi is the bank’s exposure to oil (the share of its loans accounted for by O&G firms in the pre-shock period), and Xi are bank controls. We abstract from interactions of Zi with borrower characteristics although they will be used below. We here describe our approach for the corporate loan dataset though it is essentially the same for the mortgage data. As in Khwaja and Mian (2008) we define a particular loan concept for our empirical analysis, constructed from the raw underlying loan data. Our concept of a loan is a bank-borrower pair so that, in examining the intensive margin, ∆Yij is the change in the mean individual loan committed balance within the pair, where the mean is a simple average across quarters in the relevant subperiods. We focus our attention on term loans as opposed to lines of credit to avoid complexities related to the decision to draw upon the line or whether or not and when the line was expected to be drawn upon. With term loans we have a clear picture of what funds were being loaned and when these funds were made available. In fact, as discussed below, the behavior and presence of existing lines of credit will play an important role in our analysis, so that we do not abandon their 13 information content completely. We discard loans that were 90 or more days past due and restrict our sample only to U.S. borrowers. For all econometric specifications, we remove all loans to firms in the oil and gas industry. Further details of how we cleaned the data are in the data appendix. In examining the intensive margin, we require that there be a continuous bank-firm relationship starting before or during the pre-shock period and extending into the postshock period. In examining the extensive margin we track the existence of the bank-firm pair in the pre- and post-shock periods, allowing us to apply a linear probability model where Vij is an indicator capturing entry (a relationship being initiated in the post-shock period that had not been present in the pre-shock period) or exit (a relationship that had been present in the pre-shock period but was not present or came to an end in the post-shock period). The importance of the fixed effect is that it absorbs the component of the ∆Yij that is common across i. In particular, it is intended to capture any general shift in the firm’s credit demand. This allows us to avoid concerns that the effects of the oil price shock might induce correlated credit demand and supply shocks, which would lead to endogeneity bias in the regression and confound our measurement of the credit supply effect, captured in β1 . This is the main coefficient of interest as it captures the identified effect on the lending relationship of the credit supply shock arising from the bank’s relative exposure.13 The interpretation of β1 is somewhat subtle. Consider, for example, the intensive margin regressions. The coefficient will indicate the association between a bank’s oil exposure and the change in the amount borrowed from that bank by the firm in question - the response of lending to the identified credit supply shock. However, one should not interpret the 13 If all banks systematically made adjustments in credit availability to the same firm, with the degree uncorrelated with their own exposure to oil, this would also be absorbed in the fixed effect, along with any credit demand effects. These could reflect the particular relevance of the oil shock for the credit-worthiness of a borrower and may additionally include effects interpretable as ‘systemic’ that ripple across all industries’ due to the broader effect of the shock on the economy. Nevertheless, in what follows, we will adopt the standard terminology in this literature (which focuses on separating demand from supply effects) and refer primarily to the fixed effect as absorbing firm-specific shifts in credit demand. Our identification allows us to examine the effects of the oil shock, controlling for credit demand effects, systemic supply effects or effects on lending to particular firms that are common across banks. Thus β1 relates to the marginal effect arising from banks’ relative exposure. To extrapolate from relative to systematic shocks, one must make very particular assumptions (see Chodorow-Reich (2014) for a clear assertion of such assumptions). 14 coefficient as indicating the amount by which a bank reduces its desired lending, ceteris paribus. Terms of credit are multidimensional so that even if a bank is in some sense still willing to lend the same dollar quantity, it may nevertheless tighten credit along some other margin, such as price or covenants. This could (and we argue below, did) induce a response from firms, such as seeking alternative financing, that would be encoded in any reduction in lending captured by β1 . While regression (1) allows us to analyze the bank lending channel, we complement it with a second type of regression that addresses the credit supply channel. Thus, we run a specification similar to (1), but aggregated to the level of the firm: ∆Yj = β0F + β1F Z̄j + β2F Wj + β3F X̄j + ηj , (2) where ∆Yj is the change (from pre- to post-oil shock) in an attribute Y of firm j. Z̄jF is the weighted average O&G exposure of the banks from whom firm j borrowed in the pre-shock period. The variable is meant to capture how exposed the firm was to the credit supply effects induced by the oil price shock, via the banks with whom it was associated. Wj are firm controls and X̄j are weighted averages of the characteristics of the banks with whom the firm was associated. The primary coefficient of interest in this regression is β1F as it indicates whether or not firms were systematically affected by being exposed indirectly to the oil shock via their banks. If we find that oil exposure typically induces a tightening of credit supply then one might expect a non-zero coefficient if firms are unable to substitute alternative financing. However, if the firm can substitute for any reduced financing or away from lending on worsened terms, then one might not expect any effect. In the case of mortgage lending, we can no longer apply the Khwaja and Mian (2008) approach at the borrower-bank level because we cannot identify a given household across different banks. However, this fine a level of granularity is likely unnecessary to avoid the aforementioned concerns with confounding credit demand and supply shocks for household borrowing. It is implausible that banks differ systematically with regard to lending to individuals in a way that is correlated with their oil shock exposure and in a way that is not equally well-captured in their lending patterns at the county level. Thus, since our mortgage data include borrower zip codes we then aggregate to the county-bank level before applying 15 the same fixed effects approach to absorb county-level mortgage demand effects. Summary statistics of the log change in mortgage loans within a given county-bank pair between the pre- and post-shock period are shown in table 4. The different categories of loans relate to different classes of mortgages meant to roughly mimic the broad categories included in mortgage lending questions in the Federal Reserve’s Senior Loan Officer Opinion Survey (SLOOS), discussed below. 4. Results In this section we discuss the results of our analysis. We begin by illustrating some relatively coarse but suggestive evidence that the variation in banks’ exposures induced varying responses in terms of their lending and portfolio decisions - decisions that show a coherent cross-balance sheet pattern of de-leveraging by de-risking. We then progress to our more rigorous regression analysis to confirm its essential message. 4.1. Aggregate analysis - balance sheet and lending tightness responses 4.1.1. FR-Y9C Data Turning to banks’ response to the shock, in figure 4 we show how banks’ balance sheets respond to the oil price shock over time using publicly available bank-level information from the FR-Y9C. This aggregate bank-level analysis is meant to highlight the direction and timing of the impact of the price shock and does not control for time-varying bank-specific covariates or demand factors. For each bank, we regress the logarithm of the bank-level variable on time dummies, time dummies interacted with O&G exposure, bank fixed effects, and bank-specific time trends.14 We then plot the estimates and 90th -percentile significance bands of the coefficients on the time dummies interacted with O&G exposure. Panels A and B show that the more exposed banks had lower commercial and residential lending following the shock. Panel C shows the analogous plot for agency mortgage backed securities (MBS) and is essentially a mirror image of panel B. In panel D, we depict a simple 14 Since we include both time fixed effects and time trends we need to omit two time periods. We omit 2012:Q3 and 2014:Q2, which implies the bank-specific trends are estimated over the pre-shock period, 2012:Q3 - 2014:Q2. For ease of comparison, we normalized O&G exposure to have a mean of one-standard deviation. To correct for a coding change in risk weights implemented by Basel III in 2015:Q1, the riskweighted asset ratio (panel D) for 2015:Q1 for each bank was spliced with its 2014:Q4 value. 16 measure of the riskiness of the bank’s balance sheet based on dividing total risk-weighted assets by total assets and plotting the evolution of this variable around the time of the shock. Panels E shows that there are only negligible changes to total assets. Overall, these results show a clear de-risking approach among the more exposed banks: the more exposed banks load up on relatively low-risk MBS and decrease their portfolio lending, while leaving total assets relatively unchanged. While the results of this exercise show striking portfolio responses to the oil shock, this type of aggregate analysis is prone to identification problems. In particular, it does not disentangle credit supply from credit demand. Firms whose credit demand might be affected by the oil shock could systematically be distributed among banks according to bank oil exposure. For example, firms for whom the oil shock reduces their investment opportunities and, thus, their demand for loans, might be concentrated among banks who themselves are damaged by the oil shock. Similarly for residential loans, it is possible that households demanding mortgages might disproportionately be affected in a systematic way, perhaps due to regional concentrations in the banks’ mortgage lending and the sensitivity of those regions’ economies to the oil shock. Importantly, we account for such potential endogeneity bias in the granular analysis that follows.15 4.1.2. Senior Loan Officer Opinion Survey To assess whether the oil shock induced any tightening of credit among affected banks, we use the Federal Reserve’s Senior Loan Officer Opinion Survey (SLOOS) to provide insight on how banks’ lending standards responded during the period of the oil price shock. The SLOOS is a quarterly survey where banks report how various terms and conditions have changed since the previous survey. The SLOOS is qualitative: respondents indicate whether lending standards have tightened or eased (‘somewhat’ or ‘considerably’), or remained about 15 The MBS in panel C of 4 are held on the bank balance sheets as securities, and could have been acquired in two different ways. First, the banks may originate mortgage loans and then securitize them, effectively swapping exposure to whole mortgage loans for exposure to MBS. Banks often retain the servicing rights to these mortgages and, consequently, report these loans in our Y14 mortgage data set. Thus, we can apply the loan-level granular analysis to securitized mortgages. We refer to these loans as non-portfolio mortgages below. A second way the banks can accumulate MBS is to buy the securities outright. Changes in bank liquidity requirements over our sample period may have given some banks incentive to increase their holdings of MBS, independent of the oil price shock. We deal with this possibility and its implication for bias in the robustness section. 17 the same. Importantly, the survey questions ask about lending standards for new loan applications, and thus can be interpreted as the willingness of the banks to supply credit in their current environment. To be sure the SLOOS responses conform to our constructed O&G exposure variable, we make use of the (not publicly available) underlying bank-specific data.16 In figure 5 we plot an index of tightening on loan rates and covenant requirements for commercial and industrial (C&I) loans.17 We again average individual responses of BHCs in the top and bottom quartiles of exposure. The index shows average cumulative tightening, so a decreasing index depicts BHCs that have eased terms since the last survey, an increasing index depicts tightening since the last survey, and a flat index depicts no change in lending standards. The indexes are both normalized to zero in 2014:Q2. BHCs in both groups were generally easing over this period of time. However, on average, BHCs with relatively high exposure to the oil sector (red lines in figure 5) stopped easing in late 2014 and began to tighten the terms of required loan rates and loan covenants.18 This finding, though largely qualitative, is an important complement to our regression results. The Y14 has very granular information on loan quantities and loan performance, but does not contain information on covenants. Standard sample selection issues (we only observe transacted prices) make it difficult to assess the effect of the credit supply shock on the price (interest rate) margin because of the tendency of firms to substitute away. The results from the SLOOS suggest that exposed banks shifted their credit supply function relative to the less exposed BHCs, which will provide useful insights into the sources of our results below. 4.2. Granular analysis—commercial loans The patterns discussed in the previous section are consistent with a substantial fraction of the aforementioned reduction in corporate loans among oil-exposed banks capturing a general desire to repair the balance sheet—a credit supply effect. However, even with the 16 The SLOOS may not be used for supervisory purposes, and the confidential bank-level survey responses may only be used by Federal Reserve System staff in non-supervisory functions for limited-scope economic research projects. 17 The BHCs in the indexes are ones where we could merge the bank ID in the Y14 data to the SLOOS data. Almost every Y14 bank was also located in the SLOOS. 18 For recent work on the role of covenants in the mechanisms by which banks tighten lending in the syndicated loans market, see Chodorow-Reich and Falato (2017). 18 support from the SLOOS, this is only suggestive evidence. Thus, we now move to a more rigorous regression analysis. 