Using Loan Loss Indicators by Loan Type to Sharpen the Evaluation of the Determinants and Implications of Banks Loan Loss Accruals

Transcription

1 ACCOUNTING WORKSHOP Using Loan Loss Indicators by Loan Type to Sharpen the Evaluation of the Determinants and Implications of Banks Loan Loss Accruals By Gauri Bhat* Cox School of Business Southern Methodist University Joshua Lee Florida State University College of Business Stephen G. Ryan Leonard N. Stern School of Business New York University Thursday, April 23 rd, :20 2:50 p.m. Room C06 *Speaker Paper Available in Room 447

2 Using Loan Loss Indicators by Loan Type to Sharpen the Evaluation of the Determinants and Implications of Banks Loan Loss Accruals Gauri Bhat* Cox School of Business Southern Methodist University 6212 Bishop Blvd. Dallas, TX Joshua Lee Florida State University College of Business 821 Academic Way Tallahassee, FL Stephen G. Ryan Leonard N. Stern School of Business New York University 44 West 4th Street, Suite New York, NY August 2014 * Corresponding author. We thank participants of the accounting research seminars at Boston College, Duke University, the University of Houston, the University of Michigan, Southern Methodist University, and the University of Texas at Dallas.

3 Using Loan Loss Indicators by Loan Type to Sharpen the Evaluation of the Determinants and Implications of Banks Loan Loss Accruals Abstract We show that the determinants and economic implications of the primary loan loss indicators reported in financial reports non-performing loans (NPL), the allowance and provision for loan losses (ALL and PLL), and net loan charge-offs (NLCO) vary dramatically across real estate, commercial, and consumer loan types, because these types differ in their homogeneity and collateralization and thus in the measurement of incurred losses under GAAP. Extending Wahlen (1994), we develop and estimate models of the non-discretionary and discretionary determinants of these loan loss indicators by loan type. This analysis indicates that banks exercise of discretion over PLL is largely limited to heterogeneous commercial loans, a small slice of banks loan portfolios. It also provides many insights into the bank-specific and macroeconomic drivers of banks loan loss accruals. We conduct two analyses that illustrate the increased statistical power and construct validity that results from conducting research on the implications of banks loan loss accruals by loan type. First, we show that this approach improves the accuracy of outof-sample predictions of future NLCO, more so for samples of banks that exhibit greater variation in loan portfolio composition from that of the average bank. Second, we show that this approach substantially strengthens Beatty and Liao s (2011) finding that banks that record more timely PLL exhibit lower loan origination procyclicality. While we hand collect banks ALL by loan type from their financial reports, researchers will not have to incur this sizable cost going forward as bank regulatory filings contain this information beginning in Our results illustrate the usefulness of these disaggregated disclosures for future accounting research. JEL: G21, G28, M41, M48 Keywords: provision for loan losses; disaggregation; disclosure; timeliness; procyclicality.

4 Utilizing Loan Loss Indicators by Loan Type to Sharpen the Evaluation of the Determinants and Implications of Banks Loan Loss Accruals 1. Introduction Under SEC Industry Guide 3 historically and ASU more recently, publicly traded banks disclose three loan loss indicators by loan type in their financial reports: (1) nonperforming loans (NPL), a stock measure of severe delinquencies; (2) the allowance for loan losses (ALL), a stock measure of incurred loan losses; and (3) net loan charge-offs (NLCO), a flow measure of realized loan losses. The ALL and NLCO disclosures allow users of financial reports to estimate the provision for loan losses (PLL), a flow measure of incurred loan losses, by loan type as NLCO plus the change in ALL for the type. Thus, in principle, users can conduct all of the standard analyses of banks loan portfolio quality and loan loss reserve adequacy (described in Ryan 2007) by loan type. Due to the high cost of hand collecting this financial report data, however, in practice accounting researchers investigating the determinants and economic implications of banks loan loss accruals typically employ machine-readable databases of bank regulatory filings. While these filings have included NPL and NLCO by loan type since the mid-1990s, they include ALL by type only since the first quarter of For this reason, prior accounting research on banks loan loss accruals generally has been conducted at the bank level. In this paper, we demonstrate the sizeable improvements in statistical power and construct validity that result from conducting this research by loan type, thereby illustrating the usefulness of the newly available ALL disaggregation. Researchers and other users will be able to implement this approach at low cost going forward as data on ALL by loan type accumulates in regulatory databases. 1

