Current Expected Credit Losses (CECL) Standard and Banks' Information Production

Finance and Economics Discussion Series Federal Reserve Board, Washington, D.C. ISSN 1936-2854 (Print) ISSN 2767-3898 (Online) Current Expected Credit Losses (CECL) Standard and Banks’ Information Production Sehwa Kim, Seil Kim, Anya Kleymenova, and Rongchen Li 2023-063 Please cite this paper as: Kim, Sehwa, Seil Kim, Anya Kleymenova, and Rongchen Li (2023). “Current Expected Credit Losses (CECL) Standard and Banks’ Information Production,” Finance and EconomicsDiscussionSeries2023-063. Washington: BoardofGovernorsoftheFederalReserve System, https://doi.org/10.17016/FEDS.2023.063. NOTE: Staff working papers in the Finance and Economics Discussion Series (FEDS) are preliminary materials circulated to stimulate discussion and critical comment. The analysis and conclusions set forth are those of the authors and do not indicate concurrence by other members of the research staff or the Board of Governors. References in publications to the Finance and Economics Discussion Series (other than acknowledgement) should be cleared with the author(s) to protect the tentative character of these papers.

Current Expected Credit Losses (CECL) Standard and Banks’ Information Production∗ Sehwa Kim† Seil Kim‡ Anya Kleymenova§ Rongchen Li¶ This version: July 2023 Abstract We examine whether the adoption of the current expected credit losses (CECL) model, which reflects forward-looking information in loan loss provisions (LLP), improves banks’ information production. Consistent with better information production, we find changes in CECL banks’ financial reporting and operations. First, these banks’ loan loss provisions become timelier and better reflect future local economic conditions. Second, CECL banks disclose longer, more forward-looking, and more quantitative LLP information. Lastly, they havefewerloandefaultsafteradoptingCECL.Theseimprovementsaregreaterforbanksthat invest more in CECL-related information systems and human capital and even more salient for larger banks. Our findings suggest that banks’ information production is improved under a more forward-looking accounting standard. However, these improvements are greater for banks with more resources to invest in related technology and human capital. Keywords: CurrentExpectedCreditLosses(CECL);Banks; InformationProduction; Loan Loss Provisioning JEL Classification: E32, G21, G28, M41, M48 ∗We thank Anne Beatty, Matthias Breuer, Ruby Brownen-Trinh (discussant), Alex Casas (discussant), DougDiamond,KurtGee(discussant),YadavGopalan(discussant),Jo˜aoGranja(discussant),AnilKashyap, ElenaLoutskina,YueranMa,JamesNaughton,ShivaRajgopal,JoshuaRonen,HalSchroeder,AbbieSmith, Rimmy Tomy, Cindy Vojtech, James Wang (discussant), Teng Wang, Helen Zhang (discussant), and workshopparticipantsatColumbiaUniversity,theFederalReserveBoard,the2022FederalReserveStressTesting ResearchConference,Wharton,the2022AccountingforanEver-ChangingWorldConference,the2022Fixed Income and Financial Institutions Conference, the 2022 HKUST Accounting Research Symposium, University of Virginia Darden, 2023 HARC, 2023 FARS Mid-year meeting, Ohio State University, University of Chicago Booth Banking workshop, 2023 EAA, and 9th Workshop on Accounting and Regulation for their helpful comments and suggestions. The views expressed in this study are those of the authors and do not necessarily reflect the views of the Federal Reserve Board or the Federal Reserve System. LinkUp data was obtained by Professor Sehwa Kim by agreement with the data provider. Co-author Anya Kleymenova did not have any unauthorized access to this data while working on this paper. We are grateful to Jeffrey Jou, Fanqi Liu, Cong Xiao, and Bohyun Yang for their excellent research assistance. We thank our respective institutions for research funding. All errors are our own. †Columbia University; sk4663@columbia.edu. ‡Baruch College, CUNY; seil.kim@baruch.cuny.edu. §Federal Reserve Board; anya.kleymenova@frb.gov. ¶Columbia University; rl2997@columbia.edu.

1 Introduction In response to the financial crisis of 2007–2009, the Financial Accounting Standards Board (FASB) replaced the incurred loss model (ILM) for estimating credit losses with the current expected credit losses (CECL) model.1 The adoption of the CECL model is considered to be one of the most important accounting standard changes for U.S. banks (ABA, 2016) and is expected to significantly impact banks’ reporting, compliance, and operating decisions. The CECLapproachfundamentallychangesthewaybanksevaluateandprovisionforcreditlosses because they have to provision for all expected credit losses on all outstanding loans over their entire remaining lives, as opposed to only incurred losses under the ILM. Extending the estimationofprovisionstotheremainingloanlivesrequiresbankstogeneratereasonableand supportable forecasts of future economic conditions and factor the impacts of these changing dynamics into their reported loan loss provisions (LLPs). Inthispaper, weexaminewhetherCECLadoptionaffectsbanks’informationproduction and investigate the potential channels through which these effects might arise.2 Prior studies show that banks’ information sets affect their reporting choices and operating decisions.3 Thus, understanding the impact of CECL adoption on banks’ information production processes provides insights into how and why the CECL approach could affect banks’ financial reporting and operational decision-making (e.g., risk management). We hypothesize that CECL-adopting banks would improve their information production because CECL adoptionrequiresincorporatingmoreforward-lookinginformation(e.g., macroeconomicforecasts, borrower-specific information). Since information production is not directly observable, we instead examine how CECL adoption affects banks’ LLP recognition, disclosures, and credit 1Accounting Standards Update (ASU) 2016-13 (ASC 326) was issued on June 16, 2016 (link). The new standard was set to take effect on January 1, 2020 (2023) for large public (small public and private) firms. 2Wedefinetheinformationproductionprocessasbanks’collection,analysis,organization,andreporting of information relevant to their loan portfolios. 3See, for example, Leland and Pyle (1977), Diamond (1984), Qian et al. (2015), Khan and Ozel (2016), Lisowsky et al. (2017), Howes and Weitzner (2023), and Bertomeu et al. (2023) 1

risk management. While banks are expected to exert more effort to collect, analyze, organize, and report information relevant to their loan portfolios under the CECL approach, such effects may not be salient. In particular, CECL adoption may not improve the quality of reporting and operating decisions for the following reasons. First, implementing CECL is costly because forecasting the future is inherently challenging. Moreover, industry experts have commented thatthecostofCECLimplementationishigh, especiallyforsmallerbankswithresourceconstraints (Stein, 2018; McWilliams, 2020). Second, banks often have inefficient or disjointed information systems due to mergers and acquisitions and geographic dispersion of branches. Thus, useful information on borrowers’ credit profiles often resides with loan officers and might not be shared through an internal information system (Stein, 2002; Hertzberg et al., 2010). Finally, the CECL approach grants management more discretion and judgment in estimating LLP than the ILM (Walker, 2019; Kim, 2022). If banks had incentives to exploit the ILM opportunistically, they might exercise even more discretion under the CECL approach, resulting in no improvement to their reporting and operating decisions. Hence, whether CECL adoption improves banks’ information production is an empirical question. WestudytheimpactofCECLadoptionusingU.S.bankholdingcompanies(BHCs)from 2017 to 2021, which includes three years prior to and two years after CECL implementation for large public banks. We employ a difference-in-differences research design and compare a treatment group of large public banks subject to CECL as of January 1, 2020, with a control group of small public banks and private banks not subject to CECL until 2023. To better identify the impact of CECL adoption, we exclude banks that delayed adopting CECL under the Coronavirus Aid, Relief, and Economic Security (CARES) Act exemption.4 We begin our analyses by examining the properties of banks’ LLPs. Among all reporting 4The CARES Act was signed into law in March 2020, allowing banks to delay adoption by the earlier of (1) the termination of the COVID-19 national emergency or (2) January 1, 2022. In our final sample, among public banks subject to CECL as of January 1, 2020, 42 banks elected to delay CECL adoption. As of January 1, 2022, all these banks have adopted CECL, except for two banks that merged with another bank. 2

items, we expect the most salient impact to manifest in LLPs if CECL adoption improves banks’ information production. First, we investigate whether CECL increases the timeliness of banks’ LLPs. The CECL approach requires banks to incorporate forward-looking information when estimating their provisions. Therefore, if banks produce better information about their borrowers, they would quickly react to loan quality deterioration by recognizing LLPs accordingly. Second, we examine whether CECL adopters’ LLPs is more informative about future local economic conditions. Prior studies find that banks’ loan portfolios contain useful information about local economic conditions because they collect detailed and proprietary information concerning their customers’ financial prospects (Khan and Ozel, 2016). Thus, if banks produce better information about their customers and economic conditions, we expect banks’ LLPs to reflect future local economic conditions better after CECL adoption. Consistent with CECL banks producing higher quality information, we find that they record LLPs in a timelier manner, and their LLPs better reflect future local economic conditions. Importantly, theseeffectsarestrongerforheterogeneousloans(commercialrealestate, construction, and commercial and industrial loans), which require more borrower-specific information to monitor than homogeneous loans (residential and consumer loans). Next, we examine whether the impact of CECL adoption is manifested in banks’ disclosures. Prior studies suggest firms’ internal information environments significantly affect their disclosures (Dorantes et al., 2013; Ittner and Michels, 2017; Cheng et al., 2018). Hence, if banksproduce betterinformation fortheir loanportfolios, we expectCECL banksto disclose more informative LLP-related information in their financial reports. Consistent with this prediction, we find that the LLP-related information in CECL adopters’ annual SEC 10-K filings becomes longer and contains more forward-looking and quantitative information. One potential concern regarding our LLP recognition and disclosure analyses is that we cannot distinguish between two mechanisms that could explain our findings. First, banks might already have all the information, and CECL adoption might only change banks’ reporting behavior without affecting the remaining dimensions of their information production 3

(i.e., the collection, analysis, and organization of information). Second, CECL adoption may prompt banks to exert more effort to produce forward-looking information about their customers and economic conditions. The second mechanism is arguably more intriguing as it involves real improvements in banks’ information production activities. While not refuting the existence of the first mechanism, we examine whether the second mechanism plausibly explains our findings by examining banks’ credit risk management. Prior studies suggest that monitoring borrowers is a significant part of banks’ business models (Diamond, 1984; Rajan and Winton, 1995), and banks actively collect borrower informationaspartoftheirmonitoringrole(Gustafsonetal.,2021). Moreinformationabout borrowers also leads to fewer defaults on banks’ loans due to better screening and monitoring (Ertan et al., 2017; Lisowsky et al., 2017). Therefore, we expect fewer borrower defaults among CECL banks after CECL adoption. Importantly, fewer defaults are unlikely to be driven by changes in reporting behavior but can be plausibly explained by banks producing better information. However, a major concern for the default analysis is that borrowerspecific credit risks or loan terms may drive loan default, and these characteristics are mostly unobservable to researchers. We overcome these challenges by controlling for borrowerspecific credit risks and loan-level characteristics using confidential FR Y-14Q regulatory filings. Because only the largest banks report FR Y-14Q filings, for the loan-level default analysis, we use U.S. intermediate holding companies (IHCs) of foreign banks that adopted IFRS 9 in 2018 as the control group.5 We find that CECL-adopting banks experience fewer loan defaults than IHCs after CECL adoption. These results are more salient for private borrowers andriskier loans, consistent with the impact of information production being more pronounced for more opaque and riskier borrowers (Gustafson et al., 2021). A natural follow-up question is through what channel CECL banks improve their information production. Recent studies suggest that financial institutions increasingly invest in 5WecannotuseacontrolsampleofU.S.BHCsthathavenotadoptedCECLbecausenoneofthesebanks report FR Y-14Q. 4

information technology and hire relevant experts to efficiently deal with regulatory compliance (Charoenwong et al., 2023). Also, in several comment letters, practitioners expressed that investment in information technology and human capital would be a necessary condition to implement CECL successfully (e.g., Stein, 2018). Thus, investment in information systems and human capital related to CECL adoption is a plausible channel for improved information production, which provides important policy-relevant implications. We proxy for information systems and human capital investment using banks’ job-postings data, following the approach in the literature (Hershbein and Kahn, 2018; Acemoglu et al., 2022).6 We find that CECL-related job postings mainly contain three job functions: managerial positions related to customer relationships, including collecting and evaluating customer-specific information; quantitative jobs requiring skills related to analyzing and processing the data; and auditing jobs requiring skills related to financial reporting. Thus, CECL-related positions are generally associated with banks’ information production processes of collecting, analyzing, organizing, and reporting information. Consistent with our prediction, we find that CECL adopters posted significantly more jobs related to the CECL approach over the sample period than ILM banks. Lastly, we conduct cross-sectional tests by separating CECL banks based on whether they made large or small investments in CECL-related information systems and human capital. We find that banks with more CECL-related job postings exhibit more significant improvements in their LLP recognition, disclosures, and credit risk management. Notably, these improvements are more salient for larger banks. Overall, our analyses suggest that investment in information systems and human capital is a plausible mechanism through which CECL adoption affects banks’ information production. However, these investments seem to 6Our underlying assumption is that the demand for human capital is closely associated with relatedsysteminvestmentfollowingpriorstudies. Forexample,HershbeinandKahn(2018)documentthatincreased demand for labor skills is linked to IT capital investment. We acknowledge that banks can outsource CECL-relatedfunctions,includinghiringconsultingfirmsandpurchasingcreditmodelstoprepareforCECL adoption. However, banks must also maintain internal systems and have dedicated staff to comply with the CECL approach in their daily operations. Therefore, CECL-related hiring is likely closely associated with CECL-related IT investments. 5

be more concentrated in larger banks, consistent with prior studies suggesting that larger banks have more resources to invest in technology and benefit greater because information creation, collection, and analysis have economies of scale (Wilson, 1975; Begenau et al., 2018; Charoenwong et al., 2023; Farboodi and Veldkamp, 2022). Our study makes several contributions to the literature. First, we provide empirical evidence of the economic consequences of CECL adoption, which is useful to standard setters. Several concurrent studies examine the impact of CECL adoption on lending procyclicality.7 Another stream of studies suggests that LLPs under the CECL model contain some decisionuseful information (e.g., Beatty and Liao, 2021; Wheeler, 2021; Gee et al., 2022). Our paper complements these studies by documenting evidence that CECL adoption incentivizes banks to improve their reporting and operations by producing better information about their borrowers and the underlying economic conditions. In addition, our findings suggest that CECL could provide further insights for the evaluation of credit portfolios. In particular, information gained from CECL can be used to explore loss rates in stress testing or procedures for loan-portfolio bank examinations. Thus, our findings suggest that accounting standards could help bank supervision and regulation. Our study also adds to the literature examining the effects of accounting standards on firms’ information sets.8 Shroff (2017) finds that firms’ investments are affected by GAAP changes, especially by those more likely to alter managers’ information sets. Cheng et al. (2018) find that firms affected by the accounting standard on acquired goodwill and other intangible assets (SFAS 142) provide more accurate management forecasts, consistent with managers acquiring better information while complying with a new accounting rule. Studies examining the adoption of lease accounting standards claim that firms’ investment decisions are affected by the new rule due to the change in the managers’ information set (e.g., Chen 7See,forexample,CohenandEdwards(2017),AbadandSuarez(2018),CovasandNelson(2018),Harris et al. (2018), Loudis and Ranish (2019), Chae et al. (2020), Huber (2022), Chen et al. (2022), and Lu and Nikolaev (2022). 8More broadly, our study is related to literature documenting real effects of disclosure regulation (Leuz and Wysocki, 2016; Roychowdhury et al., 2019). 6

et al., 2023; Roh, 2023). We contribute to this literature by showing an important channel through which the new accounting standard improves the adopting firms’ information environment, namely the investment in information systems and human capital. Also, our findingssuggestthatthestandard-drivenbenefitsarelikelymoresalientforlargeinstitutions with more resources to invest in technology and human capital. 2 Background, Literature, and Hypothesis 2.1 Institutional Background The financial crisis of 2007–2009 sparked a debate about banks’ financial reporting and their loan loss recognition in particular.9 Regulators and others have blamed delays in loan loss provisioning under the existing accounting standard (FAS 5, ILM) for exacerbating the severityofeconomicdownturns. Theyarguethattheits“probable”thresholdforlossaccrual and backward-looking nature induce banks to delay loss recognition in good times, creating anoverhangoflossesthatcarryforwardtobadtimes. Inresponsetothiscriticism, theFASB replaced the ILM of estimating credit losses with the CECL model in Accounting Standards Update (ASU) 2016-13 (ASC 326), effective January 1, 2020 (2023) for large public (small public and private) firms.10,11 The CECL approach mainly addresses the concerns above in two ways (Ryan, 2019). First, it eliminates the ILM’s probable condition. Under the CECL model, a bank recognizes 9See, for example, Laux and Leuz (2010), Barth and Landsman (2010), Vyas (2011), Beatty and Liao (2011), Beatty and Liao (2014), Bushman and Williams (2015), Huizinga and Laeven (2012), Kothari and Lester (2012), Acharya and Ryan (2016), Wheeler (2019), Bischof et al. (2021b), and Kim (2022). 10ASU 2016–13 was initially set to take effect in January 2020 for all SEC filers, except for smaller reporting companies. However, due to the COVID-19 pandemic, the CARES Act provided firms with an optiontodelayCECLadoptionuntiltheearlierof(1)thefirstdateofaneligiblefinancialinstitution’sfiscal year that begins after the date when the COVID-19 national emergency is terminated, or (2) January 1, 2022(asamendedbytheConsolidatedAppropriationsAct). Inaddition, theFASBfurtherpushedbackthe effective date of CECL implementation from January 2021 to January 2023 for smaller reporting companies and from January 2022 to January 2023 for private and nonprofit entities. 11In August 2020, U.S. bank regulators issued the final rule that gave banks an option to mitigate estimated regulatory capital effects of CECL for two years, followed by a three-year transition period, therefore, allowing banks to have a transition period for up to five years. 7

the amount of the expected credit losses on outstanding loans, even for those with a low probability of loss. Second, it substantially weakens the ILM’s conditions regarding when losses are incurred and can be reasonably estimated. Banks are required to incorporate reasonable and supportable forecasts of future economic conditions into their estimates of expected credit losses and recognize credit losses on outstanding loans over their entire remaining lives at inception. In particular, the CECL approach explicitly “[R]equires an entity to consider forward-looking information rather than limiting consideration to current and past events, at the date of the statement of financial position” (FASB, 2016). 2.2 Related Research Prior studies suggest the importance of banks’ information production because it influences their operating and financial reporting choices. Qian et al. (2015) find that better information production by loan officers in Chinese banks improves the forecasting power of interest rates on future outcomes. Khan and Ozel (2016) find that banks’ loan portfolios contain useful information about local economic conditions because banks collect detailed and proprietary information concerning the financial prospects of their customers. Lisowsky et al. (2017) show that banks collected less information from construction firms in the run-up to the financial crisis, which is closely associated with the lower lending standards before the housing crisis. Balakrishnan and Ertan (2021) find that banks’ loan loss provisions become timelier after improved information sharing through public credit registries. Yang (2022) suggests that insufficient loan allowances during the financial crisis are attributable to lowqualityinformationusedforprovisioning. Thesestudiescollectivelyhighlightthecriticalrole of banks’ information production in their operating and reporting choices. Therefore, understanding the impact of CECL adoption on banks’ information production process would help understand how and why CECL might affect banks’ operating and reporting choices. Several concurrent studies examine the impact of CECL adoption on banks’ lending and risk-taking. For example, some examine the effects of CECL on lending procyclicality 8

by employing either actual data under the CECL approach or simulated data under the ILM.12 These studies document mixed findings on the effects of CECL adoption on lending procyclicality, likely due to the different modeling assumptions for the simulated data or the limited data points under the CECL approach. Ballew et al. (2022) study banks’ Paycheck Protection Program (PPP) participation. They find that the intensity of participation is associated with relatively greater changes in risk-taking outside of the PPP, and this effect is concentrated in banks that have not yet adopted CECL. Analytically, Mahieux et al. (2023) investigate how provisioning models interact with bank regulation to affect banks’ risk-taking. They highlight that the expected loss model can spur credit supply and improve financial stability if intervening in banks’ operations is relatively frictionless or regulators can tailor regulatory capital to incorporate information about credit losses. Another related strand of research examines the effects of the adoption of the IFRS 9 expected credit losses (ECL) model in 2018, which occurred two years earlier than CECL adoption. Lopez-Espinosa et al. (2021) document that provisions become more predictive of future bank risk after the ECL adoption. Kim et al. (2021) document that the adoption of ECL improves loan loss recognition timeliness and thus mitigates the procyclicality of bank lending and risk-taking. Ertan (2021) shows that banks that adopted ECL reduce credit supply to small and medium-sized enterprises due to the difficulty in provisioning for more opaque borrowers. Li et al. (2023) find that because ECL banks increase monitoring under IFRS 9, borrowers shift to public debt as opposed to bank debt. Bischof et al. (2021a) find that banks strategically adjust the internal ratings of their borrowers to minimize loan loss provisions. While these studies of IFRS 9 may provide some insights into the expected effects of CECL, their findings may not be replicated under CECL because ECL differs from CECL inseveralways. The mostnotabledifferenceisthatunder ECL,loansareclassifiedintothree stages based on credit quality, and losses are estimated for different horizons depending on 12See,forexample,CohenandEdwards(2017),AbadandSuarez(2018),CovasandNelson(2018),Harris et al. (2018), Loudis and Ranish (2019), Chae et al. (2020), Huber (2022), Chen et al. (2022), and Lu and Nikolaev (2022). 9