4.2.1. Bank lending channel—intensive margin In table 5 we illustrate results from regression analysis on the change in the intensive margin. In our most basic fixed effects regression, shown in the first column of table 5, we find that β1 (in the first row) is estimated to be negative and significant at the 1 percent level, where our standard errors are two-way clustered at the bank and firm level. The coefficient estimate of −0.83 implies that a one standard deviation increase in bank exposure to oil (4.9 percent from table 2) is associated with a 4 percent decline on the intensive margin of a randomly selected loan. Thus, the bank lending channel appears to be active.19 By comparing the OLS and FE estimators on the same sample we can assess the size of any bias induced by firm credit demand. In the second and third columns of table 5 we show the estimated OLS results using the same sample as our FE analysis (firms that borrow from multiple banks before and after the shock) and also on the much larger sample of loans (corresponding to 42, 111 firms instead of 3, 541) of all firms who borrow from possibly only one bank, before and after. Details of the variables involved in the regressions are included in table 3. Within the same FE sample we see that the O&G exposure coefficient is broadly similar (−0.83 in the FE specification compared to the −0.58 using OLS and firm controls), implying that the correlation of credit demand and supply is apparently not severely biasing the OLS coefficient. If anything, the correlation appears negative: firms that experience positive demand shocks appear more concentrated among banks that experienced negative supply shocks. Thus, there is the suggestion that endogeneity bias, if there is any, may be towards attenuating the measured effect. In this case, and assuming any correlation between credit supply and demand effects is similar in the larger sample, we may expand our sample and treat the estimated OLS coefficients with reasonable confidence as a lower bound (in magnitude) on the credit supply effect. In this larger sample, we again obtain a highly statistically significant coefficient on O&G exposure, although smaller in magnitude than the FE sample 19 Similar results are obtained when we weight observations by the size of loan, indicating that our results are not being driven by the size of the loan. 19 (−0.32). This suggests that the lending channel effect is weaker for single-relationship firms or, alternatively, that any effects of tightened credit are manifested in dimensions other than loan size. 4.2.2. Bank lending channel—extensive margin In the first column of table 6 we show that, under our simplest fixed effect regression, there is a highly significant effect of the oil exposure variable on the probability of exit. In column 1, we see that a 1 percent increase in oil exposure is associated with a 0.84 percent increase in the probability of exit. In the second column, we have the results for an OLS regression (again with firm controls and for the same sample but excluding O&G firms). The coefficient (0.74) is slightly smaller than its FE equivalent, suggesting that there is little bias within the sample. When, in the third column, we expand the sample to include single bank firms the effect is lower in magnitude suggesting a stronger exit effect on multi-bank firms. The remaining three columns in table 6 include the same basic specifications as in the exit regressions, but now the dependent variable is the event of forming a new banking relationship in the post-oil shock period where none previously existed in the pre-shock period. New bank-firm relationship formation does not appear to be significantly related to bank oil exposure. Thus, our extensive margin results for firms exiting their bank relationships are broadly in line with our intensive margin findings. The bank lending channel exists and there are suggestions that its effects, within bank-firm pairs, are particularly noticeable among firms with alternative bank financing. This is captured by the fact that the effects seem smaller for our larger OLS sample, which contains firms with fewer banking relationships. 4.2.3. Bank lending channel—heterogeneous effects across firms Using our analysis to dig somewhat deeper, we assess whether the bank lending channel differs across types of firms. We show in table 7 results for the intensive and exit margins where we interact exposure with a dummy indicating a ‘small firm’, where small is defined as in Khwaja and Mian (2008) being below the 70th percentile based on average total loan size.20 Results show no differential effect on the intensive margin but a negative and significant effect on the exit regression. This means that banks more exposed to the oil shock are more likely 20 Results did not change when we included a continuous measure of firm size. 20 to stop lending to larger firms. To dig deeper into this somewhat counterintuitive result we delve into the characteristics of the firm that capture the firms’ access to alternative financing. Our results from the previous section on expanding to the single bank OLS sample already hint at a relationship between size and availability of opportunities for substitution (see table 3). In addition to this, we create a variable (external finance) that flags whether or not the firm has a CUSIP or a stock ticker. This is intended to capture the firm’s access to the capital markets, which is one of the attributes that ‘size’ is implicitly taken to proxy. To capture the ability of the firm to draw on an already existing line of credit, we construct a variable (pre-period utilization rate) that measures the share across all of the firm’s lines of credit (from any Y14 bank) that has been utilized in the pre-shock period. The lower the utilization rate in the pre-period, the more the firm can borrow on their lines in the post period. Finally, we include the number of bank relationships in our sample in the pre-shock period, given the fact that much research (see Petersen and Rajan (1994), for example) indicates non-trivial costs of switching from an established banking relationship. Most of these interactions are statistically insignificant. However, there is a statistically significant positive coefficient on the number of pre-existing banking relationships in the exit regression on the both the FE and OLS regressions. This positive estimate suggests that firms with existing alternative banking relationships in the pre-period were more apt to leave their relationship with banks who were exposed to the credit supply shock. As large firms tend to have more banking relationships, this explains the stronger exit effect on larger firms.21 As we will show later in the paper, we find strong evidence that those firms who exited in the post-shock period simultaneously entered into new banking relationships with banks that were less exposed to the oil shock. 4.2.4. Credit channel—firm-level outcomes In the previous section we demonstrated strong evidence of a bank lending channel. Banks that were particularly affected by declining oil prices restricted their credit supply. It is unclear if this ultimately should affect firm outcomes. The U.S. financial system is often 21 The coefficient on the small-firm interaction and number of banks interaction become insignificant when included in the same specification. The correlation between number of banks in the pre-period on the small size dummy is negative and statistically significant. 21 contrasted with the more bank-dependent system in Europe. It is argued that the scale and sophistication of public debt and equity markets in the U.S. reduces the dependence of firms on banks and allows them to circumvent the loss of credit from the banking sector. Of course, it is also possible that firms could substitute within the banking sector, as our interactions analysis above suggests they might. Furthermore, even if a firm’s opportunities to borrow using term loans are reduced, it may be able to substitute to other forms of financing. In table 8 we show the results of implementing regression (2), for various dependent variables. The first two columns of table 8 report estimates of the change in firms’ total term loan borrowing and total borrowing from Y14 banks, respectively. Interestingly, these regressions show a statistically insignificant relationship with the aggregate exposure variable, implying that firms that were borrowing from more exposed banks in the pre-shock period had no net change in their term loans due to the credit supply shock. This result implies that firms were able to substitute simply from other Y14 banks.22 In columns 3-5 we move beyond the Y14 borrowing to look at additional balance sheet objects from Compustat.23 Here we see no statistically significant estimates, implying essentially no relationship between changes in total liabilities (or assets, or equity) and the aggregate exposure variable. Overall, in spite of the significant adjustments that exposed BHCs were making in the wake of the oil price shock, borrowing firms (at least in the Compustat sample) were apparently able to smooth out the effects of the shift in credit supply using other means. Our firm-level estimates are potentially biased since they are estimated under OLS and recall from our estimates of the intensive margin in table 5 that there appears to be some upward endogeneity bias in the OLS estimator. Jimenez, Mian, Peydro, and Saurina Salas (2014) suggest an adjustment (described in the appendix) to the firm level estimated co22 Since the dependent variable is the change in the log of loans, we necessarily drop those firms that entered or exited the Y14 sample in the pre or post period in these regressions (to deal with zeros). We tested whether this could be biasing our result by running a regression using a Y14 exit dummy as the dependent variable in the aggregate regression. The coefficient on aggregate O&G exposure was insignificant, implying firms that exit the Y14 were no more or less likely to be banking with more exposed banks. 23 We use the Compustat variables in this case because they are obtained from the borrowing firms’ official financial statements and, presumably, will be more accurate than the bank-reported equivalents in the Y14. Note that the sample size declines substantially in columns 3-5 of table 8 due to the requirement that Compustat firms be publicly traded. Note also that for clarity we do not include the coefficients on the controls in the table. 22 efficient based on comparing the bank-firm level coefficients from FE and OLS regressions. Making this adjustment does not materially change our insights — after the adjustment, it appears that firms were able to substitute at least 80 percent of the decline in loan supply simply from other Y14 banks. Indeed, this number is likely a lower bound on the true degree of substitution, as discussed in the appendix. Overall, the results from the firm regressions suggest a substantial ability of firms to substitute alternative financing if the loan supply shock induced a tightening of credit supply by the banks from whom they were previously borrowing. Importantly, if firms can cushion themselves by borrowing from alternative sources in this way, then it breaks the vicious circle that underpins various financial accelerator models (see Shleifer and Vishny (1992) and Kiyotaki and Moore (1997), for example). If the shock’s effect on banks’ credit supply has little to no first-round effect on firms (contractions, defaults, and fire-sales) that could in turn feed back on the value of collateral and bank assets and so further damage firms and banks not initially directly exposed to the shock, then the amplification mechanisms of these models simply do not come into play. 4.2.5. Firm substitution of funding sources Our interaction results in table 7 and the firm level results of the previous section suggest substantial term-loan substitution within the set of Y14 banks. To further verify that this type of substitution is driving the aggregate term-loan result, we run the fixed-effects entry regressions of the type we ran in section 4.2.2 above, but on different subsamples of firms. In table 10, we break the population of firms into groups according to whether they experienced any relationship exit (column 1), did not experience any exit (column 2), experienced any decline in term-loan borrowing from a BHC (column 3), or did not experience any decline in borrowing from a BHC (column 4). Columns 1 and 3 demonstrate that firms that experienced any reduction in lending (by any bank, for any reason) were more likely to strike up new relationships with less-exposed banks. Conversely, firms that did not exit old relationships or experience any measurable decline in term loan funding show no such new relationship behavior. As another test, we isolate those firms that switched banks—that is, those firms that exited at least one banking relationship and entered at least one banking relationship. In column 5, we run our firm-level regressions with the dependent variable being a dummy 23 variable equal to one if the firm switched banks. The coefficient on aggregate O&G exposure is positive and statistically significant, indicating that those firms more exposed to the credit supply shock were more likely to switch banks. These results show evidence of firms substituting away from exposed banks to less exposed banks for financing. Of course, it is not necessary that firms know which banks were affected by oil. Instead, the result is likely driven by the fact that they were presumably aware of the easier terms offered by the less affected banks, as evidenced in our SLOOS results. Finally, we consider another method of financing—the firms’ undrawn credit lines. An additional reason why we observe no effect on total utilized loans (see column 2 of table 8) could be that firms drew down from an already existing line of credit. In table 9 we test whether firms using banks more exposed to the oil shock drew on their lines of credit. Specifically, we regress the log change in the utilization rate (defined as total utilized amount divided by total committed amount) between the pre and post period, on each firm’s weighted O&G exposure with controls.24 There is apparently no effect using the sample of all firms. However, the interaction with the pre-period utilization rate implies that there was an effect for firms with relatively large shares of their borrowing capacity untapped. Our credit line results emphasize the benefits of the granularity of our data as one observes that even though we have identified an overall negative shock to the suppliers of credit, a result of this in some dimensions (due to substitution) can be an increase in certain types of credit. Thus, using only limited data on credit lines or unused commitments, an analyst could easily be led astray as the data might suggest a positive credit demand shock. This alludes to a broader point we make in this paper that the credit supply response and amount of lending ultimately transacted is multifaceted. Traditionally, bank lending channel papers give the impression that a bank funding shock induces a uniform reduction in credit. However, we show that the response is more subtle. As we now illustrate, this is particularly obvious when one looks across the balance sheet, revealing the coherence of banks’ business models across various asset classes. 