5 This paper is motivated by an increased awareness among policymakers and market participants, primarily attributable to the differential timing and magnitude of the losses on different loan types experienced during the financial crisis, that banks losses and credit risk on loans should be disclosed and analyzed by loan type. This awareness has led to two significant new disclosure requirements in financial and bank regulatory reports. First, in July 2010, the FASB issued ASU , which requires lenders to disclose ALL rollforwards and credit quality indicators by loan type in the notes to their financial statements. In comment letters on the exposure draft for ASU , financial analysts encouraged the FASB to provide finer and more standardized disaggregation by loan type. 1 Second, as of the first quarter of 2013, bank regulators require banks with assets exceeding $1 billion to disclose a standardized disaggregation of ALL in their regulatory filings (Schedule HI-C of Form Y9-C for bank holding companies and Schedule RI-C of call reports for chartered banks). This disaggregation requirement reflects both bank regulators demand for supervisory purposes 2 and the expressed preferences of bank researchers and various other constituencies that use bank regulatory data. 3 1 For example, the comment letter on the June 24, 2009 exposure draft for ASU by Elizabeth Mooney of the Investors Technical Advisory Committee states that [i]t would be extremely useful to have a rollforward of the loan loss reserve account that includes the provisions, the losses and the recoveries for each loan... category... [in fact] FASB should require disclosure of this attribution of the loan loss reserve... by (standardized) category. After this statement, Ms. Mooney proposes six separate categories for real estate loans alone. 2 Bank regulators indicate they use these data to evaluate banks loan portfolio quality and loan loss reserve adequacy more accurately when they are not on site at the banks. For example, Robert Canova of the Federal Reserve Bank of Atlanta writes: During the financial crisis, details about the loan portfolio and the related allowance would have been beneficial for examiners and other users to understand the credit quality problems the banks were facing when exam staff could not be at the bank. Financial Update, Viewpoint: New Allowance for Loan Loss Schedule, Spring 2013, 3 An author of this paper participated in one of a series of meetings involving different constituencies at the Federal Reserve Bank of New York in 2012 on the topic of what additional information to include in the Y-9C filings; several participants at that meeting voiced support for disaggregated ALL disclosure. 2

6 We expect this increased awareness and newly available data to lead to significant future research on banks loan loss accruals and credit risk by loan type. Prior accounting research provides evidence that loan type significantly affects banks discretion over and the timeliness of loan loss accruals. For example, Liu and Ryan (1995), Liu, Ryan, and Wahlen (1997), and Liu and Ryan (2006) find that ALL and PLL are timelier and generally less discretionary measures of loan losses for banks with a higher percentage of homogeneous loans (i.e., loans for which loss accruals are estimated at the pool level based on historical loss statistics under FAS 5) than for banks with a high percentage of heterogeneous loans (i.e., loans for which loss accruals are estimated at the individual loan level primarily based on the judgment of loan officers under FAS 5 or FAS 114). Consistent with this differential discretion, Liu and Ryan (1995) find less (more) market anticipation and a negative (positive) market reaction to PLL for banks with a high proportion of homogeneous (heterogeneous) loans. More generally, loan type affects the relations between loan loss accruals and other, less discretionary, loan loss indicators. For example, Ryan (2011) argues and Ryan and Keeley (2013) provide evidence that collateral drives a wedge between NPL and loan loss accruals, reducing the association between these variables. An important implication of this finding for bank researchers is that commonly used measures of PLL timeliness and ALL adequacy based on the relative magnitude or timing of loan loss accruals to NPL are highly sensitive to loan collateralization. To date, researchers interested in demonstrating the effects of loan type typically have partitioned on loan portfolio composition (e.g., Liu and Ryan 1995), while researchers concerned that loan type may confound their findings typically have linearly controlled for loan portfolio composition (e.g., Beck and Narayanamoorthy 2013). At a minimum, these bank-level 3

7 approaches have yielded less interpretable and statistically powerful evidence regarding the determinants and economic implications of banks loan loss accruals than would have been the case had the researchers conducted the analysis by loan type. This may not be a major concern, however, if the tests conducted are adequately interpretable and powerful given the research question posed. In our view, a lack of power likely contributes, however, to the large number of non-results and conflicting results in the extant literature discussed by Ryan (2011) and Beatty and Liao (2013). An even more troublesome issue is that insufficient incorporation of loan type can cause researchers to ascribe results to aspects of banks loan loss accruals, such as discretion or FAS 5 s incurred loss model, that are actually attributable to loan type. Ryan and Keeley (2013) provide evidence that this issue confounds Beck and Narayanamoorthy s (2013) interpretation of their results regarding the effect of SAB 102 on banks loan loss accruals. In our view, the best and perhaps only reliable way to address this issue is to conduct the empirical analysis at the loan-type level. To develop time-series data of reasonable length for all of banks primary loan loss indicators, we hand collect disclosures of publicly traded banks annual ALL by loan type from their Form 10-K filings. To maximize comparability given banks inconsistently disaggregated disclosures, we distinguish total real estate loans, 4 commercial loans, and nonmortgage consumer (hereafter, consumer) loans, which constitute 68.6%, 18.7%, and 9.8%, respectively, of total loans for the average bank. These three types provide a good spread of loan characteristics: real estate loans may be homogeneous (residential) or somewhat to very heterogeneous (commercial) but are always collateralized; commercial loans are heterogeneous and sometimes collateralized; consumer loans are homogeneous and mostly uncollateralized. 4 Banks do not reliably disaggregate the ALL for residential mortgages and commercial real estate loans. 4