the stage, whereas under CECL, losses are estimated over the lifetime of the loan for all loans. In particular, under ECL, for loans classified as stage 1, which includes all new loans, credit losses are estimated over a one-year horizon, resulting in fewer provisions than under CECL (Lopez-Espinosa et al., 2021; Bischof et al., 2021a). Three recent papers are closely related to our study. Beatty and Liao (2021) find analyst provision forecasts incrementally predict future non-performing loans and market returns, suggesting that the incurred loss provision does not incorporate all available future loss information, especially for banks facing greater ILM constraints. The CECL approach, therefore, could remove this constraint and allow banks to incorporate their information into LLPsbetter. Similarly,Wheeler(2021)estimatesexpectedcreditlossesofloansusingvintage analysis and finds that unrecognized expected credit losses under the ILM are negatively associated with bank stock prices. Lastly, Gee et al. (2022) find that the CECL day-1 impact on credit losses improves the value relevance of credit loss allowances and their predictive ability for future credit losses. These studies suggest that LLPs and allowances under expected credit loss models contain some decision-useful information. However, our study differs from prior studies in several ways. First, by examining loan-level default from confidential regulatory filings, our study distinguishes between two explanations why CECL allowances better reflect actual loan losses and local economic conditions: the production of new information versus the unlocking of information already available internally. Second, our study suggests the investment in information systems and human capital as a potential mechanism through which expected credit loss models improve banks’ information production, which should interest standard setters, regulators, and practitioners. 2.3 Hypothesis Development We hypothesize that because CECL adoption requires incorporating more forward-looking information and forecasts of economic conditions, banks would significantly update their 10

information production process by collecting more information, investing more in information technology, and developing better forecasting models. Prior studies also suggest that accounting requirements can improve the overall quality of financial reporting and investing decisions by motivating firms to implement fuller modeling of their risk exposures (Ryan, 2017; Khan et al., 2019). Because information production is not directly observable, we instead test how CECL adoption affects banks’ LLP recognition, LLP-related disclosure, and credit risk management. With better information production, we predict that (i) LLPs become timelier and more reflective of local economic conditions, (ii) LLP-related disclosures are more informative, and (iii) credit risk management benefits from better information. While banks are expected to exert more effort in collecting, analyzing, organizing, and reporting information of their loan portfolios under CECL, such effects may not be salient and thus not improve the quality of reporting and operating decisions for several reasons. First, the implementation of CECL is costly because of the complexity of forecasting the future and integrating the forward-looking information into LLPs. Regulators and industry expertshavecommentedthatthecostofCECLimplementationishigh, especiallyforsmaller banks with resource constraints (Stein, 2018; McWilliams, 2020). Second, banks often have inefficient or disjointed information systems due to mergers and acquisitions and geographic dispersion of branches. Thus, useful information on borrowers’ credit profiles often resides with loan officers and does not end up being reported in an internal information system (Stein, 2002; Hertzberg et al., 2010). Finally, the CECL approach grants management more discretion and judgment in estimating LLP than under the ILM (Walker, 2019; Kim, 2022). If banks had incentives to exploit the ILM opportunistically, they might exercise even more discretion under the CECL approach, resulting in no improvement to their reporting and operatingdecisions. Hence,whetherCECLadoptionimprovesbanks’informationproduction processes and thus improves the quality of reporting and operating decisions is ultimately an empirical question. Given that the direction of our hypothesis is ambiguous, we state our null hypothesis as follows: 11

H : CECL adoption does not increase banks’ information production. 0 3 Data and Sample We use quarterly bank-holding company data, including both public and private banks, with available variables on their FR Y-9C filings from 2017 Q1 to 2021 Q4. This period includes three years before large public banks adopted CECL and two years afterward. We require banks to have non-missing assets, deposits, changes in non-performing loans, lagged ratio of capital to assets, and earnings before loan loss provisions and taxes. We also require banks to have at least one-quarter of observations for both pre- and post-CECL adoption periods. After implementing these data requirements, we have 357 unique banks in the sample. To clearly identify the effects of CECL adoption, we exclude 20 foreign banks with headquarters outside of the U.S. because these banks were already subject to IFRS 9 starting in 2018.13 We also exclude 53 banks with delayed adoption or adoption in different calendar quarters.14 We determine whether banks adopt or delay CECL adoption by reading their 10-K filings and cross-checking with the information available in their FR Y-9C reports.15 Banks that adopted CECL in January 2020 are defined as our treatment group, and banks that did not adopt CECL by December 2021 are our control group. The final sample consists of 5,488 bank-quarter observations representing 284 unique banks (150 CECL and 134 ILM banks). 13In loan-level analyses, we use some of these foreign banks as a control group and compare them to the U.S. CECL-adopting banks. 14In our final sample, among public banks subject to CECL as of January 1, 2020, 42 banks elected to delay CECL adoption. Among them, 15 banks adopted CECL in 2020 Q4, 18 banks adopted CECL in 2021 Q1, and seven banks adopted CECL in 2022 Q1. In Table OA.1, we examine the determinants of banks delaying CECL adoption as of 2019 Q4. Bank size is an important factor in predicting a bank’s decision on delaying CECL adoption when the CARES Act was announced. In addition, to proxy for banks’ readiness for CECL adoption, we manually collected whether a bank provided any CECL impact estimation (either range or point estimates) in their 2019 10-K, an immediate quarter before the scheduled CECL adoption. We find that whether a bank provided a CECL impact estimate is another predictor. This finding suggests that small banks not fully prepared for CECL chose to delay its adoption when given the option. In sum, our analyses suggest that significant selection bias exists for delay banks. Thus, we do not include these delay banks in our sample. 15Items BHCKJJ20-BHCKJJ28 and BHCAJJ29 are reported only by banks that adopted CECL. We use this information to determine whether and when private banks adopt CECL. No private banks adopted CECL in January 2020; hence, none are included in our treatment group. 12

For the loan-level analysis, we use FR Y-14Q regulatory filings that are collected quarterly as part of the Federal Reserve’s Dodd-Frank Act Stress Tests (DFAST) and Comprehensive Capital Analysis and Review (CCAR) for bank holding companies (BHCs), savings and loan holding companies (SLHCs), and U.S. intermediate holding companies (IHCs) of foreign bank organizations with at least $50 billion ($100 billion starting from 2019) in total assets.16 The banks that have submitted FR Y-14Q data since 2012 comprise over 85 percent of the U.S. banking sector’s total assets. The FR Y-14Q data include commercial and industrial (C&I) loans with a committed balance greater than or equal to $1 million (Caglio et al., 2022). We focus our analyses on Schedule H.1, which contains detailed information on banks’ loans to C&I borrowers. FR Y-14Q reporting banks that adopted CECL in 2020 are defined as our treatment group, and IHCs of foreign banks that adopted IFRS 9 in 2018 are our control group. The sample consists of 26 banks that adopted CECL and eight IHCs of foreign banks that adopted IFRS 9. To proxy for the investment in information systems and human capital related to the adoption of the CECL methodology, we use job posting data provided by LinkUp. The data track the daily creation and deletion dates of online job postings by U.S. firms on their websites. The LinkUp data cover 127 out of 150 CECL banks in our sample. Table1presentsthedescriptivestatisticsforoursample. PanelAprovidesthedescriptive statistics of bank-level characteristics. The mean of LLPs is 0.081 percent of beginning-ofquarter total loans. The mean of LLPs for homogeneous (heterogeneous) loans, estimated as the change in allowance plus charge-offs, is 0.040 (0.044) percent of beginning-of-quarter total loans. We define LLPs for homogeneous or heterogeneous loans as missing if a bank is under the asset threshold to report allowance by loan type or does not hold certain types of loans. Columns (9) through (14) compare the mean values of these variables for CECL and ILM banks. The mean of LLPs is higher for CECL banks. Our control variables, Size, EBLLP, Deposit, and CapRatio, significantly differ between the two groups. We include 16Our findings using confidential supervisory FR Y-14Q data have been approved for public release. 13

bank fixed effects in all our regressions to control for unobserved differences, such as the business model differences between CECL and ILM banks. In addition, in Figure 1 and Figure 2, we check for parallel trends for LLP recognition and disclosures by CECL and ILM banks before CECL adoption. We do not see evidence that provisions and disclosures of CECL and ILM banks differed before the CECL adoption. Panel B of Table 1 presents descriptive statistics of the additional loan- or borrower-level characteristics for our loan-level analyses. Similar to our discussion above, we compare U.S. CECL banks to a control group of IHCs, i.e., foreign banks that have adopted IFRS 9 by 2018. On average, U.S. CECL banks have larger and less levered borrowers and are less likely to have loans with collateral or guarantees. They are also, on average, more likely to issue loans to new borrowers and are less likely to lend to private borrowers. We check that both types of banks follow parallel trends for default rates and find that they are not significantly different prior to the implementation of CECL.17 4 Empirical Approach and Results 4.1 Information in Loan Loss Provisioning (LLP) Webeginouranalysesbyexaminingthepropertiesofbanks’LLPs, whereweexpectthemost salient changes if banks produce higher-quality information after CECL adoption. First, we examine whether the CECL approach increases the timeliness of banks’ LLPs. The CECL approach requires banks to recognize expected credit losses by incorporating forward-looking information. If banks produce better information about their customers and economic conditions, they would quickly react to loan quality deterioration by recognizing timelier LLPs. Prior studies proxy the timeliness of LLPs as a positive relationship between current LLPs 17Wereporttime-varyingloanmaturitiesinyears. Termloanstendtohavelongermaturitiesonaverage. We include loan-type fixed effects in our empirical specification to account for some of the unobserved heterogeneity that might be due to loan type (loan types consist of different types of term loans, including bridgeandasset-basedloansasreportedinFRY-14Q,weexcludecreditlinesinouranalyses). Ourfindings are also robust to using the natural logarithm of loan maturity instead. 14

and changes in future non-performing loans (Nichols et al., 2009; Beatty and Liao, 2011; Bushman and Williams, 2015; Kim, 2022). Thus, we expect the positive relationship between current LLPs and changes in future non-performing loans for the adopting banks to become stronger after CECL adoption. Also, we expect a greater impact for heterogeneous loans (commercial real estate, construction, and commercial and industrial loans) than homogeneous loans (residential and consumer loans). Banks primarily evaluate credit losses for homogeneous loans at the portfolio level and typically record LLPs as expected loan charge-offs over the next 12 months. Depending on the type of homogeneous loans, 12 months can be similar to (e.g., credit card loans) or somewhat less than (e.g., auto loans and residential mortgages) the remaining lifetime of the loan (Ryan, 2019). On the other hand, banks primarily evaluate credit losses for heterogeneous loans on a loan-by-loan basis, which requires more borrower-specific information to monitor and thus more effort to collect (Liu and Ryan, 2006; Bhat et al., 2021). Therefore, CECL likely affects heterogeneous loans more than homogeneous loans. We first examine the effects of CECL adoption on banks’ LLPs with a simple graphical analysis. In Panel A of Figure 1, we plot the average proportion of LLPs to beginning total loans for CECL and ILM banks at the quarterly frequency from 2017 Q1 to 2021 Q4. Up to 2019 Q4, both CECL and ILM banks recorded similar proportions of LLPs to loans. Notably, both groups’ LLPs show clear parallel trends until 2019 Q4. However, CECL banks increased LLPs significantly in 2020 Q1. This immediate jump is composed of the day-1 CECLadoptionimpact, estimatedasofJanuary1, 2020, andadditionalupwardadjustments during 2020 Q1, which reflect deteriorating economic conditions caused by the COVID-19 outbreak. However, CECL banks’ LLPs decreased significantly from 2020 Q2 until 2021 Q2, during which immediate government responses to mitigate the economic impact of the COVID-19 pandemic came into effect. By contrast, ILM banks show a gradual increase in LLPs from 2020 Q1 until 2020 Q2 and then a gradual decrease, consistent with these banks provisioning for losses in a less timely manner than CECL banks. 15

In Panel B and Panel C, we plot the LLP trends for homogeneous and heterogeneous loans, respectively.18 The general trends of LLP recognition for homogeneous loans are similar for both CECL and ILM banks except for the adoption quarter. By contrast, we see larger LLP recognition by CECL banks than ILM banks earlier in the COVID-19 pandemic period, followed by smaller LLP recognition by CECL banks afterward. These patterns are consistent with our prediction that the impact of CECL adoption on the timeliness of LLPs is likely larger for heterogeneous loans than homogeneous loans. Next, we formally test this hypothesis using the following model: LLP = β Treat ×Post ×∆NPL +β Treat ×Post ×∆NPL i,t 1 i t i,t+ 2 i t i,t +β Treat ×Post ×∆NPL +β Treat ×∆NPL +β Treat ×∆NPL 3 i t i,t− 4 i i,t+ 5 i i,t +β Treat ×∆NPL +β Post ×∆NPL 6 i i,t− 7 t i,t+ +β Post ×∆NPL +β Post ×∆NPL +β Treat ×Post 8 t i,t 9 t i,t− 10 i t +β ∆NPL +β ∆NPL +β ∆NPL +β X +δ +γ +ϵ , 11 i,t+ 12 i,t 13 i,t− 14 i,t t i i,t (1) where i and t index bank and year-quarter, respectively. The dependent variable, LLP , i,t is the bank’s LLPs divided by lagged total loans. We also consider three variants of the dependent variable. LLP (w/ Day 1) adds the day-1 impact that bypasses the income statement.19 LLP - Homog. and LLP - Hetero. are calculated as the quarterly change in allowance plus net charge-offs for homogeneous and heterogeneous loans. Thus, these variables contain the day-1 CECL impact as well as other adjustments to the allowance for loan losses, such as the expected credit losses on purchased credit deteriorated assets. The explanatory variable of interest is Treat ×Post ×∆NPL . Treat is an indicator i t i,t+ i 18Banks do not report LLPs by loan type in the FR Y-9C. We estimate LLPs by loan type as the change inallowanceplusnetcharge-offs. Asaresult,wecannotseparatetheday-1CECLadoptionimpactonLLPs by loan type from additional upward adjustments during 2020 Q1. Therefore, the day-1 CECL adoption impact is included in LLPs by loan type. 19We obtain the day-1 impact of CECL adoption on loan loss provisions from item BHCKJJ28 in the FR Y-9C and when it is missing from 10-Q filings. 16

that equals one if a bank adopted the CECL standard in 2020 Q1. Post is an indicator t variable that equals one for quarters after 2020. ∆NPL is the average future loan quality i,t+ change over the next two quarters, which is measured as the change in non-performing loans divided by lagged total loans. ∆NPL is the current loan quality change. ∆NPL is the i,t i,t− averagepastloanqualitychangeoverthepasttwoquarters. Thecalculationof∆NPLvaries with the choice of the dependent variable. We control for several bank characteristics, X , i,t including Size , the natural logarithm of total assets, EBLLP , the earnings before loan i,t i,t loss provisions and taxes divided by lagged loans, Deposit , total deposits divided by total i,t assets, and CapRatio , lagged ratio of capital to total assets. We include year-quarter i,t−1 fixed effects, δ , to control for economic conditions affecting all banks in each sample quarter t and bank fixed effects, γ , to account for time-invariant bank characteristics. i Table 2 reports the estimation of Equation 1. In column (1), we examine the effects of CECL adoption on LLPs of total loans without the day-1 CECL impact (i.e., provisions recognized in the income statement in each quarter). The coefficient on Treat × Post × i t ∆NPL is significantly positive (0.320, p <0.05), suggesting that LLPs of CECL banks i,t+ better reflect changes in future non-performing loans than that of ILM banks after CECL adoption. ThefindingisconsistentwithourhypothesisthatCECLbanksrecognizeexpected credit losses in a timelier manner by incorporating forward-looking information. In column (2), we examine the effects of CECL adoption on LLPs of total loans by incorporating the day-1 CECL impact and find consistent and even stronger results. The coefficient on Treat ×Post ×∆NPL is significantly positive (0.512, p <0.01), suggesting that LLPs i t i,t+ under the CECL approach, with or without the day-1 impact, contain useful information for current and future loan quality deterioration. In columns (3) and (4), we separately examine the effects of CECL adoption on LLPs of homogeneous and heterogeneous loans.20 We find that the coefficient on Treat × Post × ∆NPL is statistically insignificant for i t i,t+ 20We have fewer observations for the tests using LLPs of homogeneous and heterogeneous loans because CECL banks with assets under $5 billion are only required to report allowances by loan type semiannually after 2020. 17

homogeneous loans (-0.143, p >0.10) but is significantly positive for heterogeneous loans (0.521, p <0.01). These results indicate that the effects of CECL adoption on the timeliness of LLP recognition are mostly driven by heterogeneous loans, which is consistent with our prediction that the improvement in information production would be more substantial for loans requiring more borrower-specific information.21 Next, we examine whether CECL banks’ LLPs contain more information about local economic conditions in the states where they operate. Khan and Ozel (2016) find that banks’ loan portfolios contain useful information about local economic conditions because banks collect detailed and proprietary information about the financial prospects of their customers. Ifbanks’LLPsreflectchangesinlocaleconomicconditionsbetterduetoimproved information quality, we expect the negative relationship between current LLPs and changes in future local economic indicators to become stronger after CECL adoption. We proxy local economic conditions using the coincident index, a comprehensive measure of economic activity at the state level.22 We formally test this hypothesis using the following model: LLP =β Treat ×Post ×∆CoIndex +β Treat ×Post ×∆CoIndex i,t 1 i t s,t+ 2 i t s,t +β Treat ×Post ×∆CoIndex +β Treat ×∆CoIndex 3 i t s,t− 4 i s,t+ +β Treat ×∆CoIndex +β Treat ×∆CoIndex 5 i s,t 6 i s,t− +β Post ×∆CoIndex +β Post ×∆CoIndex +β Post ×∆CoIndex 7 t s,t+ 8 t s,t 9 t s,t− +β Treat ×Post +β ∆CoIndex +β ∆CoIndex +β ∆CoIndex 10 i t 11 s,t+ 12 s,t 13 s,t− +β X +δ +γ +ϵ , 14 i,t t i i,t (2) where i, t, and s index bank, year-quarter, and state, respectively. Same as before, the 21Inuntabulatedanalysis,wealsocomparebankswithlowandhighproportionsofheterogeneousloansin theirloanportfoliosfollowingotherstudies(e.g.,Chenetal.,2022). ConsistentwithourfindingsinTable2, we find a stronger CECL adoption impact for banks with high proportions of heterogeneous loans. 22The index is produced monthly by the Federal Reserve Bank of Philadelphia and calculated using modelswithfourstate-levelinputs: nonfarmpayrollemployment,unemploymentrate,averagehoursworked inmanufacturing,andwageandsalarydisbursementsdeflatedbytheconsumerpriceindex(KhanandOzel, 2016). 18

dependent variable is LLP and its three variants. The explanatory variable of interest is i,t Treat ×Post ×∆CoIndex . CoIndex is the average future local economic condition i t s,t+ i,t+ changes over the next two quarters, which is measured as the weighted average of the coincident index based on banks’ deposit shares in different states. CoIndex is the current s,t local economic condition change. CoIndex is the average past local economic condition i,t− changes over the past two quarters. The same set of bank characteristics, as in Equation 1, is included as control variables. We also control for ∆NPL , the changes in non-performing i,t loans divided by lagged total loans. Finally, year-quarter fixed effects, δ , and bank fixed t effects, γ , are included. i Table 3 reports the estimation of Equation 2. In column (1), we examine the effects of CECL adoption on LLPs of total loans without the day-1 CECL impact. The coefficient on Post ×∆CoIndex is significantly positive (0.035, p <0.01), suggesting banks recognize t s,t+ more provisions when future local economic conditions are better during the post period. This finding suggests that banks generally experienced difficulties incorporating future local economic conditions in their LLPs during the post period, which is likely driven by the increased uncertainty due to the pandemic.23 However, the coefficient on Treat ×Post × i t ∆CoIndex is significantly negative (-0.035, p <0.01), suggesting the positive relationship s,t+ between banks’ LLPs and future local economic conditions during the post period is almost canceled out for CECL banks; presumably, they have better capability to forecast economic 23We caveat, however, that the coefficient of Post ×∆CoIndex is sensitive to the inclusion of yeart s,t+ quarter fixed effects, likely due to a small cross-sectional variation in ∆CoIndex. Without the inclusion of year-quarter fixed effects, the coefficient of Post × ∆CoIndex is around -0.007, suggesting both t s,t+ CECL and ILM banks better incorporate future local economic conditions in their LLPs during the post period. To investigate this issue further, we utilize diagnostic tests to determine whether granular fixed effects are causing this change in the sign of the coefficient (Armstrong et al., 2022). The variation in Post × ∆CoIndex is 78.5% absorbed by year-quarter fixed effects. The impact is smaller for our t s,t+ variableofinterest, Treat ×Post ×∆CoIndex ,ofwhich43.2%ofvariationisabsorbedbyyear-quarter i t s,t+ fixed effects. The Variance Inflation Factor (VIF) statistics, which reflect the level of multicollinearity, also increase from 6.29 to 14.85 for Post ×∆CoIndex when including year-quarter fixed effects, whereas the t s,t+ VIF for Treat ×Post ×∆CoIndex is almost unchanged from 6.54 to 6.55. Thus, the interpretation i t s,t+ on Post ×∆CoIndex should be carefully made with this caveat in mind. Importantly, the coefficient t s,t+ of Treat ×Post ×∆CoIndex is generally stable across different model specifications, assuring that the i t s,t+ estimation of our variable of interest is robust. 19