24 We restricted the sample to firms with utilization rates less than 100 percent in the pre-period. 24 4.3. Residential loan analysis Our results using the corporate loan book give us interesting insights into how banks respond to shocks. A significant advantage of having access to other components of banks’ balance sheets is that we can also examine how other types of lending responded. As noted in section 4.1, panel B of figure 4 indicates that residential lending among more exposed banks lagged behind that of the less exposed. In this section, we tease out the underlying drivers of this relationship and explain how it relates to the behavior of MBS holdings in panel C of figure 4. 4.3.1. SLOOS responses—residential loans We begin by again examining responses from the SLOOS, but here in relation to residential real estate lending. Figure 6 shows how the evolution of credit terms depended on banks’ exposure after the oil price shock. Note that since the SLOOS questions on mortgages changed in 2015:Q1 we are unable to plot pre- and post-oil shock charts as we did for commercial loans or adjust for pre-shock trends. In panels A and B of figure 6 we show the cumulative tightening of terms for GSE-eligible and Government mortgages. We observe that there was a greater tendency on average to loosen terms among the more exposed banks. In contrast, in panels C and D we repeat the analysis for jumbo and non-jumbo, non-GSE lending and observe that the pattern is flipped—on average, the more exposed banks seem to have tightened relative to the less exposed.25 Immediately, therefore, we see that there are subtleties in these responses. The latter pair of panels, are apparently more intuitive given the Y9 data on residential loans in the banks’ portfolios, illustrated in figure 4. The (relative) relaxation of terms by the less exposed banks, is consistent with an expansion in their lending, as borrowers presumably 25 The GSE-eligible category of residential mortgages includes loans that meet the underwriting guidelines, including loan limit amounts, of the GSEs—Fannie Mae and Freddie Mac. The government category of residential mortgages includes loans that are insured by the Federal Housing Administration, guaranteed by the Department of Veterans Affairs, or originated under government programs, including the U.S. Department of Agriculture home loan programs. The Qualified Mortgage (QM) non-jumbo, non-GSE-eligible category of residential mortgages includes loans that satisfy the standards for a qualified mortgage and have loan balances that are below the loan limit amounts set by the GSEs but otherwise do not meet the GSE underwriting guidelines. The QM jumbo category of residential mortgages includes loans that satisfy the standards for a qualified mortgage but have loan balances that are above the loan limit amount set by the GSEs. 25 are attracted to their loan products rather than those of the more exposed banks. But what of the GSE eligible and government mortgages? As the more exposed banks loosened their lending standards with regards to GSE-eligible and government mortgages, one might have expected an expansion of mortgage lending, which would superficially be in tension with the Y9 data (i.e., panel B of figure 4). However, one must recall that GSE-eligible and government mortgages are largely intended to be securitized and shifted off the banks’ balance sheets. Thus, even if the better terms induced greater transacted originations, we would not observe them in the Y9 data on mortgages held in the banks’ portfolios. Returning to figure 4, panel C suggests that a higher rate of securitization of mortgages could explain these patterns. That is, after the shock, the more exposed banks pulled back lending (by tightening their terms) if they would ultimately have to retain those mortgages on their balance sheet, while expanding originations of the types of mortgages that they could securitize (and hold less capital against for regulatory purposes) or perhaps securitizing a higher fraction of such mortgages. This hypothesis cannot be tested further using the SLOOS, so we now turn to our Y14 data to clarify matters. 4.3.2. Residential loan regression results In table 11 we show the results of our regression analyses, where the dependent variables are those summarized in table 4. Importantly, we can partition residential loans into those originated and retained on the banks’ portfolios (referred to as ‘portfolio loans’) and those originated but either sold or securitized with only servicing rights retained (referred to as ‘non-portfolio loans’). A granular analysis at the loan level will allow us to assess whether the time-series patterns shown in panels A and B of figure 4 are statistically significant after properly controlling for bank characteristics and county-specific credit-demand factors.26 We use only those loans originated by the BHCs themselves, excluding those obtained from correspondent relationships.27 The first column of table 4 relates to total non-government mortgage originations that 26 We describe how we partitioned mortgage loans from the Y14 into different types in the data appendix. The bank level controls are the same as those in our commercial loan analysis, but we also augment them with variables capturing trends in the banks’ mortgage lending and MBS holdings (see table 3 for a description of the controls). 27 Results are essentially the same if the correspondent and other loans not originated by the banks themselves are included. They are available on request. 26 were retained in the banks’ portfolios. For these mortgages, the banks retain the credit risk associated with a default. The portfolio results generally suggest that banks with exposure to the oil shock in the corporate portfolio also reduced their balance sheet exposure to mortgages. For both the fixed effects regression in panel A and OLS regression with county controls in panel B of table 11, the estimated coefficient of -2.2 implies that a one standard deviation increase in oil exposure leads to a roughly -11.0 percent decline in portfolio lending. We estimate the balance sheet pullback to be stronger in the non-jumbo category (column 3) than in the jumbo category (column 2).28 Conversely, we see the opposite effect for the non-portfolio loans that were originated and then securitized by the banks. In this case, banks exposed to the oil shock in the corporate portfolio increased their origination and securitization activity (columns 4 and 5). The effect is particularly pronounced when we limit the mortgages to non-jumbo loans, which are more likely to be guaranteed by the GSEs. The coefficient estimate of 3.0 implies that a one standard deviation change in O&G exposure on the corporate balance sheet is associated with a nearly 14 percent increase in securitization. Similarly, in column 6 of the table we see a significant positive response of changes in loan growth for FHA and VA (i.e., “government”) mortgages—loans likely to be securitized by Ginnie Mae. The coefficient of 8.4 implies a one standard deviation increase in O&G exposure is associated with a 41 percent increase in these types of loans. The magnitude of the estimate of the shock-induced change in originate-to-securitize mortgage lending (non-portfolio mortgages) corresponds neatly with the evidence in figure 4 plotting the aggregated balance sheet data. Recall, in that figure (panel C) we observe a substantial increase in aggregate MBS for banks according to their oil exposure. Strikingly, we find no statistically significant change in overall residential loans (column 7). That is, the decline in portfolio loans (again consistent with 4, panel B) was offset by the increase in non-portfolio loans, implying no discernible change in total residential assets. The ultimate effect was a rotation of their balance sheet away from portfolio loans towards non-portfolio loans and, in turn, agency MBS. Table 12 contains the aggregate regressions, where we relate changes in county-level 28 This is reminiscent of results in Calem, Covas, and Wu (2013) where jumbo, as a share of total mortgage origination, typically rose after the liquidity shock of 2007, although our broader finding of a pull-back in portfolio mortgage lending is somewhat in tension. 27 mortgage balances to the counties’ aggregate exposure to the oil shock via the weighted exposures to the O&G sector of the banks active in those counties. We assess three different aggregate categories: portfolio loans, all Y14 loans (i.e., portfolio plus non-portfolio), and all loans from the Black Knight-McDash Analytics database, which samples loans from the universe of mortgage lenders and not just those lenders captured in the Y14. For all three levels of aggregation there is no statistically significant effect on loan quantity. We point out that these results align with the SLOOS, giving important support to the use of these surveys for insight into banks’ policies. Indeed, our results lend weight to the approach of ‘trusting the bankers’ explored in Ciccarelli, Maddaloni, and Peydro (2015), where aggregated SLOOS responses are used to identify credit supply policy. It appears that the surveys accurately reflect bank behavior (at least in the dimensions we examine) which is the assumption underlying their analysis.29 5. Robustness While the comments and answers provided by banks in the aforementioned lending surveys give support to our regression-based analysis and our overall story, in this section we carry out a set of robustness checks. First, we consider the implications of defining our exposure variable as a share of total on-balance-sheet loans, of total assets, and of equity capital (all as reported in the FR Y-9C) rather than as a share of total committed commercial loan exposure. The motivation for this (though Chodorow-Reich (2014) also scales his exposure by loans, for example) is that the size of the loan book and capital position of banks may differ, implying noise and possibly bias in our index. Table 14 shows that these changes do not substantially alter our findings. A concern is that our regression specification contains too few clusters, leading to either ‘overfitting’ of the cluster-robust variance matrix estimate or over-rejection of the critical values (Cameron and Miller (2015)). There is no definitive measure of ‘too few’, but the general consensus is around 30 clusters—close to the number of banks in our sample. To address this issue, we re-assess our standard errors using the wild cluster bootstrap recom29 Ciccarelli, Maddaloni, and Peydro (2015) only use the aggregate (across the banks responding to the SLOOS) results and also did not make use of mortgage-related responses, as the questions had not yet been incorporated into the survey. 28 mended by Cameron and Miller (2015). Since this method is computationally intensive, we tested only the two main fixed effects specifications in the paper (the fixed effects commercial intensive and extensive regressions) to get a sense of whether our results would differ. For both the intensive and extensive regressions our estimates under the wild cluster bootstrap (clustered at the bank level) were both statistically significant at the 1 percent level.30 There is also a potential concern that the predictive power of our exposure variable was somehow a coincidence. To test this concern, we run a type of permutation test using each bank’s loan exposure to other industries as placebos. Analogous to the O&G exposure variable used throughout this paper, for each of the most prevalent other 49 industries (defined by 3-digit NAICS) we create a bank-specific exposure variable analogous to our O&G exposure variable used throughout this study. We then run our main fixed-effects specification 50 times—once replicating our original specification, and 49 subsequent regressions, each time replacing O&G exposure with one of the 49 other industry exposures. To be conservative, we collected the associated t-statistic based upon the standard errors obtained from the wild cluster bootstrap. In figure 7 we show the distribution of t-statistics obtained from our intensive and extensive corporate lending regressions. Clearly, there is some sampling variability associated with this test such that, even if our O&G exposure variable were the ‘true’ exposure variable, it would not necessarily emerge as having the largest t-statistic, but we would at least hope for it to be in the tail. In fact, the results are extremely strong. In the intensive regressions, our exposure variable is a clear outlier in the left tail (where the left tail is most relevant given the limited upside to a debt contract) and in the right tail in the extensive regressions (similar relevance of the right tail). Finally, we explore the possibility that our sample period includes other, possibly confounding shocks that may have had an impact on the portfolio adjustments that we see our data. In particular, roughly contemporaneous to the oil price shock, U.S. banking regulators were in the process of phasing in new balance sheet liquidity requirements for many of the banks in our sample. The new regulations stipulated that banks maintain a liquidity coverage ratio (LCR) whereby liquid assets could be used to offset possible funding outflows 30 Although clustering should guard against an individual bank driving our results we also ran our commercial intensive and extensive fixed effect regressions repeatedly, each time removing a different bank from the sample. We retained significance in all cases and the coefficient value was consistent (average and standard deviation of −0.81 and 0.06, respectively, for intensive and 0.84 and 0.08 for the exit regressions). 29 in a period of stress. The rule may have influenced the demand for liquid assets for certain banks, which include agency MBS. In the U.S., BHCs with assets over $250 billion needed to meet 80 percent of their liquidity requirements in 2015 (i.e., in our post-shock window). BHCs with assets above $50 billion but below $250 billion were subject to a modified and less severe LCR requirement with a phase-in period in 2016 (i.e., outside our post-shock window). We can use the differential phase-in periods of the rule to control for possible LCR-related asset demand that could potentially bias our regression estimates of the coefficient on O&G exposure. We do this by constructing an indicator variable to identify banks with assets over $250 billion and subject to the LCR phase-in in 2015, and then interacting this variable with a proxy for exposure to the LCR in the pre oil shock period. This proxy consists of a measure of high-quality liquid assets (Treasury Securities + Agency MBS) which is then scaled by a measure of possible funding outflows (Total Liabilities - Deposits). The denominator of the exposure proxy reflects the notion that the LCR treats deposit funding as more stable and less likely to run off in a liquidity event. Thus, BHCs with large values for the proxy have ample liquidity, or are more reliant on deposits (or both), and thus would be less constrained by the LCR. When we include these variables among the bank controls in our regressions we find only small quantitative differences in the estimates of the sensitivity of lending to O&G exposure and no differences in statistical significance. Indeed, the coefficients on the LCR controls are not statistically significant in any of the mortgage regressions. It may still be the case that the banking sector as a whole was gradually increasing its holdings of liquid securities in order to eventually comply with the LCR, but these changes do not account for any of the differences in de-risking or lending behavior between banks with differing exposure to the O&G sector. 