8 We first extend Wahlen s (1994) bank-level models of the determinants of ΔNPL, PLL, and NLCO by developing and estimating models of the non-discretionary and discretionary determinants of these loan loss indicators by loan type. Depending on the model, the nondiscretionary determinants we examine include: current and/or lagged NPL and NLCO variables to capture loan performance and historical loss rates, respectively; the level of and growth in loans to capture effects related to loan seasoning; and the percentage growth rates in GDP and home prices to capture macroeconomic conditions. The discretionary determinants we examine include earnings before the PLL and the Tier 1 capital ratio to capture banks management of their reported profitability and regulatory capital, respectively. This descriptive analysis documents many interpretable and economically significant differences in these determinants across loan types, in the process sharpening prior findings in accounting research on banks loan loss accruals. A notable example is that this analysis shows that prior findings of discretionary management of PLL are driven by heterogeneous commercial loans, a small slice of banks loan portfolios. This result is consistent with Amir et al. s (2014) argument and findings that disaggregation improves the detection of earnings management. We then conduct two analyses to illustrate the increased statistical power and construct validity enabled by conducting research on the economic implications of banks loan loss accruals by loan type. The first of these analyses pertains to the longstanding capital markets literature examining the information content of unexpected changes in loan loss indicators (Wahlen 1994). We show that the use of loan-type-level rather than bank-level models improves the accuracy of out-of-sample predictions of future NLCO for the overall sample and particularly for subsamples that exhibit higher absolute differences of individual banks loan portfolio composition from the overall sample average loan portfolio composition. 5

9 The second analysis pertains to the recent literature examining the benefits of timely loan loss provisioning for banks loan origination procyclicality and overall risk management quality (Beatty and Liao 2011, Bushman and Williams 2012, 2013, and Bhat et al. 2014). We measure abnormal PLL timeliness for each loan type held by the bank as the ratio of ALL to the predicted value of ALL for that type. We obtain a bank-level measure of abnormal PLL timeliness by weight-averaging the loan-type-level measures by the bank s loan portfolio composition. We find that the use of a loan-type-level measure rather than a bank-level measure substantially strengthens Beatty and Liao s (2011) finding that banks that record more timely PLL exhibit reduced loan origination procyclicality, particularly during recessions. Generalizing, this analysis indicates that bank-level proxies for PLL timeliness and ALL adequacy are highly sensitive to loan portfolio composition and that loan-type-level approaches significantly improve researchers ability to estimate these constructs and identify their economic implications. 2. Modeling and Making Predictions about the Determinants of Loan Loss Indicators In this section, we first develop empirical models of the non-discretionary and discretionary determinants of three loan loss indicators: the current change in non-performing loans, ΔNPLi,t, the current provision for loan losses, PLLi,t, and the sum of net loan charge-offs over the next two years, NLCOi,t+1,t+2. The models have the same form at the bank level and three loan-type levels. 5 The loan loss indicators and most other scaled variables in the paper are 5 An alternative to our three loan-type-level models for each loan loss indicator (dependent variable) would be a single hybrid model with a bank-level dependent variable and three sets of loan-type-level explanatory variables along with the bank-level and macroeconomic explanatory variables. While we conducted these analyses, we do not tabulate them for three reasons. First, this approach would prohibit us from demonstrating differential discretionary behavior or sensitivities to macroeconomic variables by loan type. Second, this approach would not provide R 2 s for each loan type; we discuss interesting differences in this descriptive statistic across loan types. Third, there is nontrivial multicollinearity across the explanatory variables for the three loan types, which would increase the sampling 6

10 deflated by total assets at the beginning of the earliest year reflected in the definition of the unscaled variables; for example, ΔNPLi,t (PLLi,t) is deflated by total assets at the beginning of year t-1 (t). A variable name followed by (TYPE) means that the variable may be defined at either the bank level or loan-type level. Our models overlap considerably with Wahlen s (1994) bank-level models, although his models include only non-discretionary determinants. We include discretionary determinants in the PLL model to capture findings, mostly subsequent to Wahlen (1994), that banks exercise discretion over loan loss accruals. We also include macroeconomic variables in all models. We then state the primary bases for our predictions regarding how the coefficients on the determinants vary across loan types. We state general predictions rather than making specific predictions for each coefficient in each model for several reasons: the models include multiple correlated determinants, offsetting effects exist for some explanatory variables that render coefficient predictions difficult, and the total number of predictions would be large. We also state general predictions about the effect of loan type on ALL/NPL, a commonly used ratio measure of the adequacy of the ALL Models of the Determinants of the Loan Loss Indicators Wahlen s (1994) bank-level model for ΔNPLi,t regresses this variable on the beginning level of six types of loans, LOANSi,t-1(TYPE), and ΔNPLi,t-1. Our corresponding loan-type-level model is: error in the estimated coefficients, thereby rendering them harder to interpret. We do briefly describe the results of these alternative analyses in footnotes, however. 7