conditions based on better information despite the increased uncertainty. In column (2), we examine the effects of CECL adoption on LLPs of total loans by incorporating the day-1 CECL impact and find consistent results (-0.065, p <0.01). Again, these results suggest that both day-1 and subsequent LLPs of CECL banks contain useful information for current and future local economic conditions. We also further examine the effects of CECL adoption on LLPs of homogeneous and heterogeneous loans. In columns (3) and (4), we find that the coefficient on Treat × Post × ∆CoIndex is weakly significantly negative (-0.017, i t s,t+ p <0.10)forhomogeneousloansandsignificantlynegative(-0.029,p <0.01)forheterogeneous loans. These findings indicate that the effects of CECL adoption on information production regarding local economic conditions are slightly stronger for heterogeneous loans.24 However, the difference is not as salient as the results on the timeliness of LLPs.25 4.2 Information in Disclosures Prior studies suggest that firms provide more frequent and accurate disclosure when their internal information environments improve (Dorantes et al., 2013; Ittner and Michels, 2017; Cheng et al., 2018). Thus, if banks produce better information on their loan portfolios after CECL adoption, we expect CECL banks to provide more informative disclosures related to LLPintheirfinancialreports. Specifically,wetestwhetherCECLbanksprovidelonger,more forward-looking, and more quantitative information related to LLP in their 10-K filings.26 We use the number of sentences discussing LLPs to proxy the quantity of LLP-related information.27 However, an increase in the quantity of LLP-related disclosure does not 24Inuntabulatedanalysis,wealsocomparebankswithlowandhighproportionsofheterogeneousloansin their loan portfolios. Again, we find a stronger impact of CECL adoption for banks with higher proportions of heterogeneous loans. 25The less salient difference is likely because macroeconomic indicators, which are correlated with local economic conditions, are important inputs to determine LLPs for both homogeneous and heterogeneous loans. 26We focus on textual information in form 10-K rather than management guidance because the latter type of disclosure is rare in the banking industry. 27To identify LLP-related disclosures in banks’ 10-Ks, we first normalize raw filings to address issues of punctuation, inflections, and extra white spaces. Then, we search for sentences that contain LLP-related wordssuchas“provision,”“allowance,”“default,”“chargeoff,”“creditloss,”and“loanloss.” Next,wetake 20

necessarily suggest an improvement in the informativeness of such disclosure. For example, the added paragraphs could be boilerplate describing the new standard, such as how LLPs underCECLarecalculated. TofurtherexaminewhetherLLP-relateddisclosurecarrieshighquality information, we search for sentences that contain a forward-looking word (e.g., Muslu etal.,2015;Bozanicetal.,2018)orahardnumber(e.g.,Dyeretal.,2017;Blankespoor,2019) among those LLP-related sentences.28 Sentences with forward-looking words are likely to be discussions about banks’ evaluations of the macroeconomic environment and projections of indicators related to LLP calculation. Sentences with hard numbers provide quantitative information that is more specific and easier to notice, process, and compare. Appendix C provides snapshots of JP Morgan Chase’s LLP-related disclosures in its 10-K filings before and after CECL adoption. The first observation is that LLP-related discussions become longer after CECL adoption. Highlighted texts in the 2020 10-K are incrementalLLP-relateddisclosuresweintendtocaptureusingtheprocedureoutlinedabove.29 Notably, these sentences either contain forward-looking words such as “assumption,” “outlook,” “forecast,” and “scenario,” or specific macroeconomic forecasts of the unemployment rate and GDP growth (in numeric forms). This example illustrates the relevance and informativeness of LLP sentences that are forward-looking and quantitative. Panel A of Figure 2 plots the number of sentences in banks’ 10-K filings that are LLPrelated from 2017 to 2021. Panel B and Panel C of Figure 2 further plot the number of LLP sentences that contain forward-looking words and hard numbers, respectively. As LLPs in Figure 1, both groups’ LLP-related disclosures show clear parallel trends prior to CECL the union of all sentences located within the (−3,+3) window of the direct LLP-related sentences identified in the previous step to count the unique number of sentences. 28Westartbypre-specifyingalistofwordsdeemedforward-looking. Thelistcontainsthestemmedforms of the following words: “anticipate,” “believe,” “estimate,” “expect,” “forecast,” “predict,” and “target.” Next,weexpandthelistusingwordembedding. Thenaturallanguageprocessing(NLP)techniqueidentifies words that are likely to appear in the same contexts as the target words. We conduct word embedding using a large corpus of banks’ 10-K filings. The expanded list additionally includes stemmed forms of “aim,” “assumption,” “baseline,” “future,” “judgment,” “outlook,” “probably/probability,” “scenario,” and “(un)predictable.” 29The table, which provides similar information, is not highlighted because contents within HTML <table> tags are removed when processing 10-K documents. 21

adoption. However, consistent with our prediction, the average quantity and quality of LLPrelated disclosures increase for CECL banks compared to ILM banks after CECL adoption. We formally test this hypothesis using the following model: LLP Disc =β Treat ×Post +β X +δ +γ +ϵ , (3) i,t 1 i t 2 i,t t i i,t where i and t index bank and year, respectively. The dependent variable, LLP Disc , takes i,t three forms: LLP Disc is the natural logarithms of one plus the number of unique sentences falling within the (−3,+3) window of any 10-K sentence in which there is an LLP sentence; LLP Disc - Fwd. is the natural logarithms of one plus the number of sentences containing forward-looking words among such LLP-related sentences; and LLP Disc - Quant. is the natural logarithms of one plus the number of sentences containing quantitative information (i.e., hard numbers) among such LLP-related sentences. The same set of bank characteristics and fixed effects, as in Equation 2, are included. Table 4 reports the estimation of Equation 3. In columns (1) through (3), we find that the coefficient on Treat ×Post is significantly positive in all columns (0.124, p <0.01; 0.201, i t p <0.01; 0.085, p <0.01). The results suggest that managers at CECL banks provide longer, more forward-looking, and quantitative information than those at ILM banks after CECL adoption. Thesefindings suggestLLP-related disclosures are improved forCECL banks both quantitatively and qualitatively, consistent with the prior studies showing the quantity and quality of disclosures improve when firms’ internal information environments improve. 4.3 Do CECL Banks Produce Better Information? In the previous section, we show that CECL banks’ LLPs reflect future credit losses and local economic conditions better than those of ILM banks. One concern is that two different mechanisms could explain our findings. First, banks might already have all the information even before CECL adoption, and CECL adoption only affects banks’ reporting behavior be- 22

cause it eliminates restrictions on recognizing LLPs under the ILM. Second, CECL adoption prompts banks to value the forward-looking estimation task more and thus exert more effort to produce more information about their loan portfolios. While these two mechanisms can coexist, we examine whether the second mechanism plausibly explains our findings by investigating loan-level default, observable in the confidential FR Y-14Q regulatory filings. Prior studies suggest that monitoring borrowers is a major function of banks (Diamond, 1984; Rajan and Winton, 1995) and banks actively collect borrower information as part of their monitoring role (Gustafson et al., 2021). Research also suggests that more information about borrowers leads to fewer loan defaults due to better screening and monitoring (Ertan et al., 2017; Lisowsky et al., 2017). If banks screen and monitor loans better using improved information, we expect borrowers of CECL banks to exhibit fewer defaults following CECL. Furthermore, fewer defaults are unlikely to be driven by changes in reporting behavior but canbeplausiblyexplainedbybanksproducingbetterinformation. Also,examiningloan-level default instead of bank-level NPLs or charge-offs allows us to control for borrower-specific credit risks and loan terms and explore cross-sectional differences across loan characteristics. We examine the impact of CECL adoption on loan-level default, comparing large U.S. BHCs that adopted CECL in 2020 to foreign banks’ U.S. IHCs that adopted ECL under IFRS 9 in 2018. The underlying assumption is that because these foreign banks have already adopted the ECL approach, an accounting standard similar to the CECL approach, earlier than the U.S. CECL banks, they can serve as a control group. To avoid any confounding effects of IFRS 9 adoption on foreign banks, we limit our sample to 2018–2021 for this analysis. We formally test this hypothesis using the following model: Default =β Treat ×Post +X +Y +Z +δ +γ +κ +ϵ , (4) i,j,k,t 1 i t i,t j,t k,t t i k i,j,k,t where i, j, k, and t index bank, borrower, loan, and quarter, respectively. The dependent variable is Default , an indicator that equals one if a loan defaults (i.e., becomes 90 days i,j,k,t 23

past due) within four quarters of the reporting quarter.30 The same set of bank characteristics, as in Equation 2, is included. We also control for borrower characteristics, including borrower size, borrower leverage, and an indicator of whether the borrower is a private firm. We also control for loan characteristics, including the probability of default (PD) assigned by the bank, loan maturity, and indicators for whether a loan includes collateral, is syndicated, or is guaranteed. We also identify borrowers who do not have a prior lending relationship with the specific lender in our sample.31 Finally, we include year-quarter, δ , bank, γ , and t i loan-type fixed effects, κ .32 k Table 5 reports the estimation of Equation 4. In column (1), we find that the coefficient of Treat × Post is significantly negative (-0.492, p <0.01), consistent with CECL banks’ i t borrowers experiencing lower defaults. To mitigate any concern that our results are driven by treatment banks having more PPP loans than our control banks, we exclude all loans with government guarantees, including PPP loans.33 In columns (2) and (3), we divide the sample into borrowers with prior lending relationships and new borrowers to address the concern that our results are driven by banks shifting to ex-ante less risky borrowers after CECL adoption. While the decrease in default probability is greater in column (3) for new borrowers, we still find a significant decrease in default probability for existing borrowers in column (2), mitigating the concern that the decrease in loan-level default for CECL-adopting banks is solely driven by selection. In columns (4) and (5), we divide the sample into private and public borrowers. We find that the decrease in default is only significant for private borrowers (-0.529, p <0.01), consistent with a greater incremental impact of information production for more opaque borrowers. Lastly, in columns (6) and (7), we divide the sample 30Our results are robust to defining loan defaults if a loan is 30 days past due within four quarters of the reporting quarter. 31We exclude credit lines as they are rolled over from year to year and can change terms as well as loans to individuals and municipalities. 32FR Y-14Q, Schedule H1 reports different types of commercial loans. Loan-type fixed effects capture different types of term loans, such as term loans A, B, and C, bridge loans, and asset-based loans. 33Our results remain consistent if we instead compare within bank changes of pre- and post-CECL adoption periods for large U.S. BHCs that file FR Y-14Q. 24

into loans with a low and high assigned PD (defined as below or above the median). We find that the decrease in default is stronger for loans with higher PD (-0.574, p <0.01), consistent with a greater incremental impact of information production for riskier loans. 4.4 Potential Mechanism A natural follow-up question is through which channels CECL banks improve their information production. Recent studies suggest that financial institutions are increasingly investing in information technology and hiring experts to efficiently deal with regulatory monitoring, reporting, and compliance (Charoenwong et al., 2023). Similarly, Bhat et al. (2019) suggest that credit risk modeling significantly improves banks’ information about their credit losses. Arif et al. (2022) find that the quality of banks’ human capital is associated with better loan monitoring and timelier loan loss provisioning. Thus, we conjecture that the investment in information systems and human capital related to CECL adoption is a plausible channel for improved information production. We proxy for information system and human capital investment using job-postings data following the approach in the literature (Hershbein and Kahn, 2018; Acemoglu et al., 2022).34 Specifically, we search terms, including “CECL,” “CurrentExpectedCreditLosses,” “ASU2016-13,” “ASC326,” “Topic326,” and“Financial Instrument(s)CreditLoss(es)”injobdescriptions, andlabelajobpostingasaCECL-related job if it contains one of these terms.35 Figure 3 presents the number of CECL-related job postings.36 In Panel A, consistent 34We also considered alternative sources. First, we collect 10-K disclosures regarding banks’ reliance on external CECL solutions such as Moody’s Analytics. While we cannot ensure that all banks fully disclose the relevant information, we find that our results are consistent for banks that use or do not use external consultants, suggesting that the main findings of our paper are not affected by whether banks internally develop their CECL-related information system or rely on third parties. Second, we considered using FR Y- 9C information regarding accounting and auditing expenses or consulting and advisory expenses as a proxy forbanks’CECL-relatedinvestments. However, theseexpensesarereportedonlyiftheyexceedasignificant portion of other noninterest income and are missing for a large number of banks in our sample. 35Beforesearchingforpatterns,wenormalizerawjobpostingstoaddressissuesofpunctuation,inflections, and extra white spaces. 36InFigureOA.8,wechecktherepresentativenessofLinkUpdatabycomparingthemwiththejobopening data by the U.S. Bureau of Labor Statistics (BLS). The LinkUp data has fewer job postings than the BLS data because LinkUp only covers companies that list jobs on their own websites. However, the trends in the 25

with our prediction, CECL banks started posting CECL-related jobs a few years before 2020 (the adoption year), suggesting that these banks had prepared to comply with the CECL a while before the adoption. Notably, we observe a decrease in the number of CECL job posting around the outbreak of COVID-19 in early 2020. However, the number of job postings surged in 2021, suggesting that banks are increasingly investing in human capital to update and maintain the information system even after initial CECL adoption.37 To understand the characteristics of CECL-related jobs, in Appendix B, we provide summarystatisticsofthesejobpostings. InPanelA, welistthetop10CECLjobemployers. Not surprisingly, large national banks, including Wells Fargo, Bank of America, and JPMorgan Chase, comprise a significant portion of CECL-related job postings, suggesting that larger banks have better resources for investment in information technology and related-human capital.38 Also, smaller banks have argued, and regulators have acknowledged that CECL adoption is more burdensome for them.39 In Panel B, we list the top 10 CECL job titles. Most CECL job titles contain words, including Analytic, Credit Risk, and Quantitative, which are highly associated with information production. Figure 4 presents word clouds of frequently used words in CECL job titles and descriptions. The word clouds also highlight words, including analyst, credit, model, and risk, related to information production, which provides assurance that CECL-related job postings are a suitable proxy for information systems and human capital investment. In Panel C, we categorize these jobs based on the O*NET Standard Occupational Classinumber of job postings are similar in both databases, assuring that the LinkUp data well reflects the labor market demand. 37One potential concern is that observing few CECL-related job postings for ILM banks seems trivial, as thesebanksarenotsubjecttoCECLuntil2023. Toprovideanalternativebenchmark,inPanelBandPanel C of Figure 3, we define an informational job if a job shares any O*NET SOC codes with CECL-related jobs for ILM banks. Also, to mitigate bank size effects, we normalize job postings with the number of job postingsin2017Q1. WefindthatthepatternofinformationaljobpostingsbyILMbanksisrelativelystable compared to the increasing number of CECL-related job postings by CECL banks. 38We caveat that, among the top 4 commercial banks in the U.S., Citibank is not covered by the LinkUp database. However, we conjecture that Citibank has made extensive investments in CECL-related information systems and human capital. 39Forthatreason,smallerbanksarealsomorelikelytooutsourcetoconsultantsorutilizemodelsdeveloped by other banks. 26

fication (SOC).40 The SOC-based job titles and key tasks suggest that CECL jobs are mainly associated with three functions. First is managerial jobs related to customer relationships and thus likely to gather more information about them (e.g., Financial Managers). Second is quantitative jobs related to analyzing and processing the data (e.g., Financial and Investment Analysts and Credit Analysts). The last is auditing jobs related to financial reporting (e.g., Accountants and Auditors). Thus, CECL jobs generally relate to banks’ information production process of collecting, analyzing, organizing, and reporting information. To formally test the investment in information systems and human capital as a plausible mechanism, we conduct several cross-sectional tests by separating CECL banks that made large investments into CECL-related technology and human capital based on the median value of the cumulative number of CECL-related job postings from 2017 to a given yearquarter (i.e., Low- versus High-CECL Jobs). We caveat that our proxy for investment in information systems and human capital cannot be fully distinguishable from a bank size effect. However, prior research suggests greater benefits of information-related investments for larger firms because technological investments have a large fixed component and information tends to have economies of scale (Wilson, 1975; Begenau et al., 2018; Charoenwong et al., 2023; Farboodi and Veldkamp, 2022).41 Table 6 reports the estimation of Equation 1 by comparing CECL banks with low- and high-CECL job postings to ILM banks. In columns (1) through (3), we examine the effects of CECL adoption for LLPs of total loans without the day-1 CECL impact for low-CECL job CECL banks, high-CECL job CECL banks, and high-CECL jobs and large CECL banks, respectively. We find that the coefficient on Treat ×Post ×∆NPL is at least weakly i t i,t+ significant for all three columns. Notably, the magnitude of the coefficient is larger for 40TheO*NETSOCisafederalstandardusedtoclassifyoccupationsintoapproximately1,000categories. These occupations have associated data with occupational characteristics, including knowledge, skills, abilities, tasks, and general work activities. See (link.) 41We also separate Low- and High-CECL jobs based on the number of CECL-related job postings scaled by the average number of bank employees or average assets, to remove the bank size effect, although this approachdisproportionatelypenalizeslargerbanks. Wefindconsistentbutweakerdifferencesbetweenbanks with Low- and High-CECL jobs if we use the scaled number of CECL-related job postings. 27

high-CECL job CECL banks and the largest for large CECL banks with high-CECL jobs. In columns (4) through (6), we examine the effects of CECL adoption for LLPs of total loans with the day-1 CECL impact and find a similar pattern. These findings are consistent with our prediction that the CECL impacts are larger for banks with greater investment in information systems and human capital related to CECL adoption, and these effects are even more salient for larger banks.42 Table 7 reports the estimation of Equation 2 by comparing CECL banks with low- and high-CECL job postings to ILM banks. In columns (1) through (3), we examine the effects of CECL adoption on LLPs of total loans without the day-1 CECL impact. Similar to Table 6, we generally find that the magnitude of the coefficient on Treat ×Post ×∆NPL is larger i t i,t+ forhigh-CECLjobandlargeCECLbanks. Incolumns(4)through(6), wefindsimilarresults for the effects of CECL adoption for LLPs of total loans with the day-1 CECL impact.43 Table 8 reports the estimation of Equation 3 by comparing CECL banks with low- and high-CECL job postings to ILM banks. Similar to previous tables, we generally find that the magnitude of the coefficient on Treat ×Post increases with the number of CECL jobs i t and bank size.44 Lastly, we evaluate whether banks with higher CECL job postings see less default. We repeat our analyses of Equation 4 by comparing FR Y-14Q reporting U.S. CECL banks with low- and high-CECL job postings to FR Y-14Q reporting foreign banks. Table 9 presents these results and shows that FR Y-14Q reporting U.S. CECL banks with higher CECLrelated job postings experience significantly lower loan-level default (column 2). In column 42In untabulated analysis, we separately examine the effects of CECL adoption on LLPs of homogeneous and heterogeneous loans. We find a similar pattern of larger coefficients for high-CECL job banks and large CECL banks only for heterogeneous loans. 43Again, in untabulated analysis, we separately examine the effects of CECL adoption on LLPs of homogeneous and heterogeneous loans. We find a similar pattern of larger coefficients for high-CECL job CECL banks and large CECL banks for both homogeneous and heterogeneous loans. 44To reduce concern that the length of banks’ 10-Ks or LLP-related disclosures is simply a function of their size, we take the log transformation of LLP-related disclosures. With bank fixed effects, we estimate the percentage change in the number of LLP-related sentences, which mitigates a mechanical relationship between the length of 10-K filings and bank size. 28