6. De-leveraging by de-risking: Discussion In this section we draw together the various strands of our results to emphasize the cross-balance sheet coherence of banks’ responses to the shock. We also point out policy implications of our findings. 30 6.1. Portfolio optimization Together, the granular loan-level analysis corroborates the bank-level time-series patterns observed in the Y9 and the SLOOS data. More exposed banks pulled back lending, but this pullback was concentrated among loans that they ultimately would hold on their balance sheet, and against which they would be required to hold regulatory capital at a relatively high rate (depending on the LTV and other factors, the prevailing risk weights under the Basel standardized approach over the sample period ranged from 35 percent to potentially over 100 percent for mortgages and, according to the particular agency, 0 percent or 20 percent, for agency MBS). In contrast, lending that resulted in loans that could be securitized via sale to GSEs was expanded - with relaxation in terms allowing this expansion, as reflected in the SLOOS. This rebalancing would explain why in the Y9 data the MBS and residential lending diagrams were essentially mirror images as the lending panel in figure 4 only related to loans held on the balance sheet and the MBS apparently increased as a result of the additional origination of securitizable loans. The re-balancing of mortgage lending after the oil shock is an important complement to our corporate lending channel results and goes beyond the standard findings of the bank lending channel literature. Existing papers typically only have access to certain credit registers, rather than the broader insight into the balance sheets of banks that we possess.31 Focusing on the reduction in originations of loans to be held on the balance sheet, the results are consistent with a reduction in credit and, therefore, are in line with the corporate results. But the offsetting increases in originations of loans that are securitizable adds an important subtlety to the story and one that might be missed were one only looking at the residential lending data from the Y9, especially if they were considered in isolation. Banks are constantly solving complex optimization problems in their lending behaviors and there is no reason to expect that a pullback in lending will be uniform across all products. The fact that we can tie these lending book patterns to the banks’ securities holdings— in MBS specifically—also provides an interesting perspective on the connection between securitization and the availability of credit. Much work has been done on the influence of securitization on the availability of credit (see Mian and Sufi (2009) and Keys, Mukherjee, 31 One exception to this is Abbassi, Iyer, Peydr, and Tous (2016) also exploit security and loan-level data in examining a crowding out effect of banks with trading expertise shifting resources from lending to purchasing securities the prices of which had fallen during the financial crisis. 31 Seru, and Vig (2010), for example). However, we provide a novel insight into how the ability to securitize a broad class of mortgages allows banks to maintain their lending following damaging shocks. Our work provides important confirmation of theories of the increased role of securitization in recent decades, such as Loutskina (2011), which previously would have been difficult to assess given the identification problems inherent in using aggregated data. Indeed, quoting Loutskina (2011) If banks can liquidate loans to finance their liquidity need, they can also do so to finance new credit, making banks less dependent on the traditional sources of funds. The increasing liquidity of bank loans ought to change the link from banks’ funding availability (e.g., deposits) to their willingness to supply credit. . . Because securitization provides banks with an additional source of both loan financing and liquidity, it should also shield banks’ willingness to supply credit from the external cost of funds shocks. It is an interesting question why the mortgage book was a vehicle for de-risking the balance sheet. Intuitively, we might expect the mortgage rotation that they undertook to have been quite cost effective. Firstly, the MBS market is highly liquid and, secondly, it seems likely that instructing mortgage loan originators (the employees of the bank) to switch their emphasis to conforming loans, rather than, say, jumbo is a simple switch to make without requiring layoffs or dramatic changes in personnel or retraining. However, this intuition is rather speculative and, of course, we do observe some reduction in corporate lending which would free up regulatory capital (being 100% risk weighted) and resources for expanding MBS holdings. Presumably a ‘bang-bang’ approach of effecting the de-risking via only one margin of adjustment was not seen as optimal by the banks, with this interior solution reflecting the complexity of their optimization problem. 6.2. Policy implications 6.2.1. Bailouts Our results relate to the impact and desirability of ‘bailouts’ or injections of equity into a given bank.32 Since our analysis implicitly relates to a shock to net worth we can imagine 32 Again, given our assumptions, this should be interpreted as a relative (to other banks) injection. 32 a counterfactual in which this shock did not occur, or was offset with a policy that repairs the capital of the bank. Of course, one could argue that symmetry of effects may not be assured, but qualitatively, there appears to be evidence that such a policy could promote the extension of some types of credit by the bank in question and a move towards ‘riskier’ (or at least more heavily risk weighted and/or less liquid) assets.33 Since an important aspect of justifying financial support for a bank is to argue that the broader economy will benefit, the lending channel effects we identify are ceteris paribus apparently in favor of bailouts. However, counter to this are our findings on the credit channel, which suggest that the effect of any credit pullback from a given bank is minimal. Of course, the desirability of bank rescues involves a multitude of complicated factors and may also depend on a particular institution and size of shock as well as differences in the pool of relevant borrowers from those considered in our analysis. In addition, the credit channel effects of systemic banks stress may also favor a policy response. Nevertheless, our results add to the debate over this important policy tool. In order to make this insight more formal, we recast our results explicitly in terms of a shock to net worth. To do so, we run a two-stage instrumental variables (IV) specification of our model, instrumenting the bank’s change in log market capitalization (between the pre and post period) with the bank’s pre-period O&G exposure. In this way, we obtain an elasticity of the lending measures with respect to a change in market capitalization, where the change in market capitalization is attributable to the oil price shock. This allows us to map the oil shock into a common scale, in terms of its ultimate effect on net worth rather than exposure to a particular industry. In table 13 we repeat our main regressions under the IV approach. We lose some power through this procedure but qualitatively our results are confirmed and in most cases we retain statistical significance. The first-stage estimates imply that those banks more exposed to the oil-shock had reductions in their market capitalization, confirming the depiction in figure 3. Specifically, the estimates from our commercial loans regression imply that a one33 There are some suggestions of similar risk reallocation effects in relation to liquidity provision discussed in Drechsler, Drechsel, Marques-Ibanez, and Schnabl (2016), where it is shown that lender of last resort provision of liquidity to Eurozone banks by the ECB during the financial crisis seemed to allow those banks who were weakest - and for whom regulatory or endogenous capital requirements might be most binding to shift towards risky assets. Duchin and Sosyura (2014) also contains some reminiscent results in terms of banks that received TARP aid shifting towards risky securities. 33 standard deviation increase in exposure was associated with a 5 percent decline in market capitalization. In terms of quantitative implications, our results imply that a 1 percent hit to net worth leads to a 0.90 percent decline in corporate lending on the intensive margin and a 0.95 percent increase in the probability of exit. In terms of mortgage lending, the response is a 1.8 percent decline for non-government portfolio loans, 0.6 percent increase for non-government non-portfolio, and a 3.5 percent increase for government loans. Our aggregate regressions again suggest that the ultimate impact on exposed borrowers was minimal, relative to those borrowing from less exposed banks. 6.2.2. Stress testing How banks respond to stressful conditions is also of particular interest following the financial crisis. Our results suggest interesting opportunities to calibrate stress testing models of balance sheet re-optimization. Halaj (2016) provides an innovative attempt to incorporate optimizing behavior into a realistic model of bank balance sheet adjustment though most stress testing models still rely on judgment and rules of thumb in this respect (see also Birge and Judice (2013) for a simpler model featuring optimization). Since a loss of net worth is in some sense a dual of facing tightened capital requirements (in terms of a reduction in the distance of a bank from its ‘optimal’ buffer over minimum levels of capital) our work should also aid the calibration of models attempting to establish the appropriate regulatory requirements. Indeed, using aggregate and cross country data, Cohen (2013) provides interesting evidence that banks have transitioned to higher capital requirements post-crisis using some mix of a change in total assets and adjustments in riskweight profiles. Our analysis of banks’ responses to an identified shock complements these findings and puts them on a firmer empirical footing. 6.2.3. Capital and liquidity requirements In terms of regulatory constraints Cecchetti and Kashyap (2016) note that it is not obvious which of the various capital and liquidity ratios now required of financial institutions actually binds (capital relative to RWA, or simple leverage ratios, along with net stable funding ratios and requirements for sufficient high quality liquid assets). With no discernible effects in terms of overall quantities (commercial and portfolio mortgage lending declines, while securitized mortgage lending expands), but a decline in the risk-weighted asset ratio, it appears that risk weights are an important factor in a banks’ portfolio response to a net 34 worth shock and that it may not currently be the leverage ratio that binds, as opposed to risk weighted measures of leverage. Since there is also a gain in liquidity from shifting towards MBS, it is difficult to separate the impact of capital and liquidity requirements. Exploring this further is beyond the scope of this paper, though interpreting de-risking somewhat more broadly to incorporate concerns over liquidity risk (which is known to interact importantly with solvency risk) provides an interesting dimension to our results. It remains unclear whether regulatory constraints are what induces banks to ‘de-leverage by de-risking’. In Holmstrom and Tirole (1997) it is shown that endogenous minimum capital ratios for banks emerge as a result of financial frictions stemming from asymmetric information, even in the absence of regulation. This model features risk neutral agents and the concepts of different asset classes distinguished by risk does not feature. A richer model with a non-trivial portfolio problem might reveal whether the market might require the sort of de-risking patterns we observe. Without such a model, separating these endogenous constraints from regulatory requirements would be a very difficult task, however. 6.2.4. The role of GSEs Another interesting avenue for further exploration is the implication of our results for understanding the role of GSEs. Our results suggest that following a shock, the more affected banks reduced their risk profile relative to the less affected, and they did so partly by switching their mortgage origination towards loans that could be — and then were — securitized into less heavily risk weighted asset classes. The presence of a deep and liquid secondary market for mortgages allowed this portfolio reallocation to be effected at relatively low cost. The involvement of the GSEs is influential in providing this cheap technology for balance sheet reorganizations.34 One might imagine that absent this technology the banks in our paper would have amended their risk profile by a different (more costly for them) breakdown between risk weight adjustment and total asset adjustments. Since this ability to de-leverage by derisking partly relied on the GSEs, we note the interesting qualitative implication that the GSEs influence the ex post consequences of bank-specific lending and investment decisions 34 One could argue that a private securitization market could allow the same opportunities although to the extent that GSEs benefit from a special funding position (see Acharya, Richardson, Nieuwerburgh, and White (2011) for a discussion) there is still a particular relevance to examining their role. 35 in areas other than their real estate lending. This suggests, qualitatively at least, that there could be a moral hazard effect pervading all bank business lines and not just mortgages: if GSEs help insure against the consequences of banks’ general investing decisions they give them less of an incentive to be careful. Whether or not this is an important phenomenon quantitatively is an open issue. 