11 ΔNPLi,t(TYPE) = α + β1δnpli,t-1(type) + β2npli,t-2(type) + β3δloansi,t(type) + β4loansi,t-1(type) + β5gdpgrowt + β6δhpit-1+εi,t. (1) Equation (1) regresses ΔNPLi,t(TYPE) on: ΔNPLi,t-1(TYPE) and NPLt-2(TYPE) to capture firstand higher order serial correlations, respectively, of changes in NPL(TYPE); the current change in loans, ΔLOANSi,t(TYPE), and the beginning level of loans, LOANSi,t-1(TYPE), to capture the rates at which newly originated and seasoned loans, respectively, become non-performing; and the percentage growth rates in U.S. GDP during the current year, GDPGROWt, and in the S&P/Case-Shiller U.S. National Home Price Index during the prior year, ΔHPIt-1, to capture macroeconomic effects. 6 Wahlen s (1994) bank-level PLL model regresses PLLi,t on six types of loans, LOANSi,t-1(TYPE), the expected value of ΔNPLi,t from his model of that variable, NPLi,t-1, and ALLi,t-1. Our corresponding loan-type-level model is: PLLi,t(TYPE) = α + β1δnpli,t+1(type) + β2δnpli,t(type) + β3δnpli,t-1(type) + β4nlcoi,t-2,t-1(type) + β5δloansi,t(type) + β6loansi,t-1(type) + β7ebpi,t + β8tier1i,t-1 + β9gdpgrowt + β10δhpit-1+εi,t. (2) Equation (2) regresses PLLi,t(TYPE) on the following non-discretionary determinants: ΔNPLi,t+1(TYPE), ΔNPLi,t(TYPE), and ΔNPLi,t-1(TYPE) to capture the measurement of the timeliness of PLL in terms of its association with future, current, and lagged ΔNPL documented in prior research (e.g., Liu and Ryan 1995, Nichols et al. 2009, and Beatty and Liao 2011); the sum of net loan charge-offs over the prior two years, NLCOi,t-2,t-1(TYPE), to capture historical loss rates; ΔLOANSi,t(TYPE) and LOANSi,t-1(TYPE) to capture the rates at which newly originated and seasoned loans, respectively, incur losses; and GDPGROWt and ΔHPIt-1, to 6 We use one-year lagged changes in house prices because preliminary data analysis indicates the effects of these changes on the loan loss indicators appear with a lag. 8

12 capture macroeconomic effects. Reflecting prior research that provides (somewhat inconsistent) evidence that banks exercise discretion over PLL to manage income and capital (see Ryan 2011 and Beatty and Liao 2013 for recent summaries of this literature), the model also includes two discretionary determinants: the bank s current earnings before the provision for loan losses, EBPi,t, and lagged Tier 1 risk-based capital ratio, TIER1i,t.1. A positive coefficient on EBPi,t is consistent with income smoothing. A negative coefficient on TIER1i,t-1 is consistent with regulatory capital management because the tax effects of ALL increase Tier 2 risk-based capital (up to a maximum of 1.25% of risk-based assets). Equation (2) also overlaps with the PLL models in Laeven and Majnoni (2003), who interpret negative coefficients on ΔLOANSi,t(TYPE), EBPi,t, and GDPGROWi,t as evidence of procyclical loan loss provisioning. Wahlen s (1994) bank-level net loan charge-offs model regresses NLCOi,t on the same explanatory variables as in his PLL model. Our corresponding loan-type-level model is: NLCOi,t+1,t+2 = α + β1δnpli,t-1(type) + β2npli,t-2(type) + β3nlcoi,t-2,t-1(type) + β4alli,t-1(type) + β5plli,t(type) + β6δloansi,t(type) + β7loansi,t-1(type) + β8gdpgrowt + β9δhpit-1+εi,t. (3) Equation (3) regresses NLCOi,t+1,t+2 on: ΔNPLi,t-1(TYPE), NPLi,t-2(TYPE), and NLCOi,t-2,t-1(TYPE) for similar reasons as in the prior two models; beginning-of-year allowance for loan losses, ALLi,t-1(TYPE), and PLLt(TYPE), to capture the fact that banks should and logically must reserve for loans losses at least some period of time before loans are charged off; ΔLOANSi,t(TYPE) and LOANSi,t-1(TYPE) to capture the rates at which newly originated and seasoned loans, respectively, yield realized loan losses; and GDPGROWt and ΔHPIt-1 to capture macroeconomic effects. 9

13 2.2. Predictions about the Determinants of Loan Loss Indicators We now discuss salient differences in the determinants of the three loan loss indicators (ΔNPLi,t, PLLi,t, and NLCOi,t+1,t+2) for the three loan types we examine (real estate, commercial, and consumer). For simplicity, we omit subscripts and the (TYPE) notation on loan-type-level variables in this discussion. We focus on the effects of these differences on the associations captured in Equations (1)-(3), but we also describe the effects on the commonly used allowance for loan losses adequacy ratio, ALL/NPL, when it clarifies the intended points. Many of these differences we describe are observable in the panels of Figure 1, which depict the means of the three loan loss indicators deflated by total loans for the three loan types for each of the sample years. Real estate loans, depicted in Panel A, may be homogeneous (if residential) or somewhat to very heterogeneous (if commercial), but are always collateralized. Commercial loans, depicted in Panel B, are heterogeneous, and while they may be collateralized they are not reliably so. Consumer loans, depicted in Panel C, are homogeneous and typically not collateralized (e.g., credit card, student) or not reliably well collateralized (e.g., automobile). Panel A of Figure 1 also plots the time series of the macroeconomic determinants GDPGROW and ΔHPI. Predictions based on the Degree of Loan Homogeneity Loan homogeneity has the following primary effects on the loan loss indicators that are discussed by Liu and Ryan (1995) and subsequent research. Under FAS 5 banks primarily accrue for credit losses on homogeneous loans at the portfolio level using statistics. In contrast, under FAS 5 and FAS 114, banks primarily accrue for credit losses on heterogeneous loans at the loan level using input from loan officers. Because the use of statistics on large samples of loans 10