(3), we also find that this effect is more economically significant for the largest banks, even in this sample of large U.S. BHCs.45 These findings support our main results that investment in information systems and human capital is associated with lower future default risk. Overall, our analyses using the job posting data suggest that the investment in information systems and human capital is a plausible mechanism for the impact of CECL adoption on banks’ information production. These investments seem to be heterogeneous across banks and are more concentrated in larger banks, consistent with prior studies suggesting that larger banks have better resources for technology investment, and they enjoy greater benefits from those investments because information tends to have economies of scale. 5 Additional Analyses 5.1 Addressing the Effects of Loan Supply WeconductanalysestoaddresstheconcernthatCECLbanksmayreducelendingtoborrowers for which forecasting credit losses is more challenging, resulting in an improved estimate of LLP and fewer loan defaults. We examine the loan supply of U.S. CECL banks and foreign banks (i.e., those banks in the FR Y-14Q sample) using FR Y-9C data. In Table OA.6, we examine changes in the balance of loans divided by total assets by loan type and find that coefficients of Treat ×Post are all statistically insignificant, and their economic i t magnitudes are small, suggesting the loan supply is not statistically different between U.S. CECL banks and foreign banks. We also examine the change in deposits divided by total assets to examine whether U.S. CECL banks experienced funding constraints. We find that the deposit flow for U.S. CECL banks is not statistically different than that of foreign banks after CECL adoption. These analyses suggest that loan supply or funding constraints do not significantly differ between U.S. CECL banks and foreign banks after CECL adoption. 45Recall that our FR Y-14Q sample includes only 26 large U.S. banks. Therefore, there is little variation in bank size within the treatment group, which is a caveat of this cross-sectional analysis. 29

Next, we examine whether U.S. CECL banks strategically lend to less risky borrowers after CECL adoption. First, comparing borrower characteristics contained in FR Y-14Q by U.S. CECL banks before and after CECL adoption, we find that, on average, new borrowers after CECL adoption are smaller, have lower leverage, are more likely to be private, have a higher probability of default (PD) assigned by the bank, have higher loan maturity, and are more likely to have collateral but less likely to have guarantees or be syndicated. The average likelihood of default between the two time periods is not significantly different for new borrowers. The fact that U.S. CECL banks continue to lend to smaller, private, and higher-PD borrowers is inconsistent with the concern that they strategically decrease lending to borrowers that are ex-ante riskier. Second, we compare the average PD assigned by the bank to understand better what may drive lower default rates for U.S. CECL banks in the post-CECL adoption period. In Figure OA.3, we observe a timelier and sharper upward adjustment in PDs by U.S. CECL banks after CECL adoption. Given that the beginning of the post period overlapped with the pandemic, the upward adjustment in PDs is expected, but a timelier adjustment by U.S. CECL banks is consistent with these banks closely monitoring their borrowers, which likely lowers default rates. In Table OA.16, we also use only loans issued prior to 2020 and find consistent results. Overall, these additional analyses are inconsistent with a decrease in banks’ loan supply or a shift to ex-ante less risky borrowers by U.S. CECL banks. 5.2 Alternative Mechanisms In the paper, we mainly focus on the investment in information systems and human capital as a potential mechanism for the CECL adoption impact because both practitioners and regulators are concerned about this point. However, we acknowledge that other mechanisms could explain the effects of CECL adoption on banks’ reporting and operations. We explore three alternative mechanisms: (i) the quality of the existing internal information system, (ii) the earnings management incentive, and (iii) the capital ratio management incentive. 30

First, the quality of the internal information systems can affect the impact of CECL adoption because banks with weaker systems can benefit more from CECL adoption but also might find it more costly. We use three measures to identify banks with weak internal information systems before CECL adoption: indicators for whether a bank received enforcement actions, had restatements, or displayed internal control weakness (Enf , Restate , pre pre and ICW ). More specifically, we replicate Tables 6 – 8 using these separating variables pre and examine whether the coefficient on the variable of interest increases with the severity or relevance of these separating variables. For the enforcement actions, we split CECL-adopting banks into ones without any enforcement actions, with any enforcement actions, and with severe enforcement actions before CECL adoption. For the restatements, we split CECL banks into ones without any restatements, with any restatements, and with loan-related restatements before CECL adoption, which we obtain from Audit Analytics. Finally, for the internal control weakness tests, we split CECL banks into ones without any internal control weaknesses, with any internal control weaknesses, and with LLP-related internal control weaknesses before CECL adoption. Empirically,thepatternofcoefficientscangoineitherdirection. Ifbankswithbetterpre- CECL internal control systems benefit from CECL adoption more because of economies of scale, we expect a pattern of stronger results for banks with better internal control systems than for banks with worse internal control systems. On the other hand, if banks with better pre-CECL internal control systems benefit from CECL adoption less because the IT investment has a diminishing marginal benefit, we expect the opposite pattern of results. In Tables OA.7–OA.9, we conduct cross-sectional tests based on the separating variables discussed above.46 We find somewhat mixed results. For some tests, we observe stronger effectsinbankswithworseinternalcontrolspriortoCECLadoption. Butoverall,thepattern of coefficients is not strictly increasing along these cross-sections, which is inconsistent with 46We could not conduct these cross-sectional tests for loan default because little variation exists among the largest banks. 31

the hypotheses that these proxies are the main factors driving the impacts of CECL. We also test two potential mechanisms based on earnings management and capital ratio management incentives. These incentives are well-known factors in banks’ reporting and operational decisions (Liu and Ryan, 1995, 2006). For earnings management incentives, similar to the previous tests, we split CECL banks into three groups based on earnings before loan loss provisions and taxes divided by lagged loans, EBLLP . In this test, we use i,t quarterlyvariationsbecausebanks’earningsmanagementincentiveslikelyvaryeveryquarter depending on the level of earnings before LLPs. For capital ratio management incentives, we split CECL banks into three groups based on their capital-to-asset ratios as of 2019 Q4. In Tables OA.10–OA.11, we find that banks with higher earnings before LLPs and higher capital ratios tend to exhibit stronger coefficients. However, the pattern of coefficients is not strictly increasing along these cross-sections.47 In sum, our additional analyses suggest that the above alternative mechanisms are unlikely to drive our findings. However, we acknowledge exploring other mechanisms is empirically challenging, and our analyses cannot fully exclude these potential mechanisms. 5.3 Additional Robustness Tests Finally, we conduct several robustness tests. First, we apply coarsened exact matching (CEM) for our LLP analyses to mitigate concerns that differences between CECL and ILM banks may affect our inferences. With CEM, we coarsen the data by dividing observations into five evenly spaced bins of control variables (Size, EBLLP, Deposit, and CapRatio ) t−1 so that CECL and ILM banks have similarly weighted histograms of these variables. Then, the weights are applied in a weighted least squares regression. In Table OA.13, we find the regression coefficients and their statistical significance are similar to the analyses without matching. In addition, we limit the sample to 2018–2021 to balance the pre- and post-CECL periods and find similar results. These additional tests suggest that our findings are robust 47In untabulated analyses, we also use the time-varying capital ratios and find similar results. 32

to different model specifications and sample compositions. Second, we address the concern that the difference in recognizing LLPs by CECL and ILM banks could be driven by the COVID-19 pandemic, which coincided with CECL adoption. In particular, we examine the pattern of LLPs for banks that would have been subject to CECL and banks that would have been exempt from CECL around the financial crisis (2005–2010). Mimicking the treatment and control groups described in ASU 2016-13, we define CECL banks as public banks except for smaller reporting companies (SRCs) and ILM banks as smaller reporting companies and private banks as of 2007 Q4.48 In Figure OA.6, we plottheaverageproportionofLLPstobeginningtotalloansforhypotheticalCECLandILM banks at the quarterly frequency from 2005 to 2010. We see a gradual increase in LLPs for both banks during the financial crisis (2008–2009). Importantly, these patterns differ from the ones in Panel A of Figure 1 where we see an immediate jump in LLPs only for CECL banks in 2020 Q1, even before the pandemic effects are materialized. We believe this salient difference is consistent with CECL banks’ LLPs in 2020 being driven by CECL adoption, although the pandemic could amplify this impact. In addition, we replicate the timeliness of LLPs and reflection of local economic conditions in LLP analyses around the financial crisis (i.e., Post equals one for bank-quarters after 2008 and zero otherwise).49 In Table OA.14, we find that the coefficients on Treat ×Post ×∆NPL and Treat ×Post ×∆CoIndex i t i,t+ i t s,t+ are all statistically insignificant, suggesting that the timeliness of LLPs was not different for hypothetical CECL and ILM banks around the financial crisis, alleviating the concern that our findings are mainly driven by the pandemic effects. 48According to ASU 2016-13, public business entities, excluding SRCs as defined by the SEC, became subject to CECL for fiscal years beginning after December 15, 2019. 49Notethatwecouldnotruntheseanalysesseparatelyforhomogeneousandheterogeneousloansbecause allowances by loan type used to estimate LLP by loan type are reported in FR Y-9C starting in 2013. 33

6 Conclusion WeexaminewhethertheadoptionofCECLimprovesbanks’informationproduction. Wefind that, after CECL adoption, banks’ LLP becomes timelier and better reflects local economic conditions. We also find that banks provide better disclosures of LLPs in their 10-K filings and experience fewer loan-level defaults after CECL adoption. Notably, the effects of CECL on these outcomes increase with the number of CECL-related job postings, suggesting that investment in information systems and human capital is a plausible mechanism. Our findings suggest that accounting standards requiring the collection and analysis of forward-looking information can induce banks to produce and apply better information in their operating decisions. These findings also provide some important insights for banking regulation and supervision. In particular, our results that CECL leads banks to improve their evaluation and provisioning for credit losses can be used to explore loss rates in stress testing or inform procedures for loan-portfolio bank examinations. However, we also find that the CECL effects are more significant for larger banks, suggesting that the standarddriven benefits are likely more salient for large institutions with more resources to invest in technology and human capital. We caveat that our findings are based on large public banks that adopted CECL in 2020 when the COVID-19 pandemic began. A short recessionary period right after CECL adoption provides an empirical setting to observe starkly different provisioning by CECL and ILM banks. However, we do not rule out that large banks may have responded differently from small banks to the recession without CECL adoption. Also, most CECL banks opted to delay the impact of CECL on regulatory capital, a regulatory relief granted in response to the pandemic. An open question for future research is whether the information production effects of CECL adoption that we document will also manifest for small public and private banks subject to CECL adoption in 2023. 34

References Abad, J. and Suarez, J. (2018). The procyclicality of expected credit loss provisions. Working Paper. Acemoglu, D., Autor, D. H., Hazell, J., and Restrepo, P. (2022). Artificial intelligence and jobs: Evidence from online vacancies. Journal of Labor Economics, 40(51):S293–S340. Acharya, V. V. and Ryan, S. G. (2016). Banks’ financial reporting and financial system stability. Journal of Accounting Research, 54(2):277–340. American Bankers Association (ABA) (2016). Comment letter to the Financial Accounting Standards Board. Re: CECL. July 2. Arif, S., Bushman, R., Donovan, J., andGopalan, Y.(2022). Humanjudgmentsofexecutiveteams’humancapital. Working Paper. Armstrong, C., Kepler, J. D., Samuels, D., and Taylor, D. (2022). Causality redux: The evolution of empirical methods in accounting research and the growth of quasi-experiments. Journal of Accounting and Economics, 74:101521. Balakrishnan, K. and Ertan, A. (2021). Credit information sharing and loan loss recognition. The Accounting Review, 96(4):27–50. Ballew, H. B., Nicoletti, A., and Stuber, S. B. (2022). The effect of the Paycheck Protection Program and financial reporting standards on bank risk-taking. Management Science, 68(3):2363–2371. Barth,M.E.andLandsman,W.R.(2010). Howdidfinancialreportingcontributetothefinancialcrisis? European Accounting Review, 19(3):399–423. Beatty, A. and Liao, S. (2011). Do delays in expected loss recognition affect banks’ willingness to lend? Journal of Accounting and Economics, 52(1):1–20. Beatty, A. and Liao, S. (2014). Financial accounting in the banking industry: A review of the empirical literature. Journal of Accounting and Economics, 58(2):339–383. Beatty,A.andLiao,S.(2021). Whatdoanalysts’provisionforecaststellusaboutexpectedcreditlossrecognition? The Accounting Review, 96(1):1–21. Begenau, J., Farboodi, M., and Veldkamp, L. (2018). Big data in finance and the growth of large firms. Journal of Monetary Economics, 97:71–87. Bertomeu, J., Mahieux, L., and Sapra, H. (2023). Interplay between accounting and prudential regulation. The Accounting Review, 98(1):29–53. Bhat, G., Lee, J. A., and Ryan, S. G. (2021). Using loan loss indicators by loan type to sharpen the evaluation of banks’ loan loss accruals. Accounting Horizons, 35(3):69–91. Bhat, G., Ryan, S.G., andVyas, D.(2019). Theimplicationsofcreditriskmodelingforbanks’loanlossprovisions and loan-origination procyclicality. Management Science, 65(5):2116–2141. Bischof, J., Haselmann, R., Schlueter, O., and Stein, I. (2021a). Forward-looking information in bank reporting. Working Paper. Bischof, J., Laux, C., andLeuz, C.(2021b). Accountingforfinancialstability: Bankdisclosureandlossrecognition in the financial crisis. Journal of Financial Economics, 141(3):1188–1217. Blankespoor, E. (2019). The impact of information processing costs on firm disclosure choice: Evidence from the XBRL mandate. Journal of Accounting Research, 57(4):919–967. 35

Bozanic, Z., Roulstone, D. T., and Van Buskirk, A. (2018). Management earnings forecasts and other forwardlooking statements. Journal of Accounting and Economics, 65(1):1–20. Bushman, R. M. and Williams, C. D. (2015). Delayed expected loss recognition and the risk profile of banks. Journal of Accounting Research, 53(3):511–553. Caglio,C.,Darst,M.,andKalemli-O¨zcan,S.(2022). Risk-takingandmonetarypolicytransmission: Evidencefrom loans to SMEs and large firms. Working Paper. Chae,S.,Sarama,R.,Vojtech,C.M.,andWang,J.(2020). Theimpactofthecurrentexpectedcreditlossstandard (CECL) on the timing and comparability of reserves. Working Paper. Charoenwong, B., Kowaleski, Z. T., Kwan, A., and Sutherland, A. (2023). RegTech. Working Paper. Chen,C.-W.,Correia,M.M.,andUrcan,O.(2023).Accountingforleasesandcorporateinvestment.TheAccounting Review, 98(3):109–133. Chen, J., Dou, Y., Ryan, S. G., and Zou, Y. (2022). Does the current expected credit loss approach decrease the procyclicality of banks’ lending? Working Paper. Cheng,Q.,Cho,Y.J.,andYang,H.(2018). Financialreportingchangesandtheinternalinformationenvironment: Evidence from SFAS 142. Review of Accounting Studies, 23(1):347–383. Cohen, B. H. and Edwards, G. A. (2017). The new era of expected credit loss provisioning. BIS Quarterly Review, March:39–56. Covas, F. and Nelson, W. (2018). Current expected credit loss: Lessons from 2007-2009. Working Paper. Diamond, D. W. (1984). Financial intermediation and delegated monitoring. The Review of Economic Studies, 51(3):393–414. Dorantes,C.-A.,Li,C.,Peters,G.F.,andRichardson,V.J.(2013). Theeffectofenterprisesystemsimplementation on the firm information environment. Contemporary Accounting Research, 30(4):1427–1461. Dyer, T., Lang, M., andStice-Lawrence, L.(2017). Theevolutionof10-Ktextualdisclosure: EvidencefromLatent Dirichlet Allocation. Journal of Accounting and Economics, 64(2-3):221–245. Ertan, A. (2021). Expected losses, unexpected costs? Evidence from SME credit access under IFRS 9. Working Paper. Ertan, A., Loumioti, M., and Wittenberg-Moerman, R. (2017). Enhancing loan quality through transparency: Evidence from the European Central Bank loan level reporting initiative. Journal of Accounting Research, 55(4):877–918. Farboodi, M. and Veldkamp, L. (2022). A model of the data economy. Working Paper. FASB (2016). Accounting Standards Update (ASU) 2016-13 (ASC 326), Financial Instruments—Credit Losses. June 2016. Gee, K. H., Neilson, J. J., Schmidt, B., and Xie, B. (2022). Decision-usefulness of expected credit loss information under CECL. Working Paper. Gustafson, M. T., Ivanov, I. T., and Meisenzahl, R. R. (2021). Bank monitoring: Evidence from syndicated loans. Journal of Financial Economics, 139(2):452–477. Harris, T. S., Khan, U., and Nissim, D. (2018). The expected rate of credit losses on banks’ loan portfolios. The Accounting Review, 93(5):245–271. Hershbein, B.andKahn, L.B.(2018). Dorecessionsaccelerateroutine-biasedtechnologicalchange? evidencefrom vacancy postings. American Economic Review, 108(7):1737–72. 36

Hertzberg, A., Liberti, J. M., and Paravisini, D. (2010). Information and incentives inside the firm: Evidence from loan officer rotation. The Journal of Finance, 65(3):795–828. Howes, C. and Weitzner, G. (2023). Bank information production over the business cycle. Working Paper. Huber, S. J. (2022). Loan loss measurement and bank lending. Working Paper. Huizinga, H. and Laeven, L. (2012). Bank valuation and accounting discretion during a financial crisis. Journal of Financial Economics, 106(3):614–634. Ittner, C. D. and Michels, J. (2017). Risk-based forecasting and planning and management earnings forecasts. Review of Accounting Studies, 22(3):1005–1047. Khan, U. and Ozel, N. B. (2016). Real activity forecasts using loan portfolio information. Journal of Accounting Research, 54(3):895–937. Khan, U., Ryan, S. G., and Varma, A. (2019). Fair value versus amortized cost measurement and the timeliness of other-than-temporary impairments: Evidence from the insurance industry. The Accounting Review, 94(6):285– 307. Kim, J.-B., Ng, J., Wang, C., and Wu, F. (2021). The effect of the shift to an expected credit loss model on loan loss recognition timeliness. Working Paper. Kim, S. (2022). Delays in banks’ loan loss provisioning and economic downturns: Evidence from the U.S. housing market. Journal of Accounting Research, 60(3):711–754. Kothari,S.P.andLester,R.(2012). Theroleofaccountinginthefinancialcrisis: Lessonsforthefuture. Accounting Horizons, 26(2):335–351. Laux, C. and Leuz, C. (2010). Did fair-value accounting contribute to the financial crisis? Journal of Economic Perspectives, 24(1):93–118. Leland, H. E. and Pyle, D. H. (1977). Informational asymmetries, financial structure, and financial intermediation. The Journal of Finance, 32(2):371–387. Leuz, C. and Wysocki, P. D. (2016). The economics of disclosure and financial reporting regulation: Evidence and suggestions for future research. Journal of Accounting Research, 54(2):525–622. Li, X., Ng, J., andSaffar, W.(2023). Accounting-drivenbankmonitoringandfirms’debtstructure: Evidencefrom IFRS 9 adoption. Management Science, forthcoming. Lisowsky,P.,Minnis,M.,andSutherland,A.(2017). Economicgrowthandfinancialstatementverification. Journal of Accounting Research, 55(4):745–794. Liu, C.-C. and Ryan, S. G. (1995). The effect of bank loan portfolio composition on the market reaction to and anticipation of loan loss provisions. Journal of Accounting Research, 33(1):77–94. Liu, C.-C. and Ryan, S. G. (2006). Income smoothing over the business cycle: Changes in banks’ coordinated management of provisions for loan losses and loan charge-offs from the pre-1990 bust to the 1990s boom. The Accounting Review, 81(2):421–441. Lopez-Espinosa, G., Ormazabal, G., and Sakasai, Y. (2021). Switching from incurred to expected loan loss provisioning: Early evidence. Journal of Accounting Research, 59(3):757–804. Loudis, B. and Ranish, B. (2019). CECL and the credit cycle. Working Paper. Lu, Y. and Nikolaev, V. V. (2022). Expected loan loss provisioning: An empirical model. The Accounting Review, 97(7):319–346. 37

Mahieux, L., Sapra, H., and Zhang, G. (2023). CECL: Timely loan loss provisioning and bank regulation. Journal of Accounting Research, 61(1):3–46. McWilliams, J. (2020). Request for delay in transitions to and exclusions from certain accounting rules. Federal Deposit Insurance Corporation, March 19. Muslu,V.,Radhakrishnan,S.,Subramanyam,K.,andLim,D.(2015). Forward-lookingMD&Adisclosuresandthe information environment. Management Science, 61(5):931–948. Nichols, D. C., Wahlen, J. M., and Wieland, M. M. (2009). Publicly traded versus privately held: Implications for conditional conservatism in bank accounting. Review of Accounting Studies, 14(1):88–122. Qian, J., Strahan, P. E., and Yang, Z. (2015). The impact of incentives and communication costs on information production and use: Evidence from bank lending. The Journal of Finance, 70(4):1457–1493. Rajan, R. and Winton, A. (1995). Covenants and collateral as incentives to monitor. The Journal of Finance, 50(4):1113–1146. Roh, Y. (2023). Accounting human capital and implementation of accounting standards: Evidence from ASC 842. Working Paper. Roychowdhury,S.,Shroff,N.,andVerdi,R.S.(2019). Theeffectsoffinancialreportinganddisclosureoncorporate investment: A review. Journal of Accounting and Economics, 68(2–3):1–27. Ryan, S. G. (2017). Do the effects of accounting requirements on banks’ regulatory capital adequacy undermine financial stability? Annual Review of Financial Economics, 9:1–20. Ryan, S. G. (2019). The CECL approach. Banking Perspectives, Quarter 1:2–7. Shroff, N. (2017). Corporate investment and changes in GAAP. Review of Accounting Studies, 22(1):1–63. Stein, J. (2018). Comment letter to the FASB. American Bankers Association, September 18. Stein, J. C. (2002). Information production and capital allocation: Decentralized versus hierarchical firms. The Journal of Finance, 57(5):1891–1921. Vyas, D. (2011). The timeliness of accounting write-downs by U.S. financial institutions during the financial crisis of 2007–2008. Journal of Accounting Research, 49(3):823–860. Walker, M. J. (2019). Benefits and challenges of the “CECL” approach. Supervisory Research and Analysis Notes, 1:1–11. Wheeler, P. B. (2019). Loan loss accounting and procyclical bank lending: The role of direct regulatory actions. Journal of Accounting and Economics, 67(2):463–495. Wheeler, P. B. (2021). Unrecognized expected credit losses and bank share prices. Journal of Accounting Research, 59(3):805–866. Wilson, R. (1975). Informational economies of scale. Bell Journal of Economics, 6(1):184–195. Yang, L. (2022). An information quality-based explanation for loan loss allowance inadequacy during the 2008 financial crisis. Journal of Accounting and Economics, 73:101433. 38