7. Conclusions Banks damaged by exposure to industries adversely affected by the oil price decline made significant adjustments to their balance sheets, as theories of the importance of bank net worth suggest they should. We observe a significant pull-back in commercial lending by exposed banks, both on the intensive and the extensive margins. Importantly, the balance sheet adjustments by banks exposed to the oil price shock were not confined to the corporate loan portfolio where the shock might be expected to have its most direct impact. Similar derisking results are obtained with regard to mortgage lending. Exposed banks tightened credit supply for those mortgages that they would have to hold in their own portfolios. However, there was an expansion in credit for loans that could be securitized, leading to an increase in their MBS holdings. Thus, there are interesting connections between the banks’ lending book and securities holdings, constituting a tendency of the damaged banks to de-leverage by derisking. This is a novel insight relative to the traditional ‘bank lending channel’ view which emphasizes a general de-leveraging and pull back in lending. Ultimately, the relative impact on borrowers who were particularly exposed to damaged banks was apparently limited. It seems that the ability to substitute to alternative financing sources (largely) allowed borrowers to smooth through any credit tightening imposed by the exposed banks. Thus, while the bank lending channel appears to be active, the broader credit channel, at least in this case, was weak, reflecting the richness of alternative financing sources. Significant policy implications arise from our results and provide important avenues for future work. References Abbassi, P., R. Iyer, J.-L. Peydr, and F. R. Tous (2016): “Securities trading by banks and credit supply: Micro-evidence from the crisis,” Journal of Financial Economics, 121(3), 569 – 594. Abdymomunov, A. (2014): “Banking Sector Operational Losses and Macroeconomic Environment,” Unpublished paper. Acharya, V. V., T. Eisert, C. Eufinger, and C. Hirsch (2014): “Real Effects of the Sovereign Debt Crisis in Europe: Evidence from Syndicated Loans,” CEPR Discussion Papers 10108, C.E.P.R. Discussion Papers. Acharya, V. V., M. Richardson, S. V. Nieuwerburgh, and L. J. White (2011): Guaranteed to Fail: Fannie Mae, Freddie Mac, and the Debacle of Mortgage Finance, no. 9400 in Economics Books. Princeton University Press. Acharya, V. V., P. Schnabl, and G. Suarez (2013): “Securitization without risk transfer,” Journal of Financial Economics, 107(3), 515 – 536. Alloway, T. (2014): “Oil price fall starts to weigh on banks,” FT.com [Online; posted 26-November-2014]. Amiti, M., and D. E. Weinstein (2011): “Exports and Financial Shocks,” The Quarterly Journal of Economics, 126(4), 1841–1877. Ashcraft, A. B. (2005): “Are Banks Really Special? New Evidence from the FDICInduced Failure of Healthy Banks,” The American Economic Review, 95(5), 1712–1730. (2006): “New Evidence on the Lending Channel,” Journal of Money, Credit and Banking, 38(3), 751–775. Balasubramanyan, L., A. N. Berger, C. H. Bouwman, and M. M. Koepke (2016): “How Do Lead Banks Use Their Private Information about Loan Quality in the Syndicated Loan Market?,” Working Paper 16-16, Federal Reserve Bank of Cleveland. 37 Bank for International Settlements (2015): “Making supervisory stress tests more macroprudential: Considering liquidity and solvency interactions and systemic risk,” BIS Working Papers 29, Bank for International Settlements. Bernanke, B., and C. S. Lown (1991): “The Credit Crunch,” Brookings Papers on Economic Activity, 22(2), 205–248. Bernanke, B. S., and A. Blinder (1992): “The Federal Funds Rate and the Channels of Monetary Transmission,” American Economic Review, 82(4), 901–921. Berrospide, J. M., and R. Edge (2016): “The Effects of Bank Capital Requirements on Bank Lending: What Can We Learn from the Post-crisis Regulatory Reforms,” Working draft, Federal Reserve Board. Berrospide, J. M., and R. R. Meisenzahl (2015): “The Real Effects of Credit Line Drawdowns,” Finance and Economics Discussion Series 2015-7, Board of Governors of the Federal Reserve System (U.S.). Birge, J. R., and P. Judice (2013): “Long-term bank balance sheet management: Estimation and simulation of risk-factors,” Journal of Banking & Finance, 37(12), 4711 – 4720. Black, L., J. Krainer, and J. Nichols (2016): “From Origination to Renegotiation: A Comparison of Portfolio and Securitized Commercial Real Estate Loans,” The Journal of Real Estate Finance and Economics, pp. 1–31. Bocola, L. (2016): “The Pass-Through of Sovereign Risk,” Journal of Political Economy, 124(4), 879–926. Borio, C., M. Drehmann, and K. Tsatsaronis (2012): “Stress-testing macro stress testing: does it live up to expectations?,” BIS Working Papers 369, Bank for International Settlements. Bottero, M., S. Lenzu, and F. Mezzanotti (2016): “Sovereign Debt Exposure and the Bank Lending Channel: Impact on Credit Supply and the Real Economy,” Joint wb-imf seminar, World Bank - IMF. 38 Calem, P., F. Covas, and J. Wu (2013): “The Impact of the 2007 Liquidity Shock on Bank Jumbo Mortgage Lending,” Journal of Money, Credit and Banking, 45(s1), 59–91. Calem, P. S., and R. F. Sarama (2016): “Why Mortgage Borrowers Persevere: An Explanation of First and Second Lien Performance Mismatch,” Real Estate Economics. Cameron, A. C., and D. L. Miller (2015): “A practitioner’s guide to cluster-robust inference,” Journal of Human Resources, 50(2), 317–372. Cecchetti, S., and A. K. Kashyap (2016): “What binds? Interactions between Bank Capital and Liquidity Regulations,” Unpublished paper. Chava, S., and A. Purnanandam (2011): “The effect of banking crisis on bank-dependent borrowers,” Journal of Financial Economics, 99(1), 116 – 135. Chen, N.-K. (2001): “Bank net worth, asset prices and economic activity,” Journal of Monetary Economics, 48(2), 415 – 436. Chodorow-Reich, G. (2014): “The Employment Effects of Credit Market Disruptions: Firm-level Evidence from the 20089 Financial Crisis,” The Quarterly Journal of Economics, 129(1), 1–59. Chodorow-Reich, G., and A. Falato (2017): “The Loan Covenant Channel: How Bank Health Transmits to the Real Economy,” . Ciccarelli, M., A. Maddaloni, and J.-L. Peydro (2015): “Trusting the bankers: A new look at the credit channel of monetary policy,” Review of Economic Dynamics, 18(4), 979 – 1002. Cingano, F., F. Manaresi, and E. Sette (2016): “Does Credit Crunch Investment Down? New Evidence on the Real Effects of the Bank-Lending Channel,” Review of Financial Studies. Cohen, B. (2013): “How have banks adjusted to higher capital requirements?,” BIS Quarterly Review. 39 Cornett, M. M., J. J. McNutt, P. E. Strahan, and H. Tehranian (2011): “Liquidity risk management and credit supply in the financial crisis,” Journal of Financial Economics, 101(2), 297 – 312. Das, S., and A. N. R. Sy (2012): “How risky are banks’ risk weighted assets? Evidence from the financial crisis,” IMF WP/12/36. De Jonghe, O., H. Dewachter, K. Mulier, S. Ongena, and G. Schepens (2016): “Some Borrowers are More Equal than Others: Bank Funding Shocks and Credit Reallocation,” Unpublished paper. Drechsler, I., T. Drechsel, D. Marques-Ibanez, and P. Schnabl (2016): “Who Borrows from the Lender of Last Resort?,” The Journal of Finance, 71(5), 1933–1974. Duchin, R., and D. Sosyura (2014): “Safer ratios, riskier portfolios: Banks? response to government aid,” Journal of Financial Economics, 113(1), 1 – 28. Duffee, G. R., and C. Zhou (2001): “Credit derivatives in banking: Useful tools for managing risk?,” Journal of Monetary Economics, 48(1), 25 – 54. Efing, M. (2015): “Arbitraging the Basel securitization framework: Evidence from German ABS investment,” Discussion paper. Epouhe, O., and A. Hall (2016): “Payment shock in {HELOCs} at the end of the draw period,” Journal of Economics and Business, 84, 131 – 147, Special Issue on Regulating Consumer Credit. Gertler, M., and P. Karadi (2011): “A model of unconventional monetary policy,” Journal of Monetary Economics, 58(1), 17 – 34, Carnegie-Rochester Conference Series on Public Policy: The Future of Central Banking April 16-17, 2010. Gilchrist, S., and E. Zakrajšek (2012): Bank Lending and Credit Supply Shockspp. 154–176. Palgrave Macmillan UK, London. Greene, W. (2004): “Fixed effects and bias due to the incidental parameters problem in the Tobit model,” Econometric reviews, 23(2), 125–147. 40 Greenspan, A. (1999): “Do efficient financial markets mitigate financial crises?,” Speech before the 1999 Financial Markets Conference of the Federal Reserve Bank of Atlanta, Sea Island, Georgia. Greenstone, M., A. Mas, and H.-L. Nguyen (2014): “Do Credit Market Shocks affect the Real Economy? Quasi-Experimental Evidence from the Great Recession and Normal Economic Times,” Working Paper 20704, National Bureau of Economic Research. Halaj, G. (2013): “Optimal asset structure of a bank - bank reactions to stressful market conditions,” Working Paper Series 1533, European Central Bank. (2016): “Dynamic Balance Sheet Model With Liquidity Risk,” International Journal of Theoretical and Applied Finance (IJTAF), 19(07), 1–37. Hamilton, J. D. (2014): “The Changing Face of World Oil Markets,” NBER Working Papers 20355, National Bureau of Economic Research, Inc. Hancock, D., and J. A. Wilcox (1993): “Has There Been a Capital Crunch in Banking? The Effects on Bank Lending of Real Estate Market Conditions and Bank Capital Shortfalls,” Journal of Housing Economics, 3(1), 31 – 50. Heider, F., F. Saidi, and G. Schepens (2017): “Life Below Zero: Bank Lending Under Negative Policy Rates,” . Holmstrom, B., and J. Tirole (1997): “Financial Intermediation, Loanable Funds, and The Real Sector,” The Quarterly Journal of Economics, 112(3), 663–691. Iacoviello, M. (2015): “Financial Business Cycles,” Review of Economic Dynamics, 18(1), 140–164. Ivashina, V., and D. Scharfstein (2010): “Bank lending during the financial crisis of 2008,” Journal of Financial Economics, 97(3), 319 – 338, The 2007-8 financial crisis: Lessons from corporate finance. Iyer, R., J.-L. Peydro, S. da Rocha-Lopes, and A. Schoar (2014): “Interbank Liquidity Crunch and the Firm Credit Crunch: Evidence from the 2007–2009 Crisis,” Review of Financial Studies, 27(1), 347–372. 41 Jenkins, P. (2014): “Falling oil price poses new threat to banks,” FT.com [Online; posted 15-December-2014]. Jimenez, G., A. R. Mian, J.-L. Peydro, and J. Saurina Salas (2014): “The Real Effects of the Bank Lending Channel,” Unpublished paper. Jimenez, G., S. Ongena, J.-L. Peydro, and J. Saurina (2014): “Credit Demand Forever? On the Strengths of Bank and Firm Balance Sheet Channels in Good and Crisis Times,” Unpublished paper. Johnson, K. W., and R. F. Sarama (2015): “End of the Line: Behavior of HELOC Borrowers Facing Payment Changes,” Working Paper 2015-073, Federal Reserve Board. Kashyap, A. K., and J. C. Stein (1993): “Monetary Policy and Bank Lending,” NBER Working Papers 4317, National Bureau of Economic Research, Inc. Kashyap, A. K., J. C. Stein, and D. W. Wilcox (1993): “Monetary Policy and Credit Conditions: Evidence from the Composition of External Finance,” American Economic Review, 83(1), 78–98. Keys, B. J., T. Mukherjee, A. Seru, and V. Vig (2010): “Did Securitization Lead to Lax Screening? Evidence from Subprime Loans,” The Quarterly Journal of Economics, 125(1), 307–362. Khwaja, A. I., and A. Mian (2008): “Tracing the Impact of Bank Liquidity Shocks: Evidence from an Emerging Market,” American Economic Review, 98(4), 1413–42. Kiyotaki, N., and J. Moore (1997): “Credit Cycles,” Journal of Political Economy, 105(2), 211–248. Lemmon, M., and M. R. Roberts (2010): “The Response of Corporate Financing and Investment to Changes in the Supply of Credit,” Journal of Financial and Quantitative Analysis, 45, 555–587. Loutskina, E. (2011): “The role of securitization in bank liquidity and funding management,” Journal of Financial Economics, 100(3), 663 – 684. 42 Meh, C. A., and K. Moran (2010): “The role of bank capital in the propagation of shocks,” Journal of Economic Dynamics and Control, 34(3), 555 – 576. Mian, A., and A. Sufi (2009): “The Consequences of Mortgage Credit Expansion: Evidence from the U.S. Mortgage Default Crisis,” The Quarterly Journal of Economics, 124(4), 1449–1496. Minton, B. A., R. Stulz, and R. Williamson (2009): “How Much Do Banks Use Credit Derivatives to Hedge Loans?,” Journal of Financial Services Research, 35(1), 1–31. Noonan, L., and B. McLannahan (2014): “US banks hit by cheap oil as Opec warns of long-term low,” FT.com [Online; posted 23-December-2015]. Paravisini, D. (2008): “Local Bank Financial Constraints and Firm Access to External Finance,” The Journal of Finance, 63(5), 2161–2193. Peek, J., and E. S. Rosengren (1995): “The Capital Crunch: Neither a Borrower nor a Lender Be,” Journal of Money, Credit and Banking, 27(3), 625–638. (1997): “The International Transmission of Financial Shocks: The Case of Japan,” The American Economic Review, 87(4), 495–505. Petersen, M. A., and R. G. Rajan (1994): “The Benefits of Lending Relationships: Evidence from Small Business Data,” The Journal of Finance, 49(1), 3–37. Pierret, D. (2015): “Systemic Risk and the Solvency-Liquidity Nexus of Banks,” International Journal of Central Banking, 11(3), 193–227. Schnabl, P. (2012): “The International Transmission of Bank Liquidity Shocks: Evidence from an Emerging Market,” The Journal of Finance, 67(3), 897–932. Shleifer, A., and R. W. Vishny (1992): “ Liquidation Values and Debt Capacity: A Market Equilibrium Approach,” Journal of Finance, 47(4), 1343–66. SNC (2016): “Shared National Credits Program: 1st Quarter 2016 Review,” Working paper series, Board of Governors of the Federal Reserve System and Federal Deposit Insurance Company and Office of the Comptroller of the Currency. 43 Stein, J. C. (1998): “An Adverse-Selection Model of Bank Asset and Liability Management with Implications for the Transmission of Monetary Policy,” The RAND Journal of Economics, 29(3), 466–486. Stein, J. C., and A. K. Kashyap (2000): “What Do a Million Observations on Banks Say about the Transmission of Monetary Policy?,” American Economic Review, 90(3), 407–428. Sufi, A. (2009): “Bank Lines of Credit in Corporate Finance: An Empirical Analysis,” Review of Financial Studies, 22(3), 1057–1088. Van den Heuvel, S. J. (2008): “The welfare cost of bank capital requirements,” Journal of Monetary Economics, 55(2), 298–320. Vickery, J., L. Thomas, and U. Velasquez (2016): “Are Banks Being Roiled by Oil?,” Liberty Street Economics Blog [Online; posted 24-October-2016]. Wasi, N., and A. Flaaen (2015): “Record linkage using Stata: Preprocessing, linking, and reviewing utilities,” Stata Journal, 15(3), 672–697(26). Whited, T. M. (1992): “Debt, Liquidity Constraints, and Corporate Investment: Evidence from Panel Data,” The Journal of Finance, 47(4), 1425–1460. 