14 disciplines loan loss accruals, in general loan homogeneity increases the timeliness of PLL relative to changes in NPL. Hence, we predict that PLL are more strongly associated with current and future NPL changes and less strongly associated with lagged NPL changes for homogeneous real estate and consumer loans than for heterogeneous commercial loans. 7 For similar reasons, loan homogeneity mitigates banks ability to exercise discretion over loan loss accruals. Accordingly, we predict generally stronger positive relationships exist between PLL and both EBP and TIER1 for heterogeneous commercial loans than for real estate and consumer loans. 8 Homogeneous loans are predominantly comprised of consumer loans, including residential mortgages. All else being equal, consumer borrowers default at much higher rates than do commercial borrowers. This is particularly true during the favorable phases of the business cycle because the vast majority of commercial loan defaults occur in recessions or their aftermaths (Ryan 2011 and Ryan and Keeley 2013). These effects can be seen in the trends in PLL and NLCO for commercial and consumer loans reported in Panels B and C of Figure 1, respectively; compare the spikes in commercial loan PLL and NLCO in and after the 2001 recession and the financial crisis to the on-average higher and much smoother paths of these variables for consumer loans. These effects result in a more continuous flow of newly 7 The related literature empirically documents the following phenomenon that may work against finding results consistent with this prediction. Banks are more likely to allow problem heterogeneous loans than problem homogeneous loans to remain current by rolling over the loans or even lending more to the borrowers; for example, research describes how this occurred in less-developed-country lending in the 1980s (Griffin and Wallach 1991 and Elliott, Hanna, and Shaw 1991). If and when banks finally cut off such lending, the affected problem loans will tend to rapidly become NPL. Hence, banks may record PLL for these loans when they make the decision to cut off funding, i.e., while the loans are still performing, in what appears to be, but really is not, timely loan loss provisioning. 8 The related literature empirically documents the following phenomenon that may work against finding results consistent with this prediction. Relative to other loan types, consumer loans tend to experience higher loan losses and to carry higher interest yields to compensate for these losses (Ryan 2011 and Ryan and Keeley 2013). Consumer loans also exhibit more cross-sectional variation in both loan losses and interest yields for these loans; e.g., prime consumer loans experience much lower losses and so have much lower yields than do subprime consumer loans. In cross-sectional regressions, these effects work to yield a strong positive association between PLL and EBP for consumer loans. 11

15 delinquent NPL for consumer loans than for commercial loans, which yields differences in the relationships between NLCO and NPL for the two loan types. Specifically, NLCO reduce NPL for both types of loans, thereby reducing (i.e., making less positive or more negative) the association between NLCO and NPL. This effect is less for consumer loans, however, because charged-off consumer NPL are reliably replaced by a steady stream of new severely delinquent consumer loans. Bank regulatory guidance provides maximums for the number of days past due before homogeneous consumer loans and residential mortgages must be charged-off. Most importantly, closed-end consumer loans must be charged-off no later than 120 days past due, whereas openend consumer loans and residential mortgages must be charged-off no later than 180 days past due. 9 Because loans generally become non-performing at 90 days past due, non-performing closed-end (open-end) consumer loans typically remain non-performing for only 30 (90) days before being charged-off. This fast charge-off requirement artificially reduces NPL relative to PLL and NLCO for consumer loans (Ryan and Keeley 2013), as can be seen in Panel C of Figure 1. This reduction could either increase or decrease the association between NPL and the other loan loss indicators, depending on the extent to which NPL is proportionately biased downward versus becomes noisier (both likely occur). This reduction unambiguously increases ALL/NPL, however, and so we predict this ratio is higher for consumer loans than for the other loan types. Predictions based on Degree of Loan Collateralization Collateral reduces the loss given default on loans and thus drives a wedge between NPL (i.e., the entire book value of severely delinquent loans), and the currently expected and 9 Federal Financial Institutions Examination Council, Uniform Retail Credit Classification and Account Measurement Policy, June 12,

16 subsequently realized losses on those loans (i.e., current ALL or PLL and future NLCO, respectively) (Ryan 2011). This increases NPL relative to these other loan loss indicators for highly collateralized real estate loans relative to other loan types, as can be seen in Panel A of Figure 1. Predictions about Macroeconomic Determinants Loan losses for all three types of loans are sensitive to both GDPGROW and ΔHPI. However, we expect losses on real estate loans to be particularly sensitive to ΔHPI and losses on commercial loans to be particularly sensitive to GDPGROW, as can be seen in Figure 1, Panels A and B. The main difference between GDPGROW and ΔHPI is the periodicity from peak to peak of their cycles. House prices follow very long cycles (e.g., house prices rose essentially uninterrupted from the early 1990s to mid-2006), while business cycles are considerably shorter (e.g., only six years intervened between the 2001 and recessions). For this reason, we predict more positive serial correlation in the loan loss indicators for real estate loans than for other loan types, particularly in models that do not control for ΔHPI. 3. Sample Selection and Descriptive Statistics Table 1 summarizes our sample selection process and its effects on the number of observations. We collect data for all publicly traded commercial bank holding companies ( banks ) from 1994 to 2010 for which all of the variables necessary to estimate Equations (1)- (3) other than ALL by loan type, which we obtain from Form 10-K filings, as described below are available on the Federal Reserve Bank of Chicago s Y-9C files. This yields 6,136 13