Appendices A Variable Definitions Variable Definition Treat Equals one if the bank adopts CECL on January 1, 2020, and zero if the bank does not adopt CECL as of December 31, 2021. For Table 5 and Table 9, TreatequalsoneifthebankadoptsCECLonJanuary1, 2020, andzeroifthe foreign bank adopted ECL under IFRS 9 in 2018. Post Equals one for bank-quarters Q1 2019 and afterwards, and zero for bankquarters Q4 2018 and before. LLP Quarterly loan loss provisions (BHCK4230) divided by beginning total loans. LLP (w/Day 1) Quarterly loan loss provisions (BHCK4230) divided by beginning total loans but including day-1 impact for Q1 2020. LLP - Homog. Loan loss provisions for residential and consumer loans divided by beginning total loans, where provisions by loan type are estimated as ending allowance minus beginning allowance plus quarterly net charge-offs by loan type. LLP - Hetero. Loan loss provisions for construction, commercial real estate, and commercial/industrialloansdividedbybeginningtotalloans,whereprovisionsbyloan type are estimated as ending allowance minus beginning allowance plus quarterly net charge-offs by loan type. ∆NPL Ending non-performing loans (NPL) (BHCK5526 before 2018 and BHCK1403 after 2018) minus beginning NPL divided by beginning total loans. ∆NPL - Homog. Change in non-performing loans for residential and consumer loans divided by beginning total loans. ∆NPL - Hetero. Change in non-performing loans for construction, commercial real estate, and commercial/industrial loans divided by beginning total loans. ∆CoIndex Quarterly change in the weighted average of the state-level coincident index based on banks’ deposit shares in different states. Size Naturallogarithmofthebanks’beginningtotalassets(BHCK2170)inmillions. Banks with above-median total assets in a given year-quarter are considered large banks. EBLLP Earnings before loan loss provision and taxes (BHCK4301+BHCK4230) divided by beginning total loans (BHCKB528). Deposit Total deposits (BHDM6631+BHDM6636+BHFN6631+BHFN6636) divided by total assets (BHCK2170). CapRatio Total equity capital (BHCKG105) divided by total assets (BHCK2170). LLP Disc. The natural logarithm of one plus the number of LLP-related sentences in the bank’s 10-K. LLP Disc. - Fwd. The natural logarithm of one plus the number of LLP-related sentences that are forward-looking in the bank’s 10-K. Continued on next page 39

Continued from previous page Variable Definition LLP Disc. - Quant. The natural logarithm of one plus the number of LLP related sentences with quantitative information in the bank’s 10-K. SizeB Natural logarithm of the borrowers’ total assets as reported in the FR Y-14Q. Leverage Theratiooftheborrower’stotaldebtrelativetototalassetsasreportedinthe FR Y-14Q and zero otherwise. Private Equals one if a borrower is privately-held as reported in the FR Y-14Q and zero otherwise. Default Equals one if a loan defaults (i.e., 90 days past due) during the four quarters after the reporting quarter and zero otherwise. PD Probability of default for a given loan as reported in the FR Y-14Q. Maturity Loan maturity in years as reported in the FR Y-14Q. Collateral Equals one if a loan is collateralized as reported in the FR Y-14Q and zero otherwise. Guaranteed Equals one if a loan is guaranteed as reported in the FR Y-14Q and zero otherwise. Syndicated Equals one if a loan is part of a syndicate as reported in the FR Y-14Q and zero otherwise. New Borrower Equals one if a loan is originated for a borrower that has not had a previous loan with a given lender in prior quarters as reported in the FR Y-14Q and zero otherwise. CECL Jobs - Low CECL banks with a below-median number of cumulative CECL-related job postings from 2017 up to a given year-quarter. CECL Jobs - High CECL banks with an above-median number of cumulative CECL-related job postings from 2017 up to a given year-quarter. 40

B Summary Statistics of CECL-related Job Postings This appendix provides summary statistics of the CECL-related job postings on LinkUp. Panel A lists the top 10 banks with the most CECL-related job postings in 2017–2021. Panel B lists the top 10 job titles that we define as CECL-related. Panel C lists the most common SOC job classifications for CECL-related job postings and their job descriptions according to O*NET. Panel A: Top 10 CECL Job Employers Bank No. CECL Jobs % of All CECL Jobs Cum. % of All CECL Jobs Wells Fargo 1012 24.2% 24.2% Bank of America 595 14.2% 38.5% JPMorgan Chase 580 13.9% 52.4% PNC Financial 381 9.1% 61.5% SVB Financial Group 154 3.7% 65.2% Keybank 99 2.4% 67.5% American Express 95 2.3% 69.8% Discover Financial Services 74 1.8% 71.6% TD Bank 74 1.8% 73.4% Morgan Stanley 69 1.7% 75.0% Panel B: Top 10 CECL Job Titles Job Title No. CECL Jobs % of All CECL Jobs Cum. % of All CECL Jobs Credit Risk Analytics Consultant 168 4.0% 4.0% Quantitative Finance Analyst 166 4.0% 8.0% Quantitative Analytics Specialist 153 3.7% 11.7% Analytic Consultant 101 2.4% 14.1% Credit Risk Analytics Associate 46 1.1% 15.2% Credit Risk Analytics Officer 44 1.1% 16.2% QuantitativeAnalyticsConsultant 42 1.0% 17.2% Risk Analysis Specialist 42 1.0% 18.2% Credit SEC Reporting Analyst 41 1.0% 19.2% Quantitative Financial Analyst 38 0.9% 20.1% 41

Panel C: SOC Categories of CECL-related Jobs SOC Title % of CECL Jobs Top Responsibilities -Advise clients on aspects of capitalization, such as amounts, sources, or timing. Financial Analysts & -Analyze financial or operational performance of companies facing financial difficulties 13-2051.00 32.8% Investment Analysts to identify or recommend remedies. -Assess companies as investments for clients by examining company facilities. -Establish and maintain relationships with individual or business customers or provide assistance with problems these customers may encounter. 11-3031.02 Financial Managers 23.6% -Plan,direct,orcoordinatetheactivitiesofworkersinbranches,offices,ordepartments of establishments, such as branch banks, brokerage firms, risk and insurance departments, or credit departments. -Document findings of study and prepare recommendations for implementation of new systems, procedures, or organizational changes. 13-1111.00 Management Analysts 17.0% -Analyze data and other information gathered to develop solutions or alternative methods of proceeding. -Analyze credit data and financial statements to determine the degree of risk involved in extending credit or lending money. 13-2041.00 Credit Analysts 10.1% -Complete loan applications, including credit analyses and summaries of loan requests, and submit to loan committees for approval. -Use computer programs to evaluate customers’ financial status. -Collect and analyze data on customer demographics, preferences, needs, and buying Market Research & habits to identify potential markets and factors affecting product demand. 13-1161.00 3.5% Marketing Specialists -Deviseandevaluatemethodsandproceduresforcollectingdata, suchassurveys, opinion polls, or questionnaires, or arrange to obtain existing data. -Prepare detailed reports on audit findings. 13-2011.01 Accountants & Auditors 3.4% -Collect and analyze data to detect deficient controls, duplicated effort, extravagance, fraud, or non-compliance with laws, regulations, and management policies. 42

C Examples of Pre- and Post-CECL LLP Disclosures This appendix illustrates differences in LLP-related disclosures between JPMorgan Chase’s 2019 and 2020 10-Ks. We select the first page in each year’s 10-K that is specifically dedicated to discussions of LLP. The same page repeats in the financial statement footnotes. Highlighted texts reflect added LLP disclosures that are forward-looking and/or quantitative. Note that the table is not captured by the algorithm described in subsection 4.2 since all tables are dropped when processing 10-K filings. Importantly, the incremental disclosure in 2020’s 10-K persists into 2021. JP Morgan Chase 2019 10-K JP Morgan Chase 2020 10-K Management’s discussion and analysis ALLOWANCE FOR CREDIT LOSSES The Firm’s allowance for credit losses covers the retained The allowance for credit losses decreased compared with consumer and wholesale loan portfolios, as well as the December 31, 2018 driven by: Firm’s wholesale and certain consumer lending-related • an $800 million reduction in the CCB allowance for loan commitments. losses, which included $650 million in the PCI residential Refer to Critical Accounting Estimates Used by the Firm on real estate portfolio, reflecting continued improvement in pages 136–138 and Note 13 for further information on the home prices and delinquencies; $100 million in the non components of the allowance for credit losses and related credit-impaired residential real estate portfolio; and $50 management judgments. million in the business banking portfolio; as well as • a $151 million reduction for write-offs of PCI loans, At least quarterly, the allowance for credit losses is reviewed by the CRO, the CFO and the Controller of the predominantly offset by Firm. As of December 31, 2019, JPMorgan Chase deemed • a $500 million addition to the allowance for loan losses in the allowance for credit losses to be appropriate and the credit card portfolio reflecting loan growth and sufficient to absorb probable credit losses inherent in the higher loss rates as newer vintages season and become a portfolio. larger part of the portfolio, and • a $251 million addition in the wholesale allowance for credit losses driven by select client downgrades. Refer to Consumer Credit Portfolio on pages 103–107, Wholesale Credit Portfolio on pages 108–115 and Note 12 for additional information on the consumer and wholesale credit portfolios. 116 JPMorgan Chase & Co./2019 Form 10-K 43

Figure 1: Loan Loss Provisioning This figure plots the average loan loss provisioning to beginning total loans of banks that adopted CECL on January 1, 2020 (CECL) and banks not subject to CECL adoption (ILM). Panel A reports LLPs for total loans. For CECL-adopting banks, we additionally plot the LLPs with the day-1 impact for Q1 2020, which bypassestheincomestatement. PanelBandPanelCreportLLPsforhomogeneousandheterogeneousloans, respectively. For homogeneous and heterogeneous loans, LLPs are estimated as the change in the allowance plus net charge-offs for each loan type. Panel A: LLP - All Loans snaoL LLA - PLL naeM 800. 600. 400. 200. 0 200.- 2017q1 2018q1 2019q1 2020q1 2021q1 2022q1 CECL w/o Day1 CECL with Day1 ILM Panel B: LLP - Homog. Loans snaoL suoenegomoH - PLL naeM 600. 400. 200. 0 200.- Panel C: LLP - Hetero. Loans 2017q1 2018q1 2019q1 2020q1 2021q1 2022q1 CECL ILM snaoL suoenegoreteH - PLL naeM 600. 400. 200. 0 200.- 2017q1 2018q1 2019q1 2020q1 2021q1 2022q1 CECL ILM 44

Figure 2: LLP-related Disclosure This figure plots the number of LLP-related sentences in 10-Ks by banks that adopted CECL on January 1, 2020 (CECL) and banks not subject to CECL adoption (ILM). Panel A reports the number of unique sentencesfallingwithinthe(-3,+3)windowofany10-KsentenceinwhichthereisanLLPsentence. PanelB and Panel C report the number of sentences containing forward-looking words and quantitative information (i.e., hard numbers) among such LLP-related sentences, respectively. The shaded areas represents 95% confidence intervals. Panel A: LLP Disc. - All Sentences secnetneS PLL 007 036 065 094 024 053 2017 2018 2019 2020 2021 Fiscal Year CECL ILM Panel B: LLP Disc. - Fwd. Sentences gnikool-drawroF ,secnetneS PLL 051 031 011 09 07 05 Panel C: LLP Disc. - Quant. Sentences 2017 2018 2019 2020 2021 Fiscal Year CECL ILM evitatitnauQ ,secnetnetneS PLL 051 031 011 09 07 05 2017 2018 2019 2020 2021 Fiscal Year CECL ILM 45

Figure 3: Number of CECL-related Job Postings for CECL vs. ILM Banks This figure plots CECL-related and informational job postings on LinkUp by banks that adopted CECL on January 1, 2020 (CECL) and banks not subject to CECL adoption (ILM). A CECL-related job is defined if job descriptions contain one of the terms “CECL,” “Current Expected Credit Losses,” “ASU 2016-13,” “ASC326,”“Topic326,”and“FinancialInstrument(s)CreditLoss(es).” Aninformationaljobisdefinedifa jobsharesanyO*NETSOCcodeswithCECL-relatedjobs. PanelAplotsthetotalnumberofCECL-related job postings by CECL and ILM banks, Panel B plots the total number of CECL-related (informational) job postingsbyCECL(ILM)banks,andPanelCplotstheaveragenumberofCECL-related(informational)job postings by CECL (ILM) banks. Panel B and Panel C are indexed to 2017 Q1. Panel A: Total CECL Job Postings gnitsoP boJ LCEC latoT 006 004 002 0 2017q1 2018q1 2019q1 2020q1 2021q1 2022q1 date CECL ILM Panel B: Total CECL (Informational) Job Postings, Index: 2017Q1=1 1=1Q7102:xednI ,latoT 02 51 01 5 0 Panel C: Mean CECL (Informational) Job Postings, Index: 2017Q1=1 2017q1 2018q1 2019q1 2020q1 2021q1 2022q1 date CECL ILM 1=1Q7102:xednI ,naeM 51 01 5 0 2017q1 2018q1 2019q1 2020q1 2021q1 2022q1 date CECL ILM 46

Figure 4: Frequently Used Words in CECL-related Job Postings This figure plots word clouds for the most frequently used words in CECL-related job postings. The word clouds are generated using bag-of-words (BOW) document vectors. Panel A displays the words used in the job titles. Panel B displays the words used in the job descriptions. Larger font sizes indicate higher frequency. Panel A: Word Cloud: Job Titles Panel B: Word Cloud: Job Descriptions 47

Table 1: Descriptive Statistics Thistablereportsthedescriptivestatistics. VariablesexpressingLLP and∆NPLareinpercentages. PanelApresentssummarystatisticsofbank-level characteristics. Panel B presents summary statistics of the additional loan- or borrower-level characteristics for our loan-level analyses. Columns (1) to (8) provide descriptive statistics for the full sample. Columns (9) to (14) show the mean differences for the samples of CECL and control banks (ILM banks in Panel A and IHCs in Panel B). All variables are defined in Appendix A. *, **, and *** indicate statistical significance of the mean differences at the 10%, 5%, and 1% levels, respectively. (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) (13) (14) Full Sample CECL Banks Crtl. Banks Two-sample t-test N Mean Std. Dev. 10th 25th Median 75th 90th N Mean N Mean Diff. p-value Panel A. Bank-level Chars. LLP 5,488 0.081 0.196 -0.026 0.005 0.038 0.086 0.215 2,941 0.091 2,547 0.069 0.021*** <0.001 LLP (w/Day 1) 5,488 0.089 0.232 -0.026 0.005 0.037 0.086 0.223 2,941 0.105 2,547 0.070 0.035*** <0.001 LLP - Homog. 4,544 0.040 0.167 -0.022 -0.003 0.007 0.029 0.083 2,886 0.048 1,658 0.027 0.021*** <0.001 LLP - Hetero. 4,539 0.044 0.137 -0.042 -0.002 0.020 0.055 0.142 2,888 0.050 1,651 0.034 0.016*** <0.001 ∆NPL 5,488 0.004 0.197 -0.147 -0.058 -0.006 0.045 0.165 2,941 0.004 2,547 0.004 -0.000 0.975 ∆CoIndex 5,068 0.007 0.044 0.001 0.005 0.009 0.014 0.027 2,852 0.007 2,216 0.007 0.000 0.909 Size 5,488 9.084 1.579 7.328 8.057 8.757 9.845 11.125 2,941 9.930 2,547 8.106 1.824*** <0.001 EBLLP 5,488 0.008 0.012 0.004 0.005 0.006 0.008 0.012 2,941 0.009 2,547 0.008 0.001*** <0.001 Deposit 5,488 0.772 0.126 0.664 0.749 0.801 0.844 0.869 2,941 0.759 2,547 0.788 -0.029*** <0.001 CapRatio 5,488 0.116 0.039 0.082 0.095 0.110 0.128 0.150 2,941 0.120 2,547 0.112 0.008*** <0.001 LLP Disc. 851 6.233 0.479 5.897 6.125 6.297 6.471 6.644 728 6.266 123 6.037 0.229*** <0.001 LLP Disc. - Fwd. 851 4.654 0.551 4.220 4.489 4.718 4.956 5.170 728 4.701 123 4.373 0.329*** <0.001 LLP Disc. - Quant. 851 4.582 0.578 4.094 4.407 4.673 4.913 5.112 728 4.624 123 4.334 0.290*** <0.001 Panel B. Borrower- or Loan-level Chars. SizeB 688,340 18.455 2.981 14.990 16.344 17.968 20.335 22.684 620,621 18.532 67,719 17.746 0.786*** <0.001 Leverage 688,340 0.401 0.254 0.096 0.212 0.365 0.553 0.747 620,621 0.398 67,719 0.429 -0.031*** <0.001 Private 688,340 0.848 0.359 0 1 1 1 1 620,621 0.843 67,719 0.892 -0.049*** <0.001 Default 688,340 0.310 5.557 0 0 0 0 0 620,621 0.296 67,719 0.440 -0.145*** <0.001 PD 688,340 0.020 0.042 0.002 0.004 0.009 0.019 0.038 620,621 0.020 67,719 0.023 -0.003*** <0.001 Maturity 688,340 4.955 4.105 1.167 2.463 4.005 6.142 9.334 620,621 4.650 66,495 7.755 -3.105*** <0.001 Collateral 688,340 0.924 0.265 1 1 1 1 1 620,621 0.921 67,719 0.952 -0.031*** <0.001 Guaranteed 688,340 0.504 0.500 0 0 1 1 1 620,621 0.488 67,719 0.656 -0.168*** <0.001 Syndicated Loan 688,340 0.189 0.391 0 0 0 0 1 620,621 0.194 67,719 0.137 0.057*** <0.001 New Borrower 688,340 0.031 0.174 0 0 0 0 0 620,621 0.032 67,719 0.030 0.002*** 0.006 48

Table 2: Timeliness of Loan Loss Provisioning This table reports the results of estimating the timeliness of LLPs using Equation 1. The dependent variables in columns (1)–(4) are LLPs for all loans, LLPs with day-1 impact for all loans, LLPs for homogeneous loans, and LLPsforheterogeneousloans,respectively. TreatequalsoneforbanksthatadoptedCECLonJanuary1,2020and zeroforbanksthatdidnotadoptCECLasofDecember31,2021. Postequalsoneforbank-quartersafter2020and zero otherwise. ∆NPL (-Homog./Hetero.) is the change in non-performing (homogeneous/heterogeneous) loans divided by the beginning total loans. Standard errors reported in parentheses are clustered by bank. All variables are defined in Appendix A. *, **, and *** indicate statistical significance of the coefficients at the 10%, 5%, and 1% levels, respectively. (1) (2) (3) (4) Dep. Var. LLP LLP (w/ Day 1) LLP - Homog. LLP - Hetero. t t t t Treat×Post×∆NPL 0.320** 0.512*** -0.143 0.521*** t+ (0.125) (0.146) (0.438) (0.149) Treat×Post×∆NPL 0.229*** 0.350*** 0.299* 0.333* t (0.073) (0.097) (0.173) (0.201) Treat×Post×∆NPL -0.004 -0.022 0.397 -0.107 t− (0.082) (0.099) (0.246) (0.129) Treat×∆NPL 0.033 0.028 0.082 0.016 t+ (0.037) (0.037) (0.117) (0.036) Treat×∆NPL 0.031 0.031 0.023 0.033 t (0.026) (0.031) (0.085) (0.029) Treat×∆NPL -0.049 -0.045 -0.261* -0.002 t− (0.045) (0.048) (0.155) (0.027) Post×∆NPL -0.007 -0.066 0.260 -0.331*** t+ (0.077) (0.083) (0.381) (0.108) Post×∆NPL -0.028 -0.035 0.073 -0.126 t (0.051) (0.060) (0.045) (0.184) Post×∆NPL 0.068 0.092* -0.190 0.234** t− (0.045) (0.052) (0.147) (0.091) ∆NPL -0.009 -0.007 0.061** -0.017 t+ (0.013) (0.013) (0.028) (0.020) ∆NPL 0.009 0.010 0.034 0.037** t (0.011) (0.011) (0.033) (0.016) ∆NPL 0.027 0.028 0.095*** 0.053*** t− (0.018) (0.019) (0.036) (0.018) Treat×Post 0.000 0.001*** 0.000*** 0.000** (0.000) (0.000) (0.000) (0.000) Observations 4,863 4,863 4,116 4,114 Bank FE Yes Yes Yes Yes Year-Quarter FE Yes Yes Yes Yes Controls Yes Yes Yes Yes Adj. Overall R-squared 0.576 0.546 0.581 0.399 Adj. Within R-squared 0.048 0.064 0.020 0.059 49