44 Table 1: Assessing the coverage of the Y14 $ Billions Y-14 total utilized balances Total commercial loans from all BHCs Loans to U.S. nonfinancial corporates 1,170 1,691 2,386 Sources: Federal Reserve Board of Governors Table 2: Bank summary statistics - average values pre-oil shock Commercial Loan Analysis Mean Std. Dev. Total assets ($billions) Non-performing loan ratio (commercial loans) Tier-1 Capital Ratio Mean ROA (quarterly) Equity/Total assets On-balance-sheet total loans/Total assets On-balance-sheet commercial loans/Total on-balance-sheet loans On-balance-sheet residential loans/Total on-balance-sheet loans Total charge-offs/Total on-balance-sheet loans Deposit share of total liabilities Foreign bank dummy Lag ∆ log commercial loans (2012:Q3 to 2014:Q2) Lag ∆ log residential loans (2012:Q3 to 2014:Q2) Lag ∆ log MBS (2012:Q3 to 2014:Q2) Committed O&G loans/Total committed commercial loans Committed O&G loans/Total on-balance-sheet loans Committed O&G loans/Total assets Committed O&G loans/Equity Observations 474 .010 .129 .002 .116 .526 .240 .285 .006 .726 .214 .326 .059 .046 .023 .194 28 675 .007 .020 .001 .023 .225 .104 .135 .005 .231 Residential Loan Analysis Mean Std. Dev. 466 .010 .126 .002 .118 .561 .252 .306 .006 .748 .231 694 .007 .016 .001 .022 .192 .096 .116 .005 .201 -.025 .038 .060 .045 .025 .205 26 .201 .700 .049 .041 .027 .213 .662 .049 .042 .027 .210 Notes: Data are from the Y9-C and are taken as the average from the 2012:Q3 to 2014:Q2 period. Lag ∆ log commercial loans, lag ∆ log residential loans, and lag ∆ log MBS are constructed as the change between 2012:Q3 and 2014:Q2. “Committed O&G loans” are commercial loan commitments made to firms in NAICS codes 211, 213111, and 213112, taken from the Y14 data. “Committed O&G loans/Total committed commercial loans,” “Committed O&G loans/Total on-balance-sheet loans,” “Committed O&G loans/Total assets,” and “Committed O&G loans/Equity” are calculated as the quarterly average over the 2012:Q3 to 2013:Q4 period. “Committed O&G loans/Total committed commercial loans” is the main exposure variable used throughout the paper, referred to as “O&G Exposure.” 10,169 22.9 (47.0) -.05 (.47) 48,739 10.8 (41.2) -.07 (.35) 0.39 (0.49) 0.41 (0.49) Loan Variables 20,006 77,928 18.9 9.9 (41.4) (36.7) 0.26 (0.44) 17,132 0.28 (0.45) 74,407 over the pre-period. total utilized commercial loan exposure in the pre-period (2012:Q3 to 2014:Q2) divided by total committed commercial loan exposure respectively. Access to external finance = dummy for whether the firm has either a CUSIP or a ticker; Share of credit lines utilized = A “loan” is defined as a bank-firm pair. The pre and post data are averaged over 2012:Q3 to 2014:Q2 and 2015:Q1 to 2015:Q3, Notes: Listed are means and standard deviations in parentheses of variables used in the six main commercial loan samples. Entry in post-shock period Exit in post-shock period ∆ Log loan size Number of loans Loan size ($ millions, pre-shock period) Intensive Margin Exit Margin Entry Margin FE Sample OLS Sample FE Sample OLS Sample FE Sample OLS Sample Firm Variables Number of firms 3,541 42,111 7,802 65,724 5,901 60,121 Assets ($ millions) 720 179 569 178 619 182 (745) (455) (711) (454) (722) (479) Access to external finance .56 .11 .44 .11 .47 .10 (.49) (.31) (.50) (.31) (.50) (.30) Share of credit lines utilized (pre-shock period) .72 .89 .74 .88 (.28) (.22) (.29) (.23) Number of bank relationships (pre-shock period) 4.5 1.5 3.6 1.4 (3.2) (1.5) (2.7) (1.4) Table 3: Regression samples - firm and loan characteristics Table 4: Regression samples - residential loans Variable Observations Bank-County Level Variables ∆ Log Total Loans 32,832 ∆ Log Non-Government Portfolio Loans 19,929 ∆ Log Non-Government Portfolio Jumbo Loans 5,555 ∆ Log Non-Government Portfolio Non-Jumbo Loans 18,913 ∆ Log Government Portfolio Loans 3,639 ∆ Log Non-Government Non-Portfolio Loans 24,366 ∆ Log Non-Government Non-Portfolio Non-Jumbo Loans 22,892 County Level Variables ∆ Log Total LPS Loans ∆ Log Total Y-14 Loans ∆ Log Total Portfolio Y-14 Loans 2,938 2,943 2,943 Mean Std. Dev. -0.18 -0.16 0.16 -0.12 0.02 -0.09 -0.06 0.54 0.64 0.61 0.67 0.87 0.52 0.55 -0.11 -0.16 -0.17 0.14 0.12 0.41 Notes: Listed are the summary statistics of dependent variables used in the residential loan regression analysis. For each category, we aggregate the data by county-bank-quarter, and then average over the preshock and post-shock periods (defined as 2012:Q3 to 2014:Q2 and 2015:Q1 to 2015:Q3, respectively), then take the difference of the log value. County-level variables are constructed in an analogous fashion at the county level. Table 5: The bank lending channel - commercial loans, intensive margin FE O&G Exposure -0.825*** (0.162) Log Total Assets 0.060*** (0.009) ROA -14.229*** (3.994) Equity/Total Assets -1.308*** (0.486) Lag ∆ Log Commercial Loans -0.054 (0.040) NPL/Total Assets -4.523** (2.227) Tier-1 Capital/RWA 1.145 (0.729) Total Loans/Total Assets 0.457*** (0.116) Charge Offs/Total Loans -7.117*** (1.435) Commercial Loans/Total Loans -0.615*** (0.116) Residential Loans/Total Loans -0.713*** (0.181) Deposits/Total Liabilities 0.123** (0.048) Foreign Bank Dummy 0.230*** (0.038) Number of Observations 10169 R-squared 0.59 Fixed Effects Firm ∆ Log loan Size OLS OLS -0.578*** -0.323*** (0.170) (0.058) 0.040*** 0.035*** (0.009) (0.004) -6.447* -7.502*** (3.780) (1.437) -0.219 -0.979*** (0.534) (0.145) -0.041 -0.009 (0.032) (0.016) -3.805** -1.790** (1.794) (0.746) 0.410 1.328*** (0.658) (0.205) 0.191* 0.207*** (0.109) (0.042) -2.727** -1.851** (1.217) (0.736) -0.477*** -0.309*** (0.107) (0.042) -0.542*** -0.509*** (0.172) (0.053) 0.092* 0.150*** (0.048) (0.027) 0.174*** 0.143*** (0.034) (0.012) 10169 48739 0.18 0.06 Firm Controls Firm Controls Notes: The dependent variable is the change in log term-loan size. All quarterly data for a given loan were collapsed to a single pre and post period, defined as 2012:Q3 to 2014:Q2 and 2015:Q1 to 2015:Q3, respectively. We removed from the sample loans to firms in the oil and gas industry and data was restricted to term loans for consecutive quarters throughout both the pre- and post period (i.e., no exiting and re-entering). O&G exposure is constructed as total committed loans to oil and gas firms divided by total committed commercial loans in 2012 and 2013. Column 1 is run on the sample of firms that borrow term loans from multiple banks and include firm fixed effects. Column 2 is run on the same sample of firms as column 1 (but excluding firms in the O&G industry) and includes firm controls instead of firm fixed effects. Column 3 includes firms that borrow from single banks and includes firm controls. All bank controls listed are measured in the pre-shock period. Firm controls in columns 2 and 3 include industry-times-region fixed effects (across 322 industries—3-digit NAICS codes, and 9 regions—the first digit of firm’s zip code) and a dummy indicating whether the firm borrows from multiple banks. Standard errors in parentheses are two-way clustered at the bank (28 banks in total) and firm level. ∗ p < 0.10,∗∗ p < 0.05,∗∗∗ p < 0.01. Table 6: The bank lending channel- commercial loans, extensive margin Exit? FE OLS O&G Exposure 0.838*** 0.739*** (0.158) (0.193) Log Total Assets -0.027** 0.005 (0.011) (0.011) ROA 8.047 4.836 (6.041) (7.436) Equity/Total Assets 0.270 -0.326 (0.749) (0.859) Lag ∆ Log Commercial Loans 0.064* 0.059* (0.036) (0.036) NPL/Total Assets 3.159 2.233 (1.993) (2.412) Tier-1 Capital/RWA 1.241** 2.392*** (0.579) (0.792) Total Loans/Total Assets -0.215 0.047 (0.152) (0.125) Charge Offs/Total Loans 0.737 -1.640 (1.971) (1.939) Commercial Loans/Total Loans 0.475*** 0.477*** (0.138) (0.154) Residential Loans/Total Loans 0.486*** 0.295 (0.160) (0.187) Deposits/Total Liabilities -0.112 -0.119 (0.068) (0.077) Foreign Bank Dummy -0.132*** -0.095** (0.038) (0.038) Number of Observations 20006 20006 R-squared 0.60 0.11 Fixed Effects Firm Firm Controls OLS 0.109 (0.072) 0.012* (0.007) 1.044 (2.932) -0.419 (0.327) 0.011 (0.022) 3.080** (1.511) 1.247*** (0.417) 0.028 (0.060) -0.010 (1.255) 0.143* (0.075) 0.075 (0.094) 0.034 (0.051) -0.011 (0.021) 77928 0.04 Firm Controls FE -0.106 (0.176) -0.050*** (0.012) 6.584 (5.414) 0.718 (0.622) 0.053 (0.033) 7.356*** (1.942) -0.352 (0.885) -0.129 (0.139) -0.874 (1.815) -0.191 (0.142) 0.163 (0.177) -0.141** (0.060) -0.039 (0.042) 17132 0.56 Firm Entry? OLS 0.042 (0.210) -0.039** (0.017) 1.796 (6.627) -0.254 (0.741) 0.060 (0.045) 9.295*** (2.385) 0.312 (1.340) 0.065 (0.155) -2.440 (2.564) -0.157 (0.177) 0.179 (0.194) -0.215** (0.097) -0.037 (0.044) 17132 0.12 Firm Controls OLS 0.199 (0.278) -0.053** (0.026) 6.187 (7.792) 0.783 (0.840) 0.039 (0.060) 12.714*** (2.850) -0.277 (1.658) -0.172 (0.190) -3.206 (2.837) 0.154 (0.253) 0.468** (0.183) -0.339*** (0.121) -0.128*** (0.042) 74407 0.11 Firm Controls Notes: The dependent variable in columns 1-3, “exit,” is an indicator whether the term loan is not renewed and the firm exits its banking relationship in the post-shock period. The dependent variable in columns 4-6, “entry,” is an indicator if the term loan was made for the first time in the post-shock period. We removed from the sample loans to firms in the oil and gas industry. O&G exposure is constructed as total committed loans to oil and gas firms divided by total committed commercial loans in 2012 and 2013. Column 1 is run on the sample of firms that borrow term loans from multiple banks in the pre-period and include firm fixed effects. Column 2 is run on the same sample of firms as column 1 (but excluding firms in the O&G industry) and includes firm controls instead of firm fixed effects. Column 3 includes firms that borrow from single banks and includes firm controls. Column 4 is run on the sample of firms that borrow from multiple banks in the post-period and includes firm fixed effects. All bank controls listed are measured in the pre-shock period. Firm controls include industry-times-region fixed effects (across 322 industries—3-digit NAICS codes, and 9 regions—the first digit of firm’s zip code) and a dummy indicating whether the firm borrows from multiple banks. Standard errors in parenthesis are two-way clustered at the bank (28 banks in total) and firm level. ∗ p < 0.10,∗∗ p < 0.05,∗∗∗ p < 0.01. Table 7: The bank lending channel - commercial loans, interaction effects O&G Exposure O&G Exposure*(Small Firm) FE -0.904*** (0.195) 0.190 (0.217) Small Firm O&G Exposure*(External Finance Firm) ∆ Log loan size OLS FE -0.376* -1.085** (0.207) (0.482) 0.065 (0.228) -0.097*** (0.015) 0.043 (0.244) External Finance Firm O&G Exposure*(Pre-period # of Banks) -0.017 (0.026) Pre-period # of Banks O&G Exposure*(Pre-period Utilization Rate) 0.496 (0.483) Pre-period Utilization Rate Number of Observations R-squared Fixed Effects 10169 0.59 Firm 48739 0.06 Firm Controls 10169 0.59 Firm Exit? OLS -0.129 (0.365) 0.138 (0.125) -0.003 (0.010) -0.025 (0.034) 0.007*** (0.003) -0.143 (0.354) -0.092*** (0.023) 48739 0.06 Firm Controls FE 1.038*** (0.209) -0.396* (0.226) OLS 0.688** (0.278) -0.635** (0.297) 0.157*** (0.017) FE 0.930** (0.435) OLS 0.327 (0.310) 0.044 (0.292) 0.273 (0.339) -0.033* (0.019) 0.054** (0.027) -0.024*** (0.002) -0.378 (0.339) -0.153*** (0.023) 77928 0.05 Firm Controls 0.047* (0.026) -0.585 (0.426) 20006 0.60 Firm 77928 0.05 Firm Controls 20006 0.60 Firm Notes: The dependent variable for columns 1 through 4 is the change in log term-loan size and for columns 5 through 8 is an indicator whether the term loan is not renewed and the firm exits its banking relationship in the post-shock period (i.e., “exit.”). All quarterly data for a given term loan were collapsed to a single preand post-shock period, defined as 2012:Q3 to 2014:Q2 and 2015:Q1 to 2015:Q3, respectively. We removed from the sample loans to firms in the oil and gas industry. Data for the fixed-effects regressions were restricted to term loans in the sample for consecutive quarters throughout both the pre- and post-shock period (i.e., no exiting and re-entering). O&G exposure is constructed as total committed loans to oil and gas firms divided by total committed commercial loans in 2012 and 2013. Columns 1 and 3 are run on the sample of firms with term loans from multiple banks and include firm fixed effects. OLS regressions include firms that borrow from single banks and includes firm controls. Firm controls include industry-times-region fixed effects (across 322 industries—3-digit NAICS codes, and 9 regions—the first digit of firm’s zip code) and a dummy indicating whether the firm borrows from multiple banks. Small firm = total borrowing by firm from all banks is in bottom 70%, external finance firm = dummy for whether the firm has either a CUSIP or a ticker; pre-period utilization rate = total utilized commercial loan exposure in the pre-period (2012:Q3 to 2014:Q2) divided by total committed commercial loan exposure over the pre-period; pre-period # of banks = number of banks the firm has at least one loan with in the pre-shock period; length of relationship = years between date of first origination and 2014:Q2. Bank controls include pre-shock period ROA, equity/total assets, NPL/total assets, tier-1 capital/risk-weighted assets, total loans/total assets, charge offs/total loans, commercial loans/total loans, residential loans/total loans, deposits/total liabilities, lag change in log commercial loans (defined between 2012:Q3 and 2014:Q2), and a foreign bank dummy. Standard errors in parentheses are two-way clustered at the bank and firm level (28 banks in total). ∗ p < 0.10,∗∗ p < 0.05,∗∗∗ p < 0.01. Table 8: The credit channel - firm regressions Agg. O&G Exposure Number of Observations R-squared Bank Controls Firm Controls Y14 Y14 & Compustat Term Loans All Loans Liab Eqty Assets 0.090 0.058 -0.122 -0.333 0.141 (0.448) (0.565) (0.396) (0.066) (0.074) 1460 1426 1460 47167 47150 0.06 0.07 0.43 0.37 0.42 Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Notes: These regressions are run at the firm level. We removed from the sample loans to firms in the oil and gas industry. The dependent variable is the change in log total term-loan borrowing (column 1), the change in log total-loan borrowing (column 2), the change in log total liabilities (column 3), the change in log total equity (column 4), and the change in log total assets (column 5) by firm across the pre- and post-shock period. Total term loan borrowing includes all term loans across all banks in the Y14 sample. Total-loan borrowing includes utilized loans of all types (i.e., not just term loans) across all banks in the Y14 sample. Liabilities, equity, and assets are taken from Compustat, limiting the sample of firms. All quarterly data for a given loan were collapsed to a single pre and post period, defined as 2012:Q3 to 2014:Q2 and 2015:Q1 to 2015:Q3, respectively. Firm controls include industry-times-region fixed effects (across 322 industries—3-digit NAICS codes, and 9 regions—the first digit of firm’s zip code) and a dummy indicating whether the firm borrows from multiple banks. Agg-exposure is the loan-size weighted O&G exposure across all banks by the firm. Bank controls include pre-shock period loan-weighted average levels of ROA, equity/total assets, NPL/total assets, tier-1 capital/risk-weighted assets, total loans/total assets, charge offs/total loans, commercial loans/total loans, residential loans/total loans, deposits/total liabilities, lag change in log commercial loans (defined between 2012:Q3 and 2014:Q2), and a foreign bank dummy. Robust standard errors in parentheses. ∗ p < 0.10,∗∗ p < 0.05,∗∗∗ p < 0.01. Table 9: Substitute financing - credit line utilization Agg O&G Exposure Agg O&G Exposure*(Pre-Period Utilization Rate) Pre-Period Utilization Rate Number of Observations R-squared Bank Controls Firm Controls ∆ Utilization Rate OLS OLS 0.018 0.417** (0.056) (0.173) -0.521** (0.213) -0.132*** (0.014) 14205 14205 0.09 0.14 Yes Yes Yes Yes Notes: The dependent variable is the change in the utilization rate by firm. The utilization rate is defined as total utilized commercial loan exposure divided by total committed commercial loan exposure. We removed from the sample loans to firms in the oil and gas industry and restricted the sample to firms with utilization rates less than 100 percent in the pre-period. Firm controls include industry-times-region fixed effects (across 322 industries—3-digit NAICS codes, and 9 regions—the first digit of firm’s zip code) and a dummy indicating whether the firm borrows from multiple banks. Agg-exposure is the loan-size weighted O&G exposure across all banks by the firm. Bank controls include pre-shock period loan-weighted averages of ROA, equity/total assets, NPL/total assets, tier-1 capital/RWA, total loans/total assets, charge offs/total loans, commercial loans/total loans, residential loans/total loans, deposits/total liabilities, lag change in log commercial loans (defined between 2012:Q3 and 2014:Q2), and a foreign bank dummy. Robust standard errors in parentheses. ∗ p < 0.10,∗∗ p < 0.05,∗∗∗ p < 0.01. Table 10: Substitute financing - alternative banks O&G Exposure FE -0.513** (0.247) FE 0.284 (0.200) Agg. O&G Exposure Number of Observations R-squared Fixed Effects Bank Controls Sample 7277 0.41 Firm Yes Firms that exit 8447 0.42 Firm Yes Firms with no exit Entry? FE -0.346* (0.183) FE 0.421 (0.268) Switch? OLS 0.070*** (0.02) 12017 5115 77187 0.15 0.36 0.63 Firm Firm Firm Controls Yes Yes Yes Firms with Firms without All Firms loan declines loan declines Notes: Fixed-effects (FE) regressions are run at the loan level. The dependent variable of the FE regressions is an indicator if the term loan was made for the first time in the post-shock period (i.e., entry). All FE regressions are run with multiple banking relationships using firm fixed effects. Column 1 is run on the sample of firms that exited from at least one banking relationship in the post period. Column 2 is run on the sample of firms that did not exit any banking relationship. Column 3 is run on a sample firms with at least one loan that declined in size. Column 4 is run on a sample of firms without any loans that decreased in size. Bank controls include pre-shock period ROA, equity/total assets, NPL/total assets, tier-1 capital/RWA, total loans/total assets, charge offs/total loans, commercial loans/total loans, residential loans/total loans, deposits/total liabilities, lag change in log commercial loans (defined between 2012:Q3 and 2014:Q2), and a foreign bank dummy. Column 5 is a regression run at the firm level, where the dependent variable is a dummy equal to one if the firm exited at least one banking relationship and entered at least one banking relationship between the pre- and post-shock period. Bank controls are loan-size weighted in the OLS specification. We removed from the sample loans to firms in the oil and gas industry. Standard errors in parentheses are two-way clustered at the bank (28 banks in total) and firm level for FE regressions and robust standard errors reported for the OLS regression. ∗ p < 0.10,∗∗ p < 0.05,∗∗∗ p < 0.01. Table 11: The bank lending channel - residential loans All Non-Government Government Portfolio Loans Non-Portfolio Loans Jumbo Non-Jumbo All Non-Jumbo O&G Exposure -2.247*** (0.538) Number of Observations 19929 R-squared 0.23 Bank Controls Yes Fixed Effects County O&G Exposure -2.223*** (0.531) Number of Observations 19928 R-squared 0.09 Bank Controls Yes Fixed Effects State County Controls Yes Panel A: County Fixed Effects 1.073 -1.451 -2.521** (0.920) (1.019) (0.933) 5555 18913 23353 0.29 0.26 0.23 Yes Yes Yes County County County -1.522* (0.911) 5555 0.09 Yes State Yes Panel B: OLS 0.965 -2.512** (0.924) (1.039) 18912 23351 0.13 0.11 Yes Yes State State Yes Yes All Loans 3.042** (1.185) 23247 0.26 Yes County 8.426*** (1.190) 13712 0.40 Yes County -1.502 (0.944) 29056 0.24 Yes County 2.941** (1.232) 23245 0.14 Yes State Yes 8.295*** (1.133) 13710 0.23 Yes State Yes -1.518 (0.955) 29053 0.16 Yes State Yes Notes: The dependent variable is the change in log residential loans by bank-county. All quarterly data for loans was aggregated by bank-county and then collapsed to a single pre- and post-shock period, defined as 2012:Q3 to 2014:Q2 and 2015:Q1 to 2015:Q3, respectively. O&G exposure is constructed as total committed loans to oil and gas firms divided by total committed commercial loans in 2012 and 2013. In columns 1-4 the dependent variables are log changes in BHC portfolio loans, while columns 5-6 are BHC non-portfolio loans. Portfolio loans are than divided between non-government jumbo loans, non-government non-jumbo loans, and government loans (i.e., FHA and VA loans). We include only loans originated by the BHC (i.e., we removed mortgages with loan source marked as correspondent, servicing rights purchased, and bulk purchased). Bank controls include pre-shock period ROA, equity/total assets, NPL/total assets, tier-1 capital/RWA, total loans/total assets, charge offs/total loans, commercial loans/total loans, residential loans/total loans, deposits/total liabilities, lag change in log residential loan (defined between 2012:Q3 and 2014:Q2), lag change in log MBS (defined between 2012:Q3 and 2014:Q2), and a foreign bank dummy. County controls include pre-period log population, log population density, percent veterans, log housing density, percent urban housing units, percent occupied housing units, percent vacant housing units, log median house value, log median rental value, fraction of housing with 3 or more individuals, percent in poverty, unemployment rate, percent disabled, log median household income. Standard errors in parentheses are two-way clustered at the bank (26 banks in total) and county level. ∗ p < 0.10,∗∗ p < 0.05,∗∗∗ p < 0.01. Table 12: The credit channel - county regressions Y-14 LPS Portfolio All Aggregate O&G Exposure -0.477 -0.263 -0.105 (0.901) (0.428) (0.532) Number of Observations 2951 3010 3003 R-squared 0.15 0.10 0.08 Yes Yes Bank Controls Yes Fixed Effects State State State Yes Yes County Controls Yes Notes: The dependent variable is the change in log residential loans by county. All quarterly data for loans were aggregated by county and then collapsed to a single pre- and post-shock period, defined as 2012:Q3 to 2014:Q2 and 2015:Q1 to 2015:Q3, respectively. In column 1, we include only portfolio loans, and in column 2 all Y14 mortgages. In column 3, we include all loans in the LPS McDash Analytics database. Agg-exposure is the loan weighted O&G exposure across all banks in the county. Bank controls include pre-shock period (weighted to county level) ROA, ROE, NPL/total assets, tier-1 capital/RWA, total loans/total assets, charge offs/total loans, commercial loans/total loans, residential loans/total loans, deposits/total liabilities, lag change in log residential loan (defined between 2012:Q3 and 2014:Q2), lag change in log MBS (defined between 2012:Q3 and 2014:Q2), and a foreign bank dummy. County controls include pre-period log population, log population density, percent veterans, log housing density, percent urban housing units, percent occupied housing units, percent vacant housing units, log median house value, log median rental value, fraction of housing with 3 or more individuals, percent in poverty, unemployment rate, percent disabled, log median household income. Robust standard errors in parentheses. ∗ p < 0.10,∗∗ p < 0.05,∗∗∗ p < 0.01. Table 13: The bank lending channel- instrumental variables ∆ Log Market Cap O&G Exposure Number of Observations R-squared (second-stage) Bank Controls Fixed Effects ∆ Log Market Cap O&G Exposure Number of Observations R-squared (second-stage) Bank Controls Fixed Effects A: Commercial Loans ∆ Log Exit? Entry? loan size FE-IV FE-IV FE-IV Second Stage 0.901*** -0.953** 0.056 (0.328) (0.410) (0.209) First Stage -0.956*** -0.907** -0.907** (0.368) (0.353) (0.367) 9719 19222 16412 0.60 0.61 0.57 Yes Yes Yes Firm Firm Firm B: Residential Loans Non-Government Government Portfolio Non-Portfolio FE-IV FE-IV FE-IV Second Stage 1.812*** -0.581* -3.453*** (0.630) (0.316) (0.436) First Stage -1.427*** -2.772*** -2.679*** (0.448) (0.316) (0.159) 19451 22830 13514 0.24 0.24 0.41 Yes Yes Yes County County County All Loans FE-IV 1.181 (0.775) -1.749*** (0.486) 28378 0.25 Yes County Notes: All regressions were run under two-stage least squares using O&G exposure as an instrument for the change in bank’s market capitalization—defined as the difference in log average quarterly market capitalization in the post-period (2015:Q1 and 2015:Q3) and pre-period (2012:Q3 and 2014:Q2). The dependent variables are the change in log term-loan size (column 1, panel A), an indicator whether the term loan is not renewed and the firm exits its banking relationship in the post period (column 2, panel A), and an indicator if the term loan was made for the first time in the post-period (column 3, pane A), and the change in log residential loans (originated by the BHC) by county and type of loan (panel B). Bank controls include pre-shock period ROA, equity/total assets, NPL/total assets, tier-1 capital/RWA, total loans/total assets, charge offs/total loans, commercial loans/total loans, residential loans/total loans, deposits/total liabilities, and a foreign bank dummy. Residential loan regressions in include the lag change in log residential loan (defined between 2012:Q3 and 2014:Q2) and lag change in log MBS (defined between 2012:Q3 and 2014:Q2) as controls, and the commercial loan regressions include the lag change in log commercial loans (defined between 2012:Q3 and 2014:Q2) as a control. Standard errors in parentheses are two-way clustered at the bank and firm level. ∗ p < 0.10,∗∗ p < 0.05,∗∗∗ p < 0.01. Table 14: Robustness - alternative exposure measures A: Commercial Loans ∆ Log loan size Exit? Committed O&G/Total Loans -1.015*** 1.282*** (0.239) (0.197) Committed O&G/Total Assets -1.649*** 2.076*** (0.358) (0.317) 0.270*** Committed O&G/Equity -0.190*** (0.054) (0.040) 20006 20006 20006 Number of Observations 10169 10169 10169 0.60 0.60 0.60 R-squared 0.59 0.59 0.59 Fixed Effects Firm Firm Firm Firm Firm Firm Yes Yes Yes Bank Controls Yes Yes Yes All Committed O&G/Total Loans Number of Observations R-squared Bank Controls Fixed Effects Committed O&G/Total Assets Number of Observations R-squared Bank Controls Fixed Effects Committed O&G/Equity Number of Observations R-squared Bank Controls Fixed Effects B: Residential Loans Non-Government Portfolio Loans Non-Portfolio Loans Jumbo Non-Jumbo All Non-Jumbo Government All Loans -5.103*** -1.536 (0.730) (1.148) 19929 5555 0.24 0.28 Yes Yes County County -7.316*** (1.544) 18913 0.27 Yes County 4.384*** (1.701) 23353 0.24 Yes County 4.621** (2.067) 23247 0.26 Yes County 12.507*** (1.204) 13712 0.41 Yes County 1.287 (1.672) 29056 0.24 Yes County -6.335*** -3.150 (1.271) (2.096) 19929 5555 0.23 0.28 Yes Yes County County -8.726*** (2.617) 18913 0.27 Yes County 4.077 (2.814) 23353 0.23 Yes County 6.669* (3.606) 23247 0.26 Yes County 21.341*** (1.606) 13712 0.41 Yes County 0.255 (2.077) 29056 0.24 Yes County -0.865*** -0.342 (0.180) (0.279) 19929 5555 0.23 0.28 Yes Yes County County -1.163*** (0.397) 18913 0.27 Yes County 0.847*** (0.295) 23353 0.24 Yes County 1.164*** (0.319) 23247 0.26 Yes County 2.787*** (0.264) 13712 0.41 Yes County 0.169 (0.251) 29056 0.24 Yes County Notes: These tables show the main estimates under three alternative O&G exposure variables: (1) committed O&G commercial loans divided by total on-balance sheet loans, (2) committed O&G commercial loans divided by total assets, and (3) committed O&G commercial loans divided by total equity capital. The dependent variables are the change in log term-loan size (columns 1 and 2, panel A), an indicator whether the term loan is not renewed and the firm exits its banking relationship in the post period (columns 3 and 4, panel A), and the change in log residential loans (originated by the BHC) by county and type of loan (panel B). Bank controls include pre-shock period ROA, equity/total assets, NPL/total assets, tier1 capital/RWA, total loans/total assets, charge offs/total loans, commercial loans/total loans, residential loans/total loans, deposits/total liabilities, and a foreign bank dummy. Residential loan regressions in include the lag change in log residential loan (defined between 2012:Q3 and 2014:Q2) and lag change in log MBS (defined between 2012:Q3 and 2014:Q2) as controls, and the commercial loan regressions include the lag change in log commercial loans (defined between 2012:Q3 and 2014:Q2) as a control. Standard errors in parentheses are two-way clustered at the bank and firm level. ∗ p < 0.10,∗∗ p < 0.05,∗∗∗ p < 0.01. Standard errors in parentheses are two-way clustered at the bank (28 banks in total) and firm level. ∗ p < 0.10,∗∗ p < 0.05,∗∗∗ p < 0.01. Figure 1: Oil Prices and Problem Loans by Industry Dollars per barrel 100 12.0% WTI Price 10.0% 80 8.0% 60 6.0% 40 4.0% O&G Firms Non O&G Firms O&G Net-using Firms 20 2.0% 0 0.0% Percentage of Loans Past Due, Charged Off, or in Non-accrual Status 120 Notes: The red dashed line represents the price per barrel of crude West Texas Intermediate oil, collapsed to the quarterly level. Blue bars represent the percentage of loans (i.e., including undrawn commitments) to oil and gas firms (defined by NAICS codes 211, 213111, and 213112) that are either past due, charged off, or in non-accrual status (referred to as “problem loans”). Gray bars represent the percentage of loans to all non oil and gas firms that are problem loans. Orange bars represent the percentage of loans to all net users of the oil and gas industry—defined as firms in industries with less than -0.10 in net trade (i.e., make minus use) with the oil and gas industry, as defined by the 2007 BEA make and use tables—that are problem loans. Figure 2: Net Trade with NAICS 324, 211, and 213111 as a Share of Total Output by Industry 1 0.8 0.6 Net Trade Ratio 0.4 0.2 0 -0.2 -0.4 369 industries between -0.10 and 0 -0.6 -0.8 Notes: Bars represent net trade (that is, make minus use) of each industry with NAICS 324, 211, 213111 (petroleum refineries, oil and gas extraction, and drilling oil and gas wells, respectively), based on the 2007 Bureau of Economic Analysis (BEA) make and use tables (https://www.bea.gov/industry/io_annual. htm). BEA provides make and use matrix tables for 389 industries, where each industry is mapped into a three, four, or five digit NAICS code. NAICS code 213112 (support activities for oil and gas operations), included in our measure of oil and gas exposure, is not separately included in the BEA input-output table. We omit the (many) industries for which this net position is negligible. Figure 3: Stock price and market capitalization - dependence on oil exposure B: Market Capitalization A: Stock Price 30% 30% Percent Deviation from 2014:2 20% Exposure Below 25th Percentile 20% 10% 10% 0% 0% -10% -10% -20% -30% Exposure Below 25th Percentile -20% Exposure Above 75th Percentile -30% Exposure Above 75th Percentile -40% -40% -50% -50% -60% 2012:01 2012:07 2013:01 2013:07 2014:01 2014:07 2015:01 2015:07 2016:01 -60% 2012:01 2012:07 2013:01 2013:07 2014:01 2014:07 2015:01 2015:07 2016:01 Notes: This figure illustrates the impact of a bank’s oil and gas extraction (O&G) exposure to bank stock prices (panel A) and market capitalization (panel B). For each bank in each month, we calculate the deviation of its stock price or market capitalization from 2014:M6. The red dashed line represents the average deviation for banks in the upper-quarter in terms of exposure to the O&G sector, the blue dashed line represents the average deviation for banks in the bottom quartile. Figure 4: Behavior of bank balance sheet variables C: Agency MBS .75 .12 B: Residential Loans .5 .25 0 0 -.12 -.08 -.04 -.25 0 -.12 -.08 -.04 .04 .04 .08 .08 .12 A: Commercial & Industrial Loans D: Risk-Weighted Asset Ratio E: Total Assets .08 0 -.12 -.08 -.04 .04 .08 .04 0 -.12 -.08 -.04 2012.3 2013.1 2013.3 2014.1 2014.3 2015.1 2015.3 2016.1 2012.3 2013.1 2013.3 2014.1 2014.3 2015.1 2015.3 2016.1 .12 2012.3 2013.1 2013.3 2014.1 2014.3 2015.1 2015.3 2016.1 .12 2012.3 2013.1 2013.3 2014.1 2014.3 2015.1 2015.3 2016.1 2012.3 2013.1 2013.3 2014.1 2014.3 2015.1 2015.3 2016.1 Notes: This figure illustrates the impact of a bank’s oil and gas extraction (O&G) exposure to banks balance sheet variables. All data are from the FR-Y9. For each bank, we regress the logarithm of the bank-level variable on time dummies, time dummies interacted with O&G exposure, bank fixed effects, and bankspecific time trends. Shown are the coefficients and 90 percent confidence bands of the coefficients on the time dummies interacted with O&G exposure. Omitted periods in the regression are 2012:Q3 and 2014:Q2, implying bank-specific time trends capture pre-shock period trends. O&G exposure was normalized to have a mean of one standard deviation. Standard errors are clustered by bank. To correct for a coding change in risk weights implemented by Basel III in 2015:Q1, the risk-weighted asset ratio (panel D) for 2015:Q1 for each bank was spliced with its 2014:Q4 value. Figure 5: Bank Lending Standards for Commercial Loans- dependence on oil exposure. 6 A: Loan Covenants Index of Cumulative Tightening Relative to 2014:2 5 4 3 2 Exposure Above 75th Percentile 1 0 -1 -2 Exposure Below 25th Percentile -3 6 B: Loan Rates Index of Cumulative Tightening Relative to 2014:2 5 4 3 2 1 Exposure Above 75th Percentile 0 -1 -2 -3 Exposure Below 25th Percentile 2012:3 2012:4 2013:1 2013:2 2013:3 2013:4 2014:1 2014:2 2014:3 2014:4 2015:1 2015:2 2015:3 2015:4 2016:1 2016:2 2016:3 Notes: This figure illustrates the impact of a bank’s oil and gas extraction (O&G) exposure to banks lending standards for commercial loans. All data are from the Senior Loan Officer Opinion Survey (SLOOS). For each bank we measured the cumulative amount of survey responses in which the response was a tightening of the specified variable—loan covenants (panel A) or loan rates (panel B)—relative to 2014:Q2. The reddotted line represents the average for banks in the upper-quarter in terms of exposure to the O&G sector, the blue-dotted line represents the average for banks in the bottom quartile. Figure 6: Bank Lending Standards for Mortgages- dependence on oil exposure. A: GSE Eligible Exposure Below 25th Percentile -0.4 -0.6 -0.8 -1 -1.2 Exposure Above 75th Percentile -1.4 -1.6 -1.8 Index of Cumulative Tightening Relative to 2015:1 Index of Cumulative Tightening Relative to 2015:1 -0.2 -2 Exposure Below 25th Percentile -0.2 -0.4 -0.6 -0.8 Exposure Above 75th Percentile -1 -1.2 -1.4 C: Non-Jumbo, Non-GSE Eligible 0.2 Exposure Above 75th Percentile 0.1 0 -0.1 Exposure Below 25th Percentile -0.2 -0.3 D: Jumbo 0 -0.4 -0.5 Index of Cumulative Tightening Relative to 2015:1 0.3 Index of Cumulative Tightening Relative to 2015:1 B: Government 0 0 -0.2 Exposure Above 75th Percentile -0.4 -0.6 Exposure Below 25th Percentile -0.8 -1 -1.2 -0.6 2015q1 2015q2 2015q3 2015q4 2016q1 2016q2 2016q3 2015q1 2015q2 2015q3 2015q4 2016q1 2016q2 2016q3 Notes: This figure illustrates the impact of a bank’s oil and gas extraction (O&G) exposure to bank lending standards for mortgages. All data are from the Senior Loan Officer Opinion Survey (SLOOS) questions about qualifying mortgages. A positive slope indicates tightening. For each bank we measured the cumulative amount of survey responses in which the response was a tightening of the specified variable relative to 2015:Q1, the first date available. The GSE-eligible category of residential mortgages includes loans that meet the underwriting guidelines, including loan limit amounts, of the GSEs—Fannie Mae and Freddie Mac. The government category of residential mortgages includes loans that are insured by the Federal Housing Administration, guaranteed by the Department of Veterans Affairs, or originated under government programs, including the U.S. Department of Agriculture home loan programs. The non-jumbo, non-GSEeligible category of residential mortgages includes loans that satisfy the standards for a qualified mortgage and have loan balances that are below the loan limit amounts set by the GSEs but otherwise do not meet the GSE underwriting guidelines. The jumbo category of residential mortgages includes loans that satisfy the standards for a qualified mortgage but have loan balances that are above the loan limit amount set by the GSEs. The red dashed line represents the average for banks in the upper-quarter in terms of exposure to the O&G sector, the blue dashed line represents the average for banks in the bottom quartile. Figure 7: Wild Cluster Bootstrap t-Statistic Permutation Test by Industry 5 4 Frequency 3 2 2 Frequency 3 4 5 6 B: Commercial Loans, Extensive Margin with Fixed Effects 6 A: Commercial Loans, Intensive Margin with Fixed Effects KŝůĂŶĚ'ĂƐ 1 0 0 1 KŝůĂŶĚ'ĂƐ -4 -3 -2 -1 0 1 2 Wild Bootstrap t-Statistic by Industry 3 4 -4 -3 -2 -1 0 1 2 Wild Bootstrap t-Statistic by Industry 3 4 Notes: This figure shows histograms of wild cluster bootstrap t-statistics for each of the 50 most common industries (defined by 3-digit NAICS code) in the commercial loan Y14 data. Analogous to the O&G exposure variable used throughout this paper, for each of the other 49 industries we create a bank-specific exposure variable. We then run the fixed fixed-effects specification 50 times—once as our original specification, and 49 subsequent regressions, each time replacing O&G exposure with one of the 49 other industry exposures. tstatistics for each regression were constructed using the wild cluster bootstrap resampling method (clustered by bank) as described in Cameron and Miller (2015). This sequence of steps was performed twice: producing two histograms, one for the intensive margin (Panel A) and one for the extensive margin (Panel B). The wild cluster bootstrap t-statistic for the O&G exposure variable is marked in each panel. Appendix A. Data Appendix In this appendix, we discuss our methods and assumptions involved in constructing the data set of commercial loan and mortgage loan observations from the Y-14 and Y-9 databases maintained by the Federal Reserve Board of Governors (FR). The disaggregated commercial and mortgage loans are obtained from the FR Y-14Q and FR Y-14M database, respectively. Appendix A.1. Commercial Loans Appendix A.1.1. Firm ID Cleaning We use only loans made to U.S. firms. Firms are identified in Y14 by “tax identification number” (TIN), some of which had missing values. We used the field “obligorname” to help fill in those observations with missing TIN information or to identify potential problem TIN information. If an obligorname is associated with a missing TIN for some entry we look to see if, for an exact obligorname match, a TIN is provided anywhere else in the data set.35 If so, we fill in the missing TINS for that obligorname with the modal provided TIN. We corrected the remainder of the missing TINS by the following: • Standardize all obligornames using the STATA command stnd command (Wasi and Flaaen (2015)). • Flag “good TIN”—TINS uniquely associated with a standardized obligorname by a single bank. • Replace missing TINS with matched (by standardized obligorname) good TIN. • Replace remaining missing TINS with a newly created TIN assigned to standardized obligorname. We then corrected TINs associated with more than one standardized obligorname. We replaced these TINS with a matched (by standardized obligorname) “good TIN.” The remaining multiple-obligorname TINs we replaced with a newly created TIN assigned by its standardized obligorname. 35 Obligorname was set to missing if individual, trust, confidential, NA etc. Appendix A.1.2. Loan Level Cleaning We dropped loans if a duplicate loan ID (“internalcreditfacilityid”) was found within a quarter. Any IDs that are inadvertently used by more than one bank were replaced with unique IDs. We define term loans as those loans with credit facility type marked in the Y14 as “Term Loan,” “Term Loan-A,” “Term Loan-B,” “Term Loan-C,” “Term Loan-Bridge,” “Term LoanAsset Based,” and “Term Loan-DIP.” We define loans as ”Line of Credit” as those with facility type marked in the Y14 as “Revolving Credit,” “Revolving Credit Converting to Term Loan,” “Revolving Credit-Asset Based,” “Revolving Credit-DIP,” “Non-revolving Line of Credit,” “Non-revolving Line of Credit Converting to Term Loan,” and “Standby Letter of Credit.” We then cleaned the above definitions via the following: We changed loans labeled as “Term Loan” in Y14 to “Line of Credit” if the loan had a zero interest rate and had any undrawn commitments over the life of loan—these are likely credit lines. We changed loans labeled as “Line of Credit” in Y14 to “Term Loan” if the loan had a non-zero interest rate and had no undrawn commitments over any part of the life of the loan—these are likely term loans. Appendix A.1.3. Industry Codes We focus on three-digit level NAICS codes. If multiple NAICS codes are used for a given TIN, we imposed the modal code. Some banks use SIC or GICS codes, which were converted to NAICS codes using a crosswalk. NAICS code 211 (oil and gas extraction) was augmented with two six-digit codes 213111 (drilling oil and gas wells) and 213112 (support activities for oil and gas operations). To ensure that oil and gas extraction sector firms were labeled correctly we did the following sequence of steps. • Flag suggestive obligornames names—those with any of the following words in the name “gas”, “drill”, “energy”, “oil”, “propane”, “pipe”, “power”, “resource”, “material”, “petrol”, “technologies”, “operating.” • Assign firm to NAICS code 211 if at least 10 percent of its observations have NAICS code 211 and obligorname features any of the suggestive energy related words above. • Assign firm to 211 if at least 10 percent of its observations have a suggestive energy related obligorname and at least one entry has a NAICS code 211. O&G exposure bank level variable is then created. Appendix A.1.4. Final Cleaning The Y14 imposes a $1m materiality threshold from 2012:Q2 onwards. To allow for inflation and strategic behavior around the threshold, we left a buffer thus dropping those loans less than $1.2m. We dropped those observations still missing a TIN. We keep data between 2012:Q3 and 2015:Q3—data before 2012:Q3 include 19 BHCs. Log loan change variables are then created. Appendix A.2. Residential Loans A 10 percent random sample was drawn from the Y-14M. We dropped loans with missing zip codes, origination amounts, and principal balance amounts. We then winsorized principal balances at the 1 percent level. All loans past due were dropped. Appendix A.2.1. Defining Types of Mortgages Government, Portfolio, Non-Portfolio, and BHC origination categorization: • Government Loans: Loan Type (1 = FHA Residential or 2 = VA Residential) • Non-Government Portfolio Loans: Investor Type (7 = Portfolio); Loan Type (NOT 1 = FHA Residential, NOT 2 = VA Residential) • Non-Government Portfolio Loans Originated by BHC: Investor Type (7 = Portfolio); Loan Type (NOT 1 = FHA Residential, NOT 2 = VA Residential); Loan Source (NOT 3 = Correspondent, NOT 4 = Servicing Rights Purchased, NOT 5 = Bulk Purchased) • Non-Government Non-Portfolio Loans: Investor Type (NOT 7 = Portfolio); Loan Type (NOT 1 = FHA Residential, NOT 2 = VA Residential) • Non-Government Non-Portfolio Loans Originated by BHC: Investor Type (NOT 7 = Portfolio); Loan Type (NOT 1 = FHA Residential, NOT 2 = VA Residential); Loan Source (NOT 3 = Correspondent, NOT 4 = Servicing Rights Purchased, NOT 5 = Bulk Purchased) Appendix B. Descriptive Appendix Appendix B.1. Adjusting the firm-level corporate regressions A way to correct for endogeneity bias in the firm-level OLS estimator entails directly comparing the FE and OLS estimates of the loan-level regressions for term loans and adjusting the initial aggregate OLS estimate accordingly (see Jimenez, Mian, Peydro, and Saurina Salas (2014)):36 V ar(Zi ) F − (β̂1,OLS − β̂1,F E ) ∗ , βˆ1F = β̂1,OLS V ar(Z̄i ) (B.1) F where β̂1,OLS , β̂1,F E , and β̂1,OLS are aggregate OLS, loan-level FE, and loan-level OLS estii) represents the ratio of mates of the coefficient on O&G exposure, respectively, and VV ar(Z ar(Z̄i ) the in-sample variance of the loan-level and firm-level O&G exposure variables. Plugging in the coefficients from the intensive FE and OLS estimates gives a correction of -0.25, generating a coefficient of -0.16 on the aggregate term-loan regression coefficient.37 Therefore, the ) implying bias-corrected coefficient is one-fifth the size of the FE intensive effect (= −0.16 −0.83 that firms were able to substitute 80 percent of the decline in loan supply simply from other Y14 banks. The 80 percent number is likely a lower bound, for several reasons. First, even if there was not complete substitution for financing using funds from other Y14 banks, for some firms the shortfall was likely to be partly made up from loans from non-Y14 banks. Second, because firms’ total change in term loans also includes extensive effects (where the bias is apparently less severe), using the coefficients on the intensive regressions likely implies an over correction.38 Finally, we note that the availability of commercial paper facilities and other modes of external finance reinforce the sense in which our adjusted estimate is only a bound. 36 Since our FE loan-level analysis dealt with term loans, this adjustment is not suitable for the other dependent variables in table 8. 37 From table 5, β̂1,F E = −0.83 and β̂1,OLS = −0.58. The standard deviations of O&G exposure in the ar(Zi ) OLS and FE aggregate regressions are both equal to 0.37 implying VV ar( =1 Z̄i ) 38 While one could combine both the intensive and extensive effects by implementing a Tobit model, a fixed effects Tobit estimator does not exist - there is no sufficient statistic allowing fixed effects to be conditioned out of the likelihood function. Including unconditional fixed effects is also not suitable, since it suffers from the incidental parameters problem and is biased. See Greene (2004).
© Copyright 2024 Paperzz