17 bank-year observations for 810 unique banks. 10 We remove 123 bank-years for which we cannot match the Y-9C data to Compustat to obtain each bank s CIK number, which we use to obtain Form 10-K filings from the SEC s Edgar system. We search each of these filings for the disclosed ALL disaggregation (the ALL table ), removing 1,085 bank-year observations that do not contain this table; inspection of these filings suggests that these observations usually are small banks that hold only one primary loan type, real estate. Because banks do not report ALL tables using a standardized disaggregation, we examine a disaggregation into three loan types that meshes relatively well with the disclosures in most bank-year filings. Specifically, we classify each loan type disclosed in the ALL table for each bank-year into one of four categories: real estate, commercial, consumer, and other. We do not distinguish residential and commercial real estate loans because banks do not reliably distinguish these loan types in their ALL tables. We do not empirically examine other loans because they constitute only 2.9% of total loans for the median bank; untabulated analysis indicates the loan loss indicators for these mostly heterogeneous other loans have similar implications as those for commercial loans. To ensure the quality and consistency of the ALL data, we remove 59 (239) [93] bankyears that report real estate (commercial) [consumer] loans but no corresponding ALL. We also remove 75 bank-years in which the total ALL reported in the ALL table does not match the corresponding amount reported on either the Y-9C filings or Compustat. Finally, we remove 183 bank-years for which we are unable to compute a PLL for each loan type because the prior-year ALL is missing. Our final sample consists of 4,279 bank-years (519 unique banks). 10 We use the CRSP-FRB link file available from the Federal Reserve Bank of New York to identify public banks. 14

18 We estimate the PLL(TYPE) as the change in the ALL(TYPE) plus NLCO(TYPE) during the year: PLLi,t(TYPE) = ALLi,t(TYPE) ALLi,t-1(TYPE) + NLCOi,t(TYPE) + errori,t. (4) Equation (4) indicates that the estimates of PLL(TYPE) may contain errors. Errors arise when banks engage in transactions that affect ALL such as acquiring other banks, purchasing or selling loans held for investment, 11 and transferring loans between the held-for-investment and held-for-sale categories during a year. Banks report the bank-level error in their Form 10-K filings. We allocate these errors to the PLL for three loan types based on the percentage of the total ALL attributable to the types. Our results are robust to allocating this bank-level error using various methods that put some weight on the change in loans outstanding for the loan types during the year and to deleting observations with errors greater than 10% of the end-of-year ALL. Table 2 reports descriptive statistics for the bank characteristics, loan loss indicators, and macroeconomic variables included in Equations (1)-(3), by loan type where feasible. The bank characteristics and loan loss indicators typically are deflated by total assets. The table also reports these statistics for two traditional ALL adequacy indicators, ALL/NPL and ALL/NLCO. We discuss only the medians of the variables. Median loans are 73% of assets. Median real estate (commercial) [consumer] loans are 50% (12%) [5%] of assets. Growth in real estate loans is considerably higher than growth in the other loan types during the sample period, reflecting the real estate boom that occurred over the majority of the period. Median ALL is 1% 11 Securitizations and other types of sales of loans held for investment may involve retention of interests. Most retained interests are not classified as loans held for investment and thus no allowance for loan losses is recorded for these interests. Exceptions to this statement include retained seller s interests in securitizations of credit card and other revolving loans and the retained portion of loan syndications. 15

19 of assets. The composition of the ALL does not perfectly mirror the composition of the loan portfolio; in particular, reflecting the collateralization of real estate loans, the proportion of the ALL associated with these loans is much smaller than their proportion of banks loans. As mentioned earlier, ALL/NPL exhibits considerable variation across loan types, with this ratio being much higher for consumer loans than for other loan types due to the artificial reduction of NPL by the quick charge-offs of closed-end consumer loans required under bank regulatory guidance. This ratio is also much lower for real estate loans than for other loan types due to the collateralization of mortgages. Subtler variation is evident in ALL/NLCO across loan types. This ratio is somewhat lower for consumer loans than for the other loan types due to the shorter life of most consumer loans. The ratio is also somewhat higher for commercial loans than for the other loan types; this may reflect discretionary overstatement of the ALL for commercial loans because the average life of mortgages is longer than the average life of commercial loans. The pre-provision return on assets, EBP, is quite low at 1.5%. This reflects banks low operating risk and high financial leverage, which yields normal return on equity. Table 3 reports the Pearson correlations of the loan-type-level loan loss indicators. Not surprisingly, the three indicators are much more positively correlated within loan types than across loan types. As is true at the bank level, the correlations of PLL and NLCO for each of the loan types is particularly strong. 4. Empirical Results for the Determinants of the Loan Loss Indicators 4.1. Determinants of ΔNPL i,t Table 4 reports the OLS estimation of Equation (1), the model of the determinants of ΔNPLi,t. For comparison purposes, the first two columns report models regressing ΔNPLi,t for 16

20 total loans on loan loss indicators for total loans and other variables. The first (second) column reports the model excluding (including) the two macroeconomic variables. The third and fourth (fifth and sixth) [seventh and eighth] columns report analogous pairs of models regressing ΔNPLi,t(TYPE) on loan loss indicators for real estate (commercial) [consumer] loan types and other variables. 12 Unless otherwise indicated, we refer to a coefficient as significant if it differs from zero at the 5% level or better in a two-tailed test. We first discuss the estimations of the models of the determinants of ΔNPLi,t for total loans in the first two columns of Table 4. For the model excluding the macroeconomic variables reported in the first column, the coefficient on ΔNPLi,t-1 is significantly positive (t=3.2), indicating positive first-order serial correlation in ΔNPL as in Wahlen (1994). In contrast, the coefficient on NPLi,t-2 is significantly negative (t=-3.5), reflecting the fact that NPL from two years ago have likely been either charged off or cured by the end of the current year. The coefficients on ΔLOANSi,t and LOANSi,t-1 are both significantly positive (t=4.2 and 7.0, respectively), with the latter coefficient being slightly higher, consistent with newly originated and seasoned loans both becoming non-performing at positive rates, but with the rate being slightly higher for seasoned loans. The R 2 of this model is relatively modest at 6.5%. For the model including the macroeconomic variables reported in the second column, the coefficients on both GDPGROWt and ΔHPIt-1 are significantly negative (t=-7.8 and -12.4, respectively), consistent with NPL decreasing during economically more favorable times. The 12 We also estimate but do not tabulate expanded versions of Equation (1) that include bank-level ΔNPL i,t as the dependent variable and the four loan-type-level explanatory variables for each of the three loan types examined in this paper plus other loans, both with and without the two macroeconomic explanatory variables. These expanded models yield coefficients (sign, magnitude, and significance) on the loan-type-level explanatory variables that are very similar to those reported in Table 4 for real estate loans and commercial loans but not for consumer loans. The divergent results for consumer loans likely result from the facts that this loan type constitutes a relatively small percentage of banks total loans (see Table 2, Panel A) and that NPL for consumer loans is artificially low due to fast required charge-offs of retail loans under bank regulatory guidance (discussed in Section 2.2). 17

21 inclusion of the macroeconomic variables reduces the coefficient on ΔNPLi,t-1 to insignificance because these variables, particularly ΔHPIt-1, drive the first-order serial correlation of ΔNPL. In addition, this inclusion doubles the coefficient on ΔLOANSi,t because greater loan origination occurs in more favorable economic times, which depresses the coefficient on this variable in the model without the macroeconomic variables reported in the first column. The inclusion of the macroeconomic variables substantially increases the R 2 from 6.5% to 24.3%. The third and fourth columns of Table 4 report the estimation of the models of the determinants of ΔNPLi,t for real estate loans. Not surprisingly, given that real estate loans constitute over two-thirds of total loans for the average bank, the estimations of these models are similar to the corresponding estimations for total loans. The most notable of the fairly subtle differences is the coefficient on ΔNPLi,t-1 is 19% more positive for real estate loans in the third column than for total loans in the first column, reflecting the greater first-order serial correlation of ΔNPL driven by the extended length of house price cycles. Relatedly, the coefficient on ΔHPIt-1 is 24% more negative for real estate loans in the fourth column than for total loans in the second column, reflecting the greater house price sensitivity of real estate loans. The inclusion of the macroeconomic variables again reduces the coefficient on ΔNPLi,t-1 to insignificance. The fifth and sixth columns of Table 4 report the estimation of the models of the determinants of ΔNPLi,t for commercial loans. The following striking differences arise in these models for commercial loans compared to the corresponding models for real estate loans just discussed. The coefficient on ΔNPLi,t-1 is significantly negative in both the fifth and sixth columns (t=-11.0 and -11.6, respectively), not positive as for real estate loans, consistent with preexisting commercial NPL being charged off or cured within a year without being replaced by NPL for newly delinquent commercial loans. Relatedly, the coefficient on NPLi,t-2 is about two 18

22 and a half times more negative and much more significant (t=-13.7 and -13.1, respectively). In the sixth column, the coefficient on GDPGROWt is more significantly negative (t=-6.7) and the coefficient on ΔHPIt-1 is less significantly negative (t=-2.4), reflecting the greater (lesser) sensitivity of commercial loans to the business cycle (house prices). The incremental R 2 from including the macroeconomic variables is much smaller at 3.1% because house prices have a far more direct relationship to real estate loan performance than does GDP growth to commercial loan performance. The seventh and eighth columns of Table 4 report the estimation of the models of the determinants of ΔNPLi,t for consumer loans. The following striking differences arise in these models compared to those for real estate loans and commercial loans. The coefficients on ΔNPLi,t-1 and NPLi,t-2 are significantly negative at only the 10% level in both columns (e.g., t=-1.8 and -1.8, respectively), consistent with noisier NPL. The coefficient on ΔLOANSi,t is highly significantly positive (t=11.7) in fact, this variable is the primary driver of ΔNPLi,t for consumer loans reflecting the rapid migration of a reliable percentage of newly originated consumer loans to NPL status. The coefficients on the macroeconomic variables have relatively low significance. Specifically, the coefficient on GDPGROWt is less significantly negative (t=- 3.3) than the corresponding coefficients in the fourth column for real estate loans and sixth column for commercial loans. The coefficient on ΔHPIt-1 is less significantly negative (t=-3.8) than the corresponding coefficient for real estate loans. Reflecting this low significance, the incremental R 2 from including the macroeconomic variables of 1.1% is smaller than in the corresponding analyses for the other two loan types. In summary, the determinants of ΔNPLi,t vary dramatically across the three loan types. Most notably, the serial correlation of ΔNPLi,t is positive for real estate loans, insignificant for 19

23 consumer loans, and negative for commercial loans. The stronger macroeconomic determinant of ΔNPLi,t is home price growth for real estate loans and GDP growth for commercial loans; both of these determinants are relatively weak for consumer loans Determinants of PLL i,t Table 5 reports the OLS estimation of Equation (2), the model of the determinants of PLLi,t. The columnar structure of the table is the same as in Table Rather than discussing the estimations of each of the pairs of bank-level or loan-type models in sequence, as we did for the model of the determinants of ΔNPLi,t in Table 4, we instead discuss the estimated coefficients on sets of related determinants across the models in sequence. Due to the inclusion of the three ΔNPL variables in Equation (2), which are highly correlated with the two macroeconomic variables, particularly ΔHPI, the inclusion of the latter variables has relatively subtle effects on the estimation of the PLLi,t models. The most notable of these effects is to strengthen the evidence of discretionary behavior over PLL. Accordingly, we primarily discuss the models with the macroeconomic variables reported in the even-numbered columns of Table 5; unless indicated otherwise, the t-statistics mentioned pertain to those models. We point out the few places where the models without the macroeconomic variables yield notably different results. We first discuss the evidence of discretionary management of PLL. In the bank-level analysis reported in the second column, the coefficient on TIER1 is significantly negative 13 We also estimate but do not tabulate expanded versions of Equation (2) that include bank-level PLL i,t as the dependent variable and the six loan-type-level explanatory variables for each of the three loan types examined in this paper plus other loans and the two discretionary determinants of PLL i,t explanatory variables, both with and without the two macroeconomic explanatory variables. These expanded models yield coefficients (sign, magnitude, and significance) on the loan-type-level explanatory variables that are very similar to those reported in Table 5 for real estate loans and commercial loans but are typically less significant for consumer loans. The less significant results for consumer loans likely result from the facts that consumer loans constitute a relatively small percentage of banks total loans (see Table 2, Panel A). 20

24 (t=-2.8), consistent with the findings of capital management using PLL in most prior studies (e.g., Beatty et al. 1995, Collins et al. 1995, Ahmed et al. 1999, and Bushman and Williams 2012). The coefficient on EBP is insignificantly positive, inconsistent with the finding of income smoothing using PLL in most but not all prior studies (e.g., Moyer 1990, Beatty et al. 1995, Kim and Kross 1998, and Ahmed et al. 1999). As expected given the limited discretion over PLL for homogeneous loans, there is no evidence of either form of discretionary management in the analysis of real estate loans reported in the fourth column, despite the fact that real estate loans comprise more than two-thirds of the median bank s loan portfolio. Similarly, there is no evidence of capital management in the analysis of consumer loans reported in the eighth column, for similar reasons. While there is a significantly positive coefficient on EBP (t=3.6) in the consumer loans regression in the eighth column, as discussed in footnote 8 we interpret this result as attributable to consumer loans on average experiencing higher loan losses and carrying higher interest yields to compensate for these losses, as well as to the cross-sectionally variable losses and yields on these loans. In untabulated analysis, we confirm the validity of this interpretation by replacing EBPi,t with EBPi,t-1 in the bank-level and all three loan-type-level models with macroeconomic variables. We find a significant positive coefficient on EBPi,t-1 in the consumer loans model (t=2.0) but an insignificant coefficient in all of the other models. In contrast, as predicted given banks discretion over PLL for heterogeneous loans, there is strong evidence of discretionary management of PLL for commercial loans to smooth income and to raise regulatory capital. In the sixth column, the coefficient on EBP is significantly positive (t=2.6) and the coefficient on TIER1 is significantly negative (t=-2.9). Interestingly, the results for the bank-level analysis are more similar to the results for the commercial loans 21

25 analysis than for the real estate loans analysis, despite the fact that real estate loans are almost four times larger than commercial loans for the median bank. This implies that prior evidence about discretionary behavior over PLL is driven by a relatively small slice of banks loan portfolios. We now discuss the timeliness of PLL relative to changes in non-performing loans. As previously mentioned, considerable recent research views a higher association between PLLi,t and ΔNPLi,t+1 as an indicator of greater PLL timeliness, although Bushman and Williams (2012) report that this association is negative for their subsample of U.S. banks. The coefficient on ΔNPLi,t+1 is insignificant for the bank-level model reported in the second column (it is significantly positive at the 10% level in the model that excludes the macroeconomic variables in the first column) and for the real estate loans model reported in the fourth column. This coefficient is significantly positive at the 10% level in the commercial loan model reported in the sixth column (t=1.8), perhaps because banks reserve for problem but still performing commercial loans to which they have cut off lending and so expect to become NPL during the next year, as discussed in Footnote 7. In contrast, this coefficient is significantly negative at the 10% level in the consumer loan analysis (t=-1.9); this may reflect the required charge-off of consumer loans thirty to ninety days after they become non-performing. Regardless of the reasons for each of these coefficients, these results indicate that the sign of the coefficient on ΔNPLi,t+1 depends on loan portfolio composition. Relatedly, the insignificant coefficient on ΔNPLi,t+1 in the bank-level model reflects the offsetting of a positive coefficient in the commercial loans model and a negative coefficient in the consumer loans model. Hence, researchers need to exercise care in controlling for loan portfolio composition in order to interpret a higher coefficient on ΔNPLi,t+1 as indicating greater PLL timeliness. 22