Table 3: Reflection of Local Economic Conditions in Provisions ThistablereportstheresultsofestimatingtheincorporationoflocaleconomicconditionsinLLPsusingEquation2. The dependent variables in columns (1)–(4) are LLPs for all loans, LLPs with day-1 impact for all loans, LLPs for homogeneous loans, and LLPs for heterogeneous loans, respectively. Treat equals one for banks that adopted CECL on January 1, 2020 and zero for banks that did not adopt CECL as of December 31, 2021. Post equals one forbank-quartersafter 2020andzerootherwise. ∆CoIndex isthe changeinthe weightedaverageofthe state-level coincident index based on banks’ deposit shares in different states. Standard errors reported in parentheses are clustered by bank. All variables are defined in Appendix A. *, **, and *** indicate statistical significance of the coefficients at the 10%, 5%, and 1% levels, respectively. (1) (2) (3) (4) Dep. Var. LLP LLP (w/Day1) LLP - Homog. LLP - Hetero. t t t t Treat×Post×∆CoIndex -0.035*** -0.065*** -0.017* -0.029*** t+ (0.005) (0.007) (0.009) (0.008) Treat×Post×∆CoIndex -0.016* -0.016 -0.007 -0.026** t (0.009) (0.011) (0.006) (0.011) Treat×Post×∆CoIndex -0.021* -0.015 -0.009 -0.026** t− (0.012) (0.014) (0.009) (0.012) Treat×∆CoIndex 0.001 0.001 -0.001 -0.004** t+ (0.001) (0.001) (0.001) (0.002) Treat×∆CoIndex -0.001 -0.008 0.000 0.013 t (0.009) (0.010) (0.006) (0.009) Treat×∆CoIndex 0.008 -0.003 0.002 0.020 t− (0.012) (0.014) (0.009) (0.013) Post×∆CoIndex 0.035*** 0.064*** 0.033** 0.031*** t+ (0.007) (0.012) (0.015) (0.009) Post×∆CoIndex 0.009 0.007 -0.000 0.016 t (0.008) (0.010) (0.006) (0.010) Post×∆CoIndex 0.020** 0.013 0.010 0.024** t− (0.010) (0.011) (0.008) (0.012) ∆CoIndex -0.001 -0.002* 0.001 -0.001 t+ (0.001) (0.001) (0.001) (0.002) ∆CoIndex 0.007 0.014 0.011* -0.003 t (0.008) (0.009) (0.007) (0.008) ∆CoIndex -0.006 0.005 0.000 -0.018 t− (0.010) (0.012) (0.008) (0.012) Treat×Post 0.001*** 0.001*** 0.000*** 0.001*** (0.000) (0.000) (0.000) (0.000) Observations 4,738 4,738 3,941 3,938 Bank FE Yes Yes Yes Yes Year-Quarter FE Yes Yes Yes Yes Controls Yes Yes Yes Yes Adj. Overall R-squared 0.581 0.567 0.563 0.408 Adj. Within R-squared 0.083 0.122 0.029 0.052 50

Table 4: LLP-related Disclosures This table reports the results of estimating the increased LLP-related disclosure in banks’ 10-Ks using Equation 3. The dependent variables in columns (1)–(3) are the natural logarithms of one plus the number of LLP-related sentences, LLP-related forward-looking sentences, and LLP-related quantitative sentences, respectively. Treat equals one for banks that adopted CECL on January 1, 2020 and zero for banks that did not adopt CECL as of December 31, 2021. Post equals one for bank-quarters after 2020 and zero otherwise. Standard errors reported in parentheses are clustered by bank. All variables are defined in Appendix A. *, **, and *** indicate statistical significance of the coefficients at the 10%, 5%, and 1% levels, respectively. (1) (2) (3) Dep. Var. LLP Disc. LLP Disc. - Fwd. LLP Disc. - Quant. Treat×Post 0.124*** 0.201*** 0.085** (0.028) (0.037) (0.041) Size 0.152*** 0.151** 0.205*** t (0.050) (0.069) (0.075) EBLLP 0.568 -0.891 1.862* t (0.642) (1.082) (0.982) ∆NPL -1.479 -1.932 -6.527* t (2.475) (3.810) (3.326) Deposit -0.137 0.056 0.018 t (0.218) (0.272) (0.314) CapRatio 0.189 0.865 0.165 t−1 (0.434) (0.730) (0.614) Observations 851 851 851 Firm FE Yes Yes Yes Year FE Yes Yes Yes Adj. Overall R-squared 0.933 0.909 0.873 Adj. Within R-squared 0.055 0.058 0.019 51

Table 5: Loan-level Default This table reports the results of estimating the decrease in loan-level default using Equation 4. Treat equals one for FR Y-14Q reporting banks that adopted CECL on January 1, 2020 and zero for FR Y-14Q reporting foreign banksthatadoptedIFRS9in2018. Postequalsoneforbank-quartersafter2020andzerootherwise. Observations start in 2018 to incorporate IFRS adoption of ECL. Standard errors reported in parentheses are clustered by bank. All variables are defined in Appendix A. *, **, and *** indicate statistical significance of the coefficients at the 10%, 5%, and 1% levels, respectively. (1) (2) (3) (4) (5) (6) (7) Dep. Var. Default Default Default Default Default Default Default Split. Vars. All Existing vs. New Private vs. Public High vs. Low Borrowers Borrowers PD Treat×Post -0.492*** -0.493*** -0.707*** -0.529*** -0.076 -0.574*** -0.085 (0.090) (0.093) (0.203) (0.096) (0.231) (0.114) (0.138) Size -0.187 -0.229 -0.521 -0.044 -0.520 -0.121 -0.275 t (0.255) (0.244) (0.379) (0.237) (0.515) (0.278) (0.286) EBLLP 24.616** 26.219** 4.925 24.299** 27.642 33.886*** 3.900 t (10.514) (10.778) (31.262) (10.450) (30.776) (10.546) (17.604) Deposit 0.005 -0.088 0.319 0.039 -0.967 -0.531 1.391 t (0.564) (0.589) (1.521) (0.632) (1.700) (0.760) (0.902) CapRatio -3.837 -3.710 -7.326 -3.268 -4.433 -4.943 -2.227 t−1 (3.273) (3.160) (7.107) (2.804) (6.460) (3.459) (2.837) SizeB 0.035 0.033 0.082*** 0.013 0.084 0.008 0.060** t (0.023) (0.023) (0.026) (0.012) (0.065) (0.018) (0.026) Leverage 0.143*** 0.132*** 0.309*** 0.110** 0.200 0.088 0.307** t (0.043) (0.043) (0.091) (0.053) (0.198) (0.058) (0.145) Private -0.331*** -0.328*** -0.611 -0.304*** -0.323*** (0.073) (0.076) (0.371) (0.078) (0.102) PD 10.242*** 10.386*** 5.704* 10.203*** 9.723* 10.446*** -15.033 t (2.168) (2.166) (3.103) (2.202) (5.459) (2.198) (19.126) Maturity 0.000 0.000 -0.000*** -0.000*** 0.000 0.000 0.000 t (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) Collateral 0.042 0.039 0.119 -0.027 0.078 -0.064 0.190*** (0.045) (0.044) (0.167) (0.053) (0.105) (0.090) (0.061) Guaranteed -0.033 -0.033 -0.031 0.015 -0.230 0.007 -0.118 (0.047) (0.047) (0.059) (0.026) (0.181) (0.033) (0.089) Syndicated Loan -0.346*** -0.342*** -0.461*** -0.155*** -0.862*** -0.229*** -0.458*** (0.110) (0.112) (0.162) (0.042) (0.281) (0.071) (0.145) New Borrower 0.047 -0.009 0.497 0.002 0.193 (0.065) (0.042) (0.370) (0.045) (0.133) Observations 700,410 678,116 22,293 593,242 107,168 476,687 223,723 Bank FE Yes Yes Yes Yes Yes Yes Yes Loan Type FE Yes Yes Yes Yes Yes Yes Yes Year-Quarter FE Yes Yes Yes Yes Yes Yes Yes Adj. Overall R-squared 0.008 0.008 0.008 0.008 0.013 0.010 0.004 Adj. Within R-squared 0.006 0.006 0.005 0.007 0.004 0.008 0.002 52

Table 6: CECL-induced Information Production: Timeliness This table replicates Table 2, estimating the timeliness of LLPs using Equation 1 by comparing CECL banks with low- and high-CECL job postings to ILM banks. CECL jobs are calculated as the cumulative number of CECLrelated job postings from 2017 to a given year-quarter. Large banks are CECL banks with above-median total assets in a given year-quarter. Standard errors reported in parentheses are clustered by bank. All variables are defined in Appendix A. *, **, and *** indicate statistical significance of the coefficients at the 10%, 5%, and 1% levels, respectively. (1) (2) (3) (4) (5) (6) Dep. Var. LLP LLP (w/ Day 1) t t CECL Bank Split. CECL Jobs Low High High Low High High Size All All Large All All Large Treat×Post×∆NPL 0.325* 0.589*** 0.747** 0.558*** 0.873*** 1.169*** t+ (0.168) (0.226) (0.307) (0.199) (0.266) (0.339) Treat×Post×∆NPL 0.140* 0.432*** 0.454*** 0.300** 0.566*** 0.655*** t (0.073) (0.110) (0.119) (0.130) (0.127) (0.142) Treat×Post×∆NPL -0.026 0.082 0.273 0.030 0.027 0.278 t− (0.102) (0.164) (0.190) (0.115) (0.208) (0.216) Treat×∆NPL 0.029 0.013 0.003 0.028 0.015 -0.022 t+ (0.057) (0.051) (0.072) (0.057) (0.060) (0.077) Treat×∆NPL 0.010 -0.002 -0.034 0.012 0.004 -0.064** t (0.029) (0.032) (0.028) (0.032) (0.051) (0.032) Treat×∆NPL -0.094 0.052 -0.022 -0.081 0.041 -0.073 t− (0.072) (0.051) (0.046) (0.076) (0.067) (0.063) Post×∆NPL 0.045 0.024 0.028 0.011 -0.015 -0.008 t+ (0.075) (0.077) (0.077) (0.078) (0.082) (0.082) Post×∆NPL -0.006 -0.015 -0.016 -0.012 -0.022 -0.022 t (0.049) (0.050) (0.050) (0.055) (0.057) (0.056) Post×∆NPL 0.055 0.067 0.063 0.070 0.082* 0.077 t− (0.042) (0.043) (0.042) (0.045) (0.048) (0.047) ∆NPL -0.009 -0.010 -0.010 -0.007 -0.009 -0.009 t+ (0.013) (0.013) (0.014) (0.013) (0.013) (0.013) ∆NPL 0.009 0.009 0.009 0.009 0.009 0.009 t (0.011) (0.011) (0.011) (0.011) (0.011) (0.011) ∆NPL 0.029 0.024 0.024 0.031 0.025 0.025 t− (0.018) (0.018) (0.018) (0.019) (0.018) (0.018) Treat×Post 0.000 0.000 -0.000 0.001*** 0.001*** 0.001*** (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) Observations 3,648 3,039 2,870 3,648 3,039 2,870 Firm FE Yes Yes Yes Yes Yes Yes Year-Quarter FE Yes Yes Yes Yes Yes Yes Controls Yes Yes Yes Yes Yes Yes Adj. Overall R-squared 0.540 0.593 0.601 0.522 0.537 0.547 Adj. Within R-squared 0.0448 0.0682 0.0711 0.0734 0.0769 0.0867 53

Table 7: CECL-induced Information Production: Local Economic Conditions This table replicates Table 3, estimating the incorporation of local economic conditions in LLPs using Equation 2 by comparing CECL banks with low- and high-CECL job postings to ILM banks. CECL jobs are calculated as the cumulative number of CECL-related job postings from 2017 to a given year-quarter. Large banks are CECL banks with above-median total assets in a given year-quarter. Standard errors reported in parentheses are clustered by bank. All variables are defined in Appendix A. *, **, and *** indicate statistical significance of the coefficients at the 10%, 5%, and 1% levels, respectively. (1) (2) (3) (4) (5) (6) Dep. Var. LLP LLP (w/ Day 1) t t CECL Bank Split. CECL Jobs Low High High Low High High Size All All Large All All Large Treat×Post×∆CoIndex -0.027*** -0.047*** -0.049*** -0.053*** -0.078*** -0.086*** t+ (0.005) (0.008) (0.009) (0.007) (0.012) (0.015) Treat×Post×∆CoIndex -0.024** -0.016 -0.024 -0.029** -0.022 -0.021 t (0.011) (0.014) (0.018) (0.012) (0.017) (0.022) Treat×Post×∆CoIndex -0.029* -0.018 -0.033 -0.030* -0.017 -0.014 t− (0.015) (0.020) (0.028) (0.016) (0.022) (0.031) Treat×∆CoIndex 0.001 0.001 0.001 0.001 0.001 0.001 t+ (0.001) (0.001) (0.001) (0.001) (0.001) (0.001) Treat×∆CoIndex 0.012 -0.006 -0.001 0.011 -0.007 -0.012 t (0.010) (0.014) (0.018) (0.012) (0.017) (0.022) Treat×∆CoIndex 0.020 -0.002 0.011 0.018 -0.008 -0.015 t− (0.015) (0.021) (0.029) (0.016) (0.023) (0.031) Post×∆CoIndex 0.018*** 0.030*** 0.027*** 0.034*** 0.051*** 0.045*** t+ (0.006) (0.008) (0.009) (0.009) (0.014) (0.014) Post×∆CoIndex 0.009 0.011 0.011 0.008 0.008 0.009 t (0.008) (0.009) (0.009) (0.009) (0.010) (0.010) Post×∆CoIndex 0.020** 0.017* 0.018* 0.015 0.009 0.012 t− (0.010) (0.010) (0.010) (0.011) (0.011) (0.011) CoIndex -0.000 -0.002 -0.001 -0.001 -0.003** -0.002* t+ (0.001) (0.001) (0.001) (0.001) (0.001) (0.001) CoIndex -0.000 0.002 0.000 0.004 0.008 0.006 t (0.009) (0.009) (0.009) (0.009) (0.010) (0.010) CoIndex -0.012 -0.007 -0.008 -0.005 0.004 0.001 t− (0.010) (0.011) (0.011) (0.012) (0.013) (0.013) Treat×Post 0.001*** 0.001*** 0.001** 0.002*** 0.002*** 0.002*** (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) Observations 3,507 2,885 2,708 3,507 2,885 2,708 Firm FE Yes Yes Yes Yes Yes Yes Year-Quarter FE Yes Yes Yes Yes Yes Yes Controls Yes Yes Yes Yes Yes Yes Adj. Overall R-squared 0.535 0.599 0.609 0.530 0.558 0.568 Adj. Within R-squared 0.0546 0.116 0.126 0.104 0.146 0.158 54

Table 8: CECL-induced Information Production: LLP Disclosures This table replicates Table 4, estimating the LLP-related disclosure using Equation 3 by comparing CECL banks with low- and high-CECL job postings to ILM banks. CECL jobs are calculated as the cumulative number of CECL-related job postings from 2017 to a given year-quarter. Large banks are CECL banks with above-median total assets in a given year-quarter. Treat equals one for banks that adopted CECL on January 1, 2020 and zero for banks that did not adopt CECL as of December 31, 2021. Post equals one for bank-quarters after 2020 and zero otherwise. Standard errors reported in parentheses are clustered by bank. All variables are defined in Appendix A. *, **, and *** indicate statistical significance of the coefficients at the 10%, 5%, and 1% levels, respectively. (1) (2) (3) (4) (5) (6) (7) (8) (9) Dep. Var. LLP Disc. LLP Disc. - Fwd. LLP Disc. - Quant. CECL Bank Split. CECL Jobs Low High High Low High High Low High High Bank Size All All Large All All Large All All Large Treat×Post 0.115*** 0.133*** 0.170*** 0.183*** 0.203*** 0.241*** 0.073 0.100* 0.129** (0.029) (0.034) (0.032) (0.038) (0.045) (0.046) (0.044) (0.054) (0.057) Observations 483 361 307 483 361 307 483 361 307 Firm FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Year FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Controls Yes Yes Yes Yes Yes Yes Yes Yes Yes Adj. Overall R-squared 0.954 0.963 0.976 0.921 0.946 0.953 0.916 0.922 0.930 Adj. Within R-squared 0.087 0.094 0.176 0.093 0.123 0.178 0.016 0.022 0.023 55

Table 9: CECL-induced Information Production: Loan-level Default This table reports the results of estimating changes in loan-level default using Equation 4 by comparing FR Y-14Q reporting CECL banks with low- and high-CECL job postings to FR Y-14Q reporting foreign banks. CECL jobs are calculated as the cumulative number of CECL-related job positions from 2017 to a given year-quarter. Large banks are FR Y-14Q reporting CECL banks with above-median total assets in a given quarter. Treat equals one for FR Y-14Q reporting banks that adopted CECL on January 1, 2020 and zero for FR Y-14Q reporting foreign banksthatadoptedIFRS9in2018. Postequalsoneforbank-quartersafter2020andzerootherwise. Observations start in 2018 to incorporate IFRS adoption of ECL. Standard errors reported in parentheses are clustered by bank. All variables are defined in Appendix A. *, **, and *** indicate statistical significance of the coefficients at the 10%, 5%, and 1% levels, respectively. (1) (2) (3) Dep. Var. Default Default Default CECL Bank Split. CECL Jobs Low High High Bank Size All All Large Treat×Post -0.402*** -0.509*** -0.480*** (0.088) (0.087) (0.110) Observations 230,290 486,499 389,076 Bank FE Yes Yes Yes Loan Type FE Yes Yes Yes Year-Quarter FE Yes Yes Yes Controls Yes Yes Yes Adj. Overall R-squared 0.011 0.007 0.005 Adj. Within R-squared 0.009 0.006 0.004 56

Appendix Document (Not Intended for Publication) Current Expected Credit Losses (CECL) Standard and Banks’ Information Production This version: July 2023

Table of Contents Figures Figure OA.1 Loan Loss Provisioning by Delay Banks 1 Figure OA.2 Number of CECL-related Job Postings for Delay Banks 2 Figure OA.3 Average PD Rates for All, Existing, and New Borrowers 3 Figure OA.4 Average PDs for High and Low Default Risk Borrowers 4 Figure OA.5 Size Distributions of CECL Banks Purchasing External Models 5 Figure OA.6 Loan Loss Provisioning around the Financial Crisis 6 Figure OA.7 Loan Loss Provisioning by Foreign Banks 7 Figure OA.8 Time Trends of Job Postings 8 Tables Table OA.1 Understanding Banks’ Delayed CECL Adoption 9 Table OA.2 Timeliness of Loan Loss Provisioning, Various Fixed Effects 11 Table OA.3 Reflection of Local Economic Conditions, Various Fixed Effects 12 Table OA.4 External vs. Internal Modeling: Descriptives 13 Table OA.5 External vs. Internal Modelling, Regressions 15 Table OA.6 Changes in Loan Supplies 21 Table OA.7 Weakness in Internal Control System, Enforcement 22 Table OA.8 Weakness in Internal Control System, ICW 25 Table OA.9 Weakness in Internal Control System, Restatement 28 Table OA.10 Reporting Manipulation Incentives, EBLLP 31 Table OA.11 Reporting Manipulation Incentives, Capital Ratio 35 Table OA.12 Comparison of New and Existing Borrowers 39 Table OA.13 Informativeness of Loan Loss Provisioning, Matched Sample 40 Table OA.14 Loan Loss Provisioning around the Financial Crisis 42 Table OA.15 Loan-level Default, robustness 44 Table OA.16 Loan-Level Default, Focusing on Loans issued prior to 2020 47

Figure OA.1: Loan Loss Provisioning by Delay Banks This figure plots the average loan loss provisions to beginning total loans for banks that delayed CECL adoption under the Coronavirus Aid, Relief, and Economic Security (CARES) Act exemption and adopted CECL later. We divide the delayed adoption banks into three groups based on their delayed adoption date (2020 Q4, 2021 Q1, and 2022 Q1).Panel A reports LLPs with the day-1 impact for total loans. Panel B reports LLPs without the day-1 impactfortotalloans. PanelCreportsLLPsforhomogeneousloans. PanelDreportsLLPsforheterogeneousloans. For homogeneous and heterogeneous loans, LLPs are estimated as the change in allowance plus net charge-offs for each loan type. Panel A: LLP with Day1 - All Loans snaoL llA - 1yaD htiw PLL naeM 800. 600. 400. 200. 0 200.- Panel B: LLP without Day1 - All Loans 2017q1 2018q1 2019q1 2020q1 2021q1 2022q1 date 2020 Q4 2021 Q1 2022 Q1 snaoL llA - 1yaD tuohtiw PLL naeM 800. 600. 400. 200. 0 200.- 2017q1 2018q1 2019q1 2020q1 2021q1 2022q1 date 2020 Q4 2021 Q1 2022 Q1 Panel C: LLP - Homog. Loans snaoL .gomoH - PLL naeM 800. 600. 400. 200. 0 200.- Panel D: LLP - Hetero. Loans 2017q1 2018q1 2019q1 2020q1 2021q1 2022q1 date 2020 Q4 2021 Q1 2022 Q1 snaoL .oreteH - PLL naeM 800. 600. 400. 200. 0 200.- 2017q1 2018q1 2019q1 2020q1 2021q1 2022q1 date 2020 Q4 2021 Q1 2022 Q1 1

Figure OA.2: Number of CECL-related Job Postings for Delay Banks This figure plots CECL-related and informational job postings on LinkUp by banks that delayed CECL adoption. A CECL-related job is defined if job descriptions contain one of the terms “CECL,” “Current Expected Credit Losses,” “ASU 2016-13,” “ASC 326,” “Topic 326,” and “Financial Instrument(s) Credit Loss(es).” Panel A plots the total number of CECL-related job postings by CECL banks, Delay banks, and ILM banks, Panel B plots the total number of CECL-related job postings by Delay banks and ILM banks. Panel A: Total CECL Job Postings by All Banks gnitsoP boJ LCEC latoT 006 004 002 0 2017q1 2018q1 2019q1 2020q1 2021q1 2022q1 date CECL Delayed ILM Panel B: Total CECL Job Postings by Delay and ILM Banks gnitsoP boJ LCEC latoT 01 8 6 4 2 0 2017q1 2018q1 2019q1 2020q1 2021q1 2022q1 date Delayed ILM 2

Figure OA.3: Average PD Rates for All, Existing, and New Borrowers Thisfigureplotstheaverageprobabilityofdefault(PD)ratingsforFRY-14QreportingbanksthatadoptedCECL on January 1, 2020 relative to foreign banks that adopted IFRS 9 in 2018. Panel A reports average default rates forallloans. PanelBreportsaveragePDsforallexistingloans. PanelCreportsaveragePDsfornewloansdefined as the first loan by a given borrower with a given lender. Panel A: All Borrowers Panel B: Existing Borrowers Panel C: New Borrowers 3

Figure OA.4: Average PDs for High and Low Default Risk Borrowers Thisfigureplotstheaverageprobabilityofdefault(PD)ratingsforFRY-14QreportingbanksthatadoptedCECL on January 1, 2020 relative to foreign banks that adopted IFRS 9 in 2018. Panel A reports average default rates for borrowers with high PD rates (high default risk borrowers). Panel B reports average PD rates for borrowers with low PD rates (low default risk borrowers). Panel A: Mean PD Rates – High Default Risk Borrowers Panel B: Mean PD Rates – Low Default Risk Borrowers 4

Figure OA.5: Size Distributions of CECL Banks Purchasing External Models ThisfigurecomparesthehistogramsofbanksizesbetweenallCECLbanksandCECLbanksrelyonexternalmodeling products. Size is measured using lnAssets. We identify banks’ purchase of external modeling products by running textual analyses on banks’ annual reports (10-Ks). In specific, we search for case-insensitive keywords including“Moody’sAnalytics,”“ImpairmentStudio,”“ValuCast,”and“Valuant,”aswellasmentioningof“Moody’s” and either one of “analytic,” “forecast,” “scenario,” and “baseline” in the same sentence. A bank is regarded as a purchaser of external models if any of the above criteria is met in the post-period. 52 02 51 01 5 0 6 8 10 12 14 16 Size CECL Banks External Purchases 5

Figure OA.6: Loan Loss Provisioning around the Financial Crisis This figure compares the average loan loss provisioning to beginning total loans of hypothetical groups of banks that would have been subject to CECL vs. banks that would have been exempt from CECL around the financial crisis period (2005–2010) had CECL been implemented then. Following the implementation of ASU 2016-13, we define CECL banks as public banks except for smaller reporting companies and ILM banks as smaller reporting companies and private banks as of 2007 Q4. We assume that the hypothetical adoption date is January 1, 2008, the onset of the financial crisis. snaoL llA sisirC laicnaniF eht dnuora PLL naeM 800. 600. 400. 200. 0 200.- 2005q1 2006q3 2008q1 2009q3 2011q1 date CECL ILM 6

Figure OA.7: Loan Loss Provisioning by Foreign Banks This figure plots the average loan loss provisioning to beginning total loans by U.S. CECL banks (CECL), U.S. intermediate holding companies of foreign banks that adopted IFRS 9 in 2018 (Foreign), and ILM banks (ILM). Panel A reports LLPs without the day-1 impact for Q1 2020 for total loans. Panel B reports LLPs with the day-1 impact for Q1 2020 for total loans. Panel A: LLP without Day1 - All Loans 1yaD o/w snaoL llA - PLL 800. 600. 400. 200. 0 200.- 2017q1 2018q1 2019q1 2020q1 2021q1 2022q1 date CECL Foreign ILM Panel B: LLP with Day1 - All Loans 1yaD htiw snaoL llA - PLL 800. 600. 400. 200. 0 200.- 2017q1 2018q1 2019q1 2020q1 2021q1 2022q1 date CECL Foreign ILM 7

Figure OA.8: Time Trends of Job Postings This figure plots the number of job openings reported by the Bureau of Labor Statistics (left axis in thousands) and the number of job postings in LinkUp (right axis in thousands). Panel A plots the LinkUp numbers for all industries, and Panel B plots the LinkUp numbers for banks only. Panel A: BLS Openings vs. LinkUp Postings – All 00021 00001 0008 0006 0004 0052 0002 0051 0001 005 2017m1 2018m7 2020m1 2021m7 BLS Job Openings Linkup Job Postings - All Panel B: BLS Openings vs. LinkUp Postings – Banks 00021 00001 0008 0006 0004 07 06 05 04 03 02 2017m1 2018m7 2020m1 2021m7 BLS Job Openings Linkup Job Postings - Banks 8

Table OA.1: Understanding Banks’ Delayed CECL Adoption This table provides various descriptive analyses of banks that delayed CECL adoption under the Coronavirus Aid, Relief, and Economic Security (CARES) Act exemption. Panel A reports the dates of delayed adoption. Panel B compares summary statistics of banks that adopted CECL as of January 1, 2020, to banks that delayed adoption. In Panel B, ∆NPL is in percentages. Panel C reports the estimation of a determinants model predicting the delay of CECL adoption. Delay equals one if the bank delayed CECL adoption under the CARES Act and zero if the bank adopts CECL as of January 1, 2020. CECL Est. equals one if the bank provides an estimation of day 1 adoption effects in their 10-K prior to 2020 and zero otherwise. Homog% is the percentage of homogeneous loan types divided by total loans. Hetero% is the percentage of heterogeneous loan types divided by total loans. All other variables are defined in Appendix A of the paper. *, **, and *** indicate statistical significance of the coefficients at the 10%, 5%, and 1% levels, respectively. Panel A: Adoption Date by Delay Banks Adoption Date No. of Banks 2020 Q4 15 2021 Q1 18 2022 Q1 7 Merged 2 Total 42 Panel B: Descriptive Statistics (1) (2) (3) (4) (5) Delay CECL Diff. t-test Mean P50 Mean P50 t-stat Size 8.534 8.516 9.962 9.605 -5.81*** EBLLP 0.006 0.006 0.009 0.006 -2.14** Deposit 0.787 0.804 0.744 0.768 2.12** CapRatio 0.119 0.117 0.127 0.124 -1.23 ∆NPL -0.019 -0.008 0.005 0.001 -0.98 Homog% 0.401 0.388 0.372 0.348 0.81 Hetero% 0.579 0.579 0.542 0.574 0.99 CECL Est. 0.744 1.000 0.900 1.000 -2.60*** Obs. 39 150 9

Table OA.1: Understanding Banks Delayed CECL Adoption, continued Panel C: Determinants of Delaying CECL Adoption (1) (2) (3) Dep Var. Delay Delay Delay Model LPM Logit Probit CECL Est. -0.240*** -1.389** -0.799** (0.086) (0.622) (0.366) Size -0.117*** -1.861*** -1.053*** (0.021) (0.404) (0.222) EBLLP -6.451 -181.516* -97.211* (5.294) (108.149) (57.869) ∆NPL -23.686 -144.352 -92.815 (21.028) (171.925) (103.598) Deposit -0.485 -0.608 -0.136 (0.334) (3.479) (2.123) CapRatio -2.036** -1.449 -1.030 (0.835) (11.592) (6.612) Homog% 0.063 -0.134 -0.071 (0.142) (1.184) (0.705) Constant 2.192*** 18.381*** 10.234*** (0.435) (4.962) (2.815) Observations 189 189 189 Adjusted R-squared 0.227 Pseudo R-squared 0.324 0.325 10

Table OA.2: Timeliness of Loan Loss Provisioning, Various Fixed Effects ThistablereportstheresultsofestimatingthetimelinessofLLPsusingseveralvariationsofEquation1. Thedependentvariableincolumns(1)–(5)isLLPs forallloans. Thedependentvariableincolumns(6)–(10)isLLPswithday-1impactforallloans. TreatequalsoneforbanksthatadoptedCECLonJanuary 1, 2020 and zero for banks that did not adopt CECL as of December 31, 2021. Post equals one for bank-quarters after 2020 and zero otherwise. ∆NPL is the change in non-performing loans divided by the beginning total loans. Standard errors reported in parentheses are clustered by bank. All variables are defined in Appendix A. *, **, and *** indicate statistical significance of the coefficients at the 10%, 5%, and 1% levels, respectively. (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) Dep. Var. LLP LLP (w/Day1) t t Treat×Post×∆NPL 0.412*** 0.426*** 0.370** 0.479*** 0.320** 0.617*** 0.637*** 0.575*** 0.713*** 0.512*** t+ (0.146) (0.144) (0.144) (0.135) (0.125) (0.170) (0.162) (0.163) (0.156) (0.146) Treat×Post×∆NPL 0.279*** 0.271*** 0.310*** 0.229*** 0.443*** 0.432*** 0.475*** 0.350*** t (0.080) (0.076) (0.078) (0.073) (0.099) (0.096) (0.097) (0.097) Treat×Post×∆NPL 0.170** 0.110 0.220** -0.004 0.174 0.114 0.241** -0.022 t− (0.086) (0.099) (0.088) (0.082) (0.109) (0.111) (0.114) (0.099) Treat×∆NPL 0.067 0.076 0.086 0.043 0.033 0.067 0.076 0.087 0.039 0.028 t+ (0.061) (0.075) (0.063) (0.039) (0.037) (0.061) (0.075) (0.062) (0.041) (0.037) Treat×∆NPL 0.020 0.044 0.048 0.031 0.020 0.045 0.053 0.031 t (0.025) (0.028) (0.030) (0.026) (0.025) (0.029) (0.035) (0.031) Treat×∆NPL -0.038 -0.019 -0.057 -0.049 -0.038 -0.019 -0.056 -0.045 t− (0.049) (0.054) (0.054) (0.045) (0.049) (0.055) (0.059) (0.048) Post×∆NPL 0.129* 0.133** 0.147** 0.188*** -0.007 0.125* 0.129** 0.145** 0.179*** -0.066 t+ (0.067) (0.062) (0.060) (0.066) (0.077) (0.068) (0.064) (0.060) (0.068) (0.083) Post×∆NPL 0.050 0.050 0.069 -0.028 0.044 0.044 0.059 -0.035 t (0.042) (0.044) (0.046) (0.051) (0.046) (0.049) (0.051) (0.060) Post×∆NPL 0.051 0.044 0.071 0.068 0.051 0.044 0.067 0.092* t− (0.044) (0.045) (0.045) (0.045) (0.045) (0.045) (0.047) (0.052) Observations 4,902 4,864 4,864 4,863 4,863 4,902 4,864 4,864 4,863 4,863 FirmFE No No No Yes Yes No No No Yes Yes Year-QuarterFE No No No No Yes No No No No Yes Controls No No Yes Yes Yes No No Yes Yes Yes Adj. OverallR-squared 0.052 0.084 0.151 0.464 0.576 0.072 0.109 0.165 0.425 0.546 Adj. WithinR-squared 0.154 0.048 0.183 0.064 11

Table OA.3: Reflection of Local Economic Conditions, Various Fixed Effects This table reports the results of estimating the incorporation of local economic conditions in LLPs using several variations of Equation 2. The dependent variable in columns (1)–(5) is LLPs for all loans. The dependent variable in columns (6)–(10) is LLPs with day-1 impact for all loans. Treat equals one for banks that adopted CECL on January 1, 2020 and zero for banks that did not adopt CECL as of December 31, 2021. Post equals one for bank-quarters after2020andzerootherwise. ∆CoIndexisthechangeintheweightedaverageofthestate-levelcoincidentindexbasedonbanks’depositsharesindifferent states. Standarderrorsreportedinparenthesesareclusteredbybank. AllvariablesaredefinedinAppendixA. *,**,and***indicatestatisticalsignificance of the coefficients at the 10%, 5%, and 1% levels, respectively. (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) Dep. Var. LLP LLP (w/Day1) t t Treat×Post×∆CoIndex -0.014*** -0.035*** -0.033*** -0.034*** -0.035*** -0.036*** -0.064*** -0.062*** -0.064*** -0.065*** t+ (0.004) (0.005) (0.005) (0.005) (0.005) (0.005) (0.007) (0.007) (0.008) (0.007) Treat×Post×∆CoIndex -0.028* -0.024 -0.019** -0.016* -0.034** -0.031* -0.022** -0.016 t (0.017) (0.016) (0.009) (0.009) (0.017) (0.016) (0.010) (0.011) Treat×Post×∆CoIndex -0.059* -0.037 -0.018 -0.021* -0.064** -0.040* -0.014 -0.015 t− (0.032) (0.024) (0.013) (0.012) (0.032) (0.024) (0.014) (0.014) Treat×∆CoIndex 0.001 0.000 0.001 0.000 0.001 0.001 0.000 0.001 0.001 0.001 t+ (0.001) (0.001) (0.001) (0.001) (0.001) (0.001) (0.001) (0.001) (0.001) (0.001) Treat×∆CoIndex 0.011 0.007 0.002 -0.001 0.011 0.008 -0.001 -0.008 t (0.016) (0.016) (0.008) (0.009) (0.016) (0.016) (0.009) (0.010) Treat×∆CoIndex 0.047 0.025 0.005 0.008 0.047 0.024 -0.003 -0.003 t− (0.032) (0.025) (0.013) (0.012) (0.032) (0.025) (0.014) (0.014) Post×∆CoIndex 0.003* -0.007*** -0.007*** -0.006*** 0.035*** 0.003 -0.008*** -0.008*** -0.006** 0.064*** t+ (0.002) (0.002) (0.002) (0.002) (0.007) (0.002) (0.003) (0.003) (0.003) (0.012) Post×∆CoIndex -0.002 -0.000 0.001 0.009 -0.002 0.000 0.001 0.007 t (0.007) (0.008) (0.007) (0.008) (0.007) (0.009) (0.008) (0.010) Post×∆CoIndex 0.016 0.008 0.000 0.020** 0.016 0.008 -0.002 0.013 t− (0.019) (0.013) (0.008) (0.010) (0.019) (0.013) (0.010) (0.011) Observations 4,774 4,739 4,739 4,738 4,738 4,774 4,739 4,739 4,738 4,738 FirmFE No No No Yes Yes No No No Yes Yes Year-QuarterFE No No No No Yes No No No No Yes Controls No No Yes Yes Yes No No Yes Yes Yes Adj. OverallR-squared 0.016 0.147 0.219 0.510 0.581 0.053 0.189 0.248 0.481 0.567 Adj. WithinR-squared 0.246 0.083 0.278 0.122 12

Table OA.4: External vs. Internal Modeling, Descriptives This table provides various descriptive analyses of banks that explicitly mention their purchases of external (i.e., third-party)modelingproducts. PanelAreportstheby-sizedistributionofexternalmodelpurchasesamongCECL banks. Panel B reports the estimation of a determinants model predicting CECL banks’ purchase of external models. Homog% is the percentage of homogeneous loan types divided by total loans. Enforce equals one if the bank was the target of regulator’s enforcement actions in the pre-period. Restate equals one if the bank has issued restatements in the pre-period. ICW equals one if the bank was identified was internal control weakness in the pre-period. All other variables are defined in Appendix A. *, **, and *** indicate statistical significance of the coefficients at the 10%, 5%, and 1% levels, respectively. Panel A: Distributions of External Model Purchases Size bracket No. Banks No. External Purchases Smallest (1) 30 3 2 30 7 3 30 11 4 30 10 Largest (5) 29 5 Total 149 36 13

Table OA.4: External vs. Internal Modeling, Descriptives, continued Panel B: Determinants of External Model Purchases (1) (2) (3) Dep Var. External External External Model LPM Logit Probit lnAssets 0.002 0.030 0.049 (0.032) (0.115) (0.196) EBLLP -10.033 -67.766 -116.468 (6.282) (41.408) (71.327) ∆NPL -7.136 -14.606 -40.821 (27.094) (96.681) (175.158) ALLL -0.160 -13.664 -23.948 (6.731) (33.899) (58.087) Deposit -0.419 -1.666 -2.899 (0.444) (1.668) (2.832) CapRatio 0.018 -1.364 -2.322 (1.124) (5.262) (9.135) Homog% -0.297 -1.378* -2.452* (0.212) (0.791) (1.414) Enfore 0.082 0.286 0.514 (0.082) (0.273) (0.476) Restate 0.044 0.171 0.256 (0.133) (0.426) (0.725) ICW 0.350** 0.950** 1.546** (0.142) (0.438) (0.712) Observations 149 149 149 R-squared 0.093 Pseudo R-squared 0.094 0.094 14

Table OA.5: External vs. Internal Modelling, Regressions This table first repeats Table 2, Table 3, and Table 4 by comparing CECL banks relying on internal vs. external models with ILM banks, then repeats the cross-sectional analyses Table 6, Table 7, and Table 8 by excluding CECL banks who rely on external models. Treat equals one for banks that adopted CECL on January 1, 2020 and zero for banks that did not adopt CECL as of December 31, 2021. Post equals one for bank-quarters after 2020 and zero otherwise. ∆NPL is the change in non-performing loans divided by the beginning total loans. Standard errors reported in parentheses are clustered by bank. All variables are defined in Appendix A. *, **, and *** indicate statistical significance of the coefficients at the 10%, 5%, and 1% levels, respectively. Panel A: Timeliness of Loan Loss Provisioning (1) (2) (3) (4) (5) (6) (7) (8) Dep. Var. LLP LLP (w/ Day 1) LLP - Homog. LLP - Hetero. t t t t CECL Bank Split. Modeling Internal External Internal External Internal External Internal External Treat×Post×∆NPL 0.331** 0.440** 0.501*** 0.796*** -0.234 0.214 0.546*** 0.565** t+ (0.136) (0.182) (0.162) (0.219) (0.495) (0.450) (0.160) (0.267) Treat×Post×∆NPL 0.248*** 0.292* 0.365*** 0.481** 0.195 0.659** 0.320 0.445 t (0.068) (0.167) (0.103) (0.189) (0.207) (0.321) (0.204) (0.279) Treat×Post×∆NPL -0.019 0.172 -0.016 0.095 0.414 0.326 -0.133 0.101 t− (0.090) (0.148) (0.110) (0.211) (0.301) (0.351) (0.133) (0.217) Observations 4,209 2,884 4,209 2,884 3,462 2,185 3,460 2,180 Firm FE Yes Yes Yes Yes Yes Yes Yes Yes Year-Quarter FE Yes Yes Yes Yes Yes Yes Yes Yes Controls Yes Yes Yes Yes Yes Yes Yes Yes Adj. Overall R-squared 0.593 0.503 0.546 0.499 0.592 0.651 0.371 0.351 Adj. Within R-squared 0.049 0.054 0.059 0.097 0.012 0.116 0.065 0.074 15

Table OA.5: External vs. Internal Modelling, Regressions, continued Panel B: CECL-induced Information Production: Timeliness, excl. External Modeling (1) (2) (3) (4) (5) (6) Dep. Var. LLP LLP (w/ Day 1) t t CECL Bank Split. CECL Jobs Low High High Low High High Bank Size All All Large All All Large Treat×Post×∆NPL 0.403** 0.481 0.651 0.581*** 0.846** 1.159** t+ (0.181) (0.304) (0.426) (0.215) (0.371) (0.484) Treat×Post×∆NPL 0.227*** 0.428*** 0.518*** 0.361** 0.579*** 0.707*** t (0.067) (0.101) (0.137) (0.143) (0.121) (0.166) Treat×Post×∆NPL -0.024 0.062 0.101 0.017 0.088 0.151 t− (0.109) (0.185) (0.257) (0.130) (0.218) (0.320) Observations 3,340 2,811 2,687 3,340 2,811 2,687 Firm FE Yes Yes Yes Yes Yes Yes Year-Quarter FE Yes Yes Yes Yes Yes Yes Controls Yes Yes Yes Yes Yes Yes Adj. Overall R-squared 0.557 0.605 0.610 0.531 0.542 0.549 Adj. Within R-squared 0.061 0.058 0.067 0.077 0.066 0.081 16

Table OA.5: External vs. Internal Modelling, Regressions, continued Panel C: Reflection of Local Economic Conditions in Provisions (1) (2) (3) (4) (5) (6) (7) (8) Dep. Var. LLP LLP (w/ Day 1) LLP - Homog. LLP - Hetero. t t t t CECL Bank Split. Modeling Internal External Internal External Internal External Internal External Treat×Post×∆CoIndex -0.032*** -0.047*** -0.061*** -0.081*** -0.019** -0.013 -0.025*** -0.050*** t+ (0.005) (0.012) (0.008) (0.016) (0.009) (0.009) (0.007) (0.013) Treat×Post×∆CoIndex -0.015 -0.016 -0.013 -0.023 -0.005 -0.010 -0.026** -0.027 t (0.009) (0.019) (0.011) (0.017) (0.006) (0.008) (0.010) (0.019) Treat×Post×∆CoIndex -0.022* -0.018 -0.014 -0.021 -0.006 -0.012 -0.027** -0.020 t− (0.012) (0.025) (0.014) (0.022) (0.009) (0.013) (0.012) (0.024) Observations 4,107 2,683 4,107 2,683 3,310 1,938 3,307 1,932 Firm FE Yes Yes Yes Yes Yes Yes Yes Yes Year-Quarter FE Yes Yes Yes Yes Yes Yes Yes Yes Controls Yes Yes Yes Yes Yes Yes Yes Yes Adj. Overall R-squared 0.595 0.505 0.563 0.524 0.572 0.602 0.372 0.391 Adj. Within R-squared 0.078 0.096 0.111 0.172 0.022 0.048 0.045 0.121 17

Table OA.5: External vs. Internal Modelling, Regressions, continued Panel D: CECL-induced Information Production: Local Economic Conditions, excl. External Modeling (1) (2) (3) (4) (5) (6) Dep. Var. LLP LLP (w/ Day 1) t t CECL Bank Split. CECL Jobs Low High High Low High High Bank Size All All Large All All Large Treat×Post×∆CoIndex -0.023*** -0.042*** -0.046*** -0.045*** -0.075*** -0.084*** t+ (0.004) (0.009) (0.011) (0.007) (0.015) (0.018) Treat×Post×∆CoIndex -0.023** -0.025 -0.029 -0.027** -0.026 -0.025 t (0.011) (0.016) (0.019) (0.013) (0.020) (0.024) Treat×Post×∆CoIndex -0.028** -0.021 -0.025 -0.029* -0.006 0.003 t− (0.014) (0.027) (0.034) (0.015) (0.030) (0.037) Observations 3,194 2,672 2,538 3,194 2,672 2,538 Firm FE Yes Yes Yes Yes Yes Yes Year-Quarter FE Yes Yes Yes Yes Yes Yes Controls Yes Yes Yes Yes Yes Yes Adj. Overall R-squared 0.540 0.612 0.617 0.525 0.562 0.568 Adj. Within R-squared 0.0461 0.116 0.127 0.0816 0.139 0.153 18

Table OA.5: External vs. Internal Modelling, Regressions, continued Panel E: LLP-related Disclosures (1) (2) (3) (4) (5) (6) Dep. Var. LLP Disc. LLP Disc. - Fwd. LLP Disc. - Quant. CECL Bank Split. Modeling Internal External Internal External Internal External Treat×Post 0.117*** 0.144*** 0.183*** 0.246*** 0.070* 0.124* (0.028) (0.043) (0.036) (0.060) (0.042) (0.066) Observations 669 302 669 302 669 302 Firm FE Yes Yes Yes Yes Yes Yes Year FE Yes Yes Yes Yes Yes Yes Controls Yes Yes Yes Yes Yes Yes Adj. Overall R-squared 0.957 0.873 0.940 0.828 0.913 0.801 Adj. Within R-squared 0.052 0.071 0.071 0.106 0.011 0.016 19

Table OA.5: External vs. Internal Modelling, Regressions, continued Panel F: CECL-induced Information Production: LLP Disclosures, excl. External Modeling (1) (2) (3) (4) (5) (6) (7) (8) (9) Dep. Var. LLP Disc. LLP Disc. - Fwd. LLP Disc. - Quant. CECL Bank Split. CECL Jobs Low High High Low High High Low High High Bank Size All All Large All All Large All All Large Treat×Post 0.119*** 0.134*** 0.172*** 0.186*** 0.197*** 0.236*** 0.083* 0.074 0.108* (0.030) (0.041) (0.037) (0.040) (0.051) (0.051) (0.047) (0.062) (0.063) Observations 409 294 254 409 294 254 409 294 254 Firm FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Year FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Controls Yes Yes Yes Yes Yes Yes Yes Yes Yes Adj. Overall R-squared 0.961 0.967 0.977 0.926 0.955 0.961 0.917 0.924 0.932 Adj. Within R-squared 0.096 0.078 0.143 0.112 0.154 0.205 0.020 0.006 -0.000 20

Table OA.6: Changes in Loan Supplies This table reports effects of CECL adoption on banks’ loan supplies. Dependent variables in columns (1) - (6) are the changes in loans (scaled by lagged assets) for all loans, construction loans, commercial real estate loans, residential real estate loans, C&I loans, and consumer loans, respectively. Dependent variable in column (7) is the change in total deposits (scaled by lagged assets). All control variables are defined in Appendix A. *, **, and *** indicate statistical significance of the coefficients at the 10%, 5%, and 1% levels, respectively. (1) (2) (3) (4) (5) (6) (7) ∆Loan ∆Loan ∆Loan ∆Loan ∆Loan ∆Loan ∆Deposit VARIABLES All Construction Comm. RE Resi. RE C&I Consumer Treat×Post -0.001 -0.000 -0.001 0.002 -0.003 -0.001 -0.002 (0.005) (0.000) (0.001) (0.002) (0.002) (0.002) (0.006) Observations 637 557 595 616 637 637 637 Firm FE Yes Yes Yes Yes Yes Yes Yes Year-Quarter FE Yes Yes Yes Yes Yes Yes Yes Controls Yes Yes Yes Yes Yes Yes Yes Adj. Overall R-squared 0.287 0.0142 0.0652 0.0775 0.268 0.176 0.327 Adj. Within R-squared 0.047 0.018 0.056 0.048 0.011 0.010 0.096 21

Table OA.7: Weakness in Internal Control System, Enforcement This table repeats the cross-sectional analyses Table 6, Table 7, and Table 8 by comparing CECL banks with and without (severe) enforcement records to ILM banks. Treat equals one for banks that adopted CECL on January 1, 2020 and zero for banks that did not adopt CECL as of December 31, 2021. Post equals one for bank-quarters after 2020 and zero otherwise. ∆NPL is the change in non-performing loans divided by the beginning total loans. Standard errorsreportedinparenthesesareclusteredbybank. AllvariablesaredefinedinAppendixA. *,**,and***indicatestatisticalsignificanceofthecoefficients at the 10%, 5%, and 1% levels, respectively. Panel A: Timeliness of Loan Loss Provisioning (1) (2) (3) (4) (5) (6) Dep. Var. LLP LLP (w/ Day 1) t t CECL Bank Split. Enforcement No Yes Severe No Yes Severe Treat×Post×∆NPL 0.279** 0.470** -0.161 0.349** 0.833*** 0.097 t+ (0.137) (0.194) (0.312) (0.159) (0.254) (0.449) Treat×Post×∆NPL 0.185** 0.307*** 0.332* 0.215** 0.568*** 0.516*** t (0.085) (0.101) (0.185) (0.108) (0.125) (0.187) Treat×Post×∆NPL -0.033 0.044 -0.068 -0.051 -0.023 -0.064 t− (0.114) (0.098) (0.177) (0.137) (0.135) (0.202) Observations 3,588 3,511 2,431 3,588 3,511 2,431 Firm FE Yes Yes Yes Yes Yes Yes Year-Quarter FE Yes Yes Yes Yes Yes Yes Controls Yes Yes Yes Yes Yes Yes Adj. Overall R-squared 0.596 0.523 0.544 0.544 0.523 0.532 Adj. Within R-squared 0.033 0.077 0.046 0.033 0.137 0.074 22

Table OA.7: Weakness in Internal Control System, Enforcement, continued Panel B: Reflection of Local Economic Conditions in Provisions (1) (2) (3) (4) (5) (6) Dep. Var. LLP LLP (w/ Day 1) t t CECL Bank Split. Enforcement No Yes Severe No Yes Severe Treat×Post×∆CoIndex -0.034*** -0.035*** -0.061*** -0.059*** -0.071*** -0.119*** t+ (0.005) (0.007) (0.016) (0.008) (0.011) (0.023) Treat×Post×∆CoIndex -0.023** -0.008 -0.037 -0.019 -0.015 -0.042 t (0.011) (0.010) (0.034) (0.013) (0.011) (0.032) Treat×Post×∆CoIndex -0.030** -0.011 -0.052 -0.016 -0.013 -0.058 t− (0.014) (0.015) (0.044) (0.016) (0.016) (0.039) Observations 3,460 3,339 2,265 3,460 3,339 2,265 Firm FE Yes Yes Yes Yes Yes Yes Year-Quarter FE Yes Yes Yes Yes Yes Yes Controls Yes Yes Yes Yes Yes Yes Adj. Overall R-squared 0.604 0.511 0.540 0.569 0.521 0.547 Adj. Within R-squared 0.078 0.094 0.079 0.106 0.161 0.142 23

Table OA.7: Weakness in Internal Control System, Enforcement, continued Panel C: LLP-related Disclosures (1) (2) (3) (4) (5) (6) (7) (8) (9) Dep. Var. LLP Disc. LLP Disc. - Fwd. LLP Disc. - Quant. CECL Bank Split. Enforcement No Yes Severe No Yes Severe No Yes Severe Treat×Post 0.113*** 0.133*** 0.227** 0.189*** 0.215*** 0.275* 0.086** 0.080 0.202 (0.029) (0.036) (0.097) (0.038) (0.047) (0.141) (0.043) (0.054) (0.156) Observations 499 475 177 499 475 177 499 475 177 Firm FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Year FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Controls Yes Yes Yes Yes Yes Yes Yes Yes Yes Adj. Overall R-squared 0.918 0.946 0.926 0.876 0.924 0.898 0.876 0.886 0.845 Adj. Within R-squared 0.076 0.063 0.079 0.094 0.083 0.070 0.027 0.004 -0.011 24

Table OA.8: Weakness in Internal Control System, ICW Thistablerepeatsthecross-sectionalanalysesTable6,Table7,andTable8bycomparingCECLbankswithandwithout(LLP)internalcontrolweaknessto ILM banks. Treat equals one for banks that adopted CECL on January 1, 2020 and zero for banks that did not adopt CECL as of December 31, 2021. Post equals one for bank-quarters after 2020 and zero otherwise. ∆NPL is the change in non-performing loans divided by the beginning total loans. Standard errorsreportedinparenthesesareclusteredbybank. AllvariablesaredefinedinAppendixA. *,**,and***indicatestatisticalsignificanceofthecoefficients at the 10%, 5%, and 1% levels, respectively. Panel A: Timeliness of Loan Loss Provisioning (1) (2) (3) (4) (5) (6) Dep. Var. LLP LLP (w/ Day 1) t t CECL Bank Split. ICW No Yes LLP No Yes LLP Treat×Post×∆NPL 0.323** 0.470** 0.535 0.524*** 0.664*** 0.566 t+ (0.133) (0.228) (0.333) (0.157) (0.254) (0.390) Treat×Post×∆NPL 0.269*** 0.004 0.017 0.408*** 0.048 0.016 t (0.077) (0.100) (0.086) (0.102) (0.143) (0.102) Treat×Post×∆NPL -0.032 0.272* 0.500** -0.036 0.114 0.092 t− (0.084) (0.160) (0.192) (0.100) (0.259) (0.415) Observations 4,683 2,416 2,308 4,683 2,416 2,308 Firm FE Yes Yes Yes Yes Yes Yes Year-Quarter FE Yes Yes Yes Yes Yes Yes Controls Yes Yes Yes Yes Yes Yes Adj. Overall R-squared 0.578 0.556 0.558 0.546 0.536 0.538 Adj. Within R-squared 0.048 0.050 0.051 0.065 0.072 0.053 25

Table OA.8: Weakness in Internal Control System, ICW, continued Panel B: Reflection of Local Economic Conditions in Provisions (1) (2) (3) (4) (5) (6) Dep. Var. LLP LLP (w/ Day 1) t t CECL Bank Split. ICW No Yes LLP No Yes LLP Treat×Post×∆CoIndex -0.036*** -0.017 0.010 -0.065*** -0.075*** -0.034** t+ (0.005) (0.011) (0.007) (0.007) (0.019) (0.015) Treat×Post×∆CoIndex -0.016* -0.011 -0.003 -0.015 -0.022 -0.028 t (0.009) (0.026) (0.033) (0.011) (0.025) (0.026) Treat×Post×∆CoIndex -0.025** 0.059* 0.062 -0.018 0.060* 0.014 t− (0.012) (0.030) (0.044) (0.014) (0.034) (0.073) Observations 4,568 2,231 2,118 4,568 2,231 2,118 Firm FE Yes Yes Yes Yes Yes Yes Year-Quarter FE Yes Yes Yes Yes Yes Yes Controls Yes Yes Yes Yes Yes Yes Adj. Overall R-squared 0.584 0.528 0.531 0.567 0.513 0.507 Adj. Within R-squared 0.085 0.019 0.020 0.123 0.061 0.023 26

Table OA.8: Weakness in Internal Control System, ICW, continued Panel C: LLP-related Disclosures (1) (2) (3) (4) (5) (6) (7) (8) (9) Dep. Var. LLP Disc. LLP Disc. - Fwd. LLP Disc. - Quant. CECL Bank Split. ICW No Yes LLP No Yes LLP No Yes LLP Treat×Post 0.129*** 0.063* 0.037 0.208*** 0.131** 0.096 0.087** 0.052 0.107 (0.029) (0.036) (0.067) (0.037) (0.054) (0.067) (0.042) (0.064) (0.096) Observations 801 173 143 801 173 143 801 173 143 Firm FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Year FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Controls Yes Yes Yes Yes Yes Yes Yes Yes Yes Adj. Overall R-squared 0.938 0.960 0.959 0.909 0.933 0.928 0.876 0.930 0.936 Adj. Within R-squared 0.056 0.038 0.001 0.062 0.052 0.001 0.022 -0.014 -0.017 27

Table OA.9: Weakness in Internal Control System, Restatement This table repeats the cross-sectional analyses Table 6, Table 7, and Table 8 by comparing CECL banks with and without (loan) restatement to ILM banks. Treat equals one for banks that adopted CECL on January 1, 2020 and zero for banks that did not adopt CECL as of December 31, 2021. Post equals one for bank-quarters after 2020 and zero otherwise. ∆NPL is the change in non-performing loans divided by the beginning total loans. Standard errors reported in parentheses are clustered by bank. All variables are defined in Appendix A. *, **, and *** indicate statistical significance of the coefficients at the 10%, 5%, and 1% levels, respectively. Panel A: Timeliness of Loan Loss Provisioning (1) (2) (3) (4) (5) (6) Dep. Var. LLP LLP (w/ Day 1) t t CECL Bank Split. Restatement Non-Res. Res. Res. (Loan) Non-Res. Res. Res. (Loan) Treat×Post×∆NPL 0.216** 1.969*** -0.239 0.385*** 2.564*** 1.657 t+ (0.103) (0.322) (0.674) (0.119) (0.286) (1.989) Treat×Post×∆NPL 0.236*** 0.107 0.241 0.360*** 0.213 0.766** t (0.076) (0.109) (0.349) (0.102) (0.144) (0.347) Treat×Post×∆NPL 0.013 -0.237 -0.066 -0.012 -0.144 0.474** t− (0.078) (0.433) (0.126) (0.099) (0.425) (0.230) Observations 4,647 2,452 2,308 4,647 2,452 2,308 Firm FE Yes Yes Yes Yes Yes Yes Year-Quarter FE Yes Yes Yes Yes Yes Yes Controls Yes Yes Yes Yes Yes Yes Adj. Overall R-squared 0.583 0.569 0.555 0.553 0.538 0.541 Adj. Within R-squared 0.043 0.135 0.015 0.061 0.156 0.030 28

Table OA.9: Weakness in Internal Control System, Restatement, continued Panel B: Reflection of Local Economic Conditions in Provisions (1) (2) (3) (4) (5) (6) Dep. Var. LLP LLP (w/ Day 1) t t CECL Bank Split. Restatement Non-Res. Res. Res. (Loan) Non-Res. Res. Res. (Loan) Treat×Post×∆CoIndex -0.032*** -0.085*** -0.050*** -0.062*** -0.132*** -0.103*** t+ (0.004) (0.027) (0.009) (0.007) (0.037) (0.015) Treat×Post×∆CoIndex -0.016* -0.043 -0.078* -0.016 -0.052 -0.113*** t (0.009) (0.033) (0.042) (0.010) (0.042) (0.034) Treat×Post×∆CoIndex -0.022* -0.024 -0.009 -0.017 -0.024 -0.052*** t− (0.012) (0.026) (0.021) (0.014) (0.025) (0.018) Observations 4,530 2,269 2,137 4,530 2,269 2,137 Firm FE Yes Yes Yes Yes Yes Yes Year-Quarter FE Yes Yes Yes Yes Yes Yes Controls Yes Yes Yes Yes Yes Yes Adj. Overall R-squared 0.589 0.530 0.543 0.576 0.509 0.541 Adj. Within R-squared 0.080 0.097 0.027 0.122 0.141 0.069 29

Table OA.9: Weakness in Internal Control System, Restatement, continued Panel C: LLP-related Disclosures (1) (2) (3) (4) (5) (6) (7) (8) (9) Dep. Var. LLP Disc. LLP Disc. - Fwd. LLP Disc. - Quant. CECL Bank Split. Restatement No Yes Loan No Yes Loan No Yes Loan Treat×Post 0.122*** 0.117 0.168** 0.201*** 0.208** 0.229 0.073* 0.152 0.209 (0.028) (0.083) (0.069) (0.037) (0.085) (0.148) (0.042) (0.103) (0.124) Observations 793 181 143 793 181 143 793 181 143 Firm FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Year FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Controls Yes Yes Yes Yes Yes Yes Yes Yes Yes Adj. Overall R-squared 0.944 0.913 0.948 0.914 0.893 0.895 0.885 0.877 0.906 Adj. Within R-squared 0.049 0.079 0.047 0.061 0.082 0.039 0.014 0.065 0.000 30

Table OA.10: Reporting Manipulation Incentives, EBLLP This table repeats the cross-sectional analyses Table 6, Table 7, Table 8, and Table 9 by comparing CECL banks with low, medium, and high EBLLP to ILM banks. Treat equals one for banks that adopted CECL on January 1, 2020 and zero for banks that did not adopt CECL as of December 31, 2021. Post equals one for bank-quarters after 2020 and zero otherwise. ∆NPL is the change in non-performing loans divided by the beginning total loans. Standard errorsreportedinparenthesesareclusteredbybank. AllvariablesaredefinedinAppendixA. *,**,and***indicatestatisticalsignificanceofthecoefficients at the 10%, 5%, and 1% levels, respectively. Panel A: Timeliness of Loan Loss Provisioning (1) (2) (3) (4) (5) (6) Dep. Var. LLP LLP (w/ Day 1) t t CECL Bank Split. EBLLP Low Medium High Low Medium High Treat×Post×∆NPL 0.274 0.513*** 0.405** 0.431** 0.827*** 0.581*** t+ (0.177) (0.148) (0.176) (0.214) (0.190) (0.214) Treat×Post×∆NPL 0.253** 0.295* 0.201** 0.341** 0.443** 0.384*** t (0.098) (0.169) (0.089) (0.132) (0.172) (0.141) Treat×Post×∆NPL -0.010 0.088 0.063 -0.032 0.075 0.072 t− (0.101) (0.180) (0.127) (0.124) (0.199) (0.172) Observations 3,097 3,101 3,095 3,097 3,101 3,095 Firm FE Yes Yes Yes Yes Yes Yes Year-Quarter FE Yes Yes Yes Yes Yes Yes Controls Yes Yes Yes Yes Yes Yes Adj. Overall R-squared 0.540 0.537 0.607 0.515 0.531 0.556 Adj. Within R-squared 0.044 0.062 0.049 0.058 0.100 0.063 31

Table OA.10: Reporting Manipulation Incentives, EBLLP, continued Panel B: Reflection of Local Economic Conditions in Provisions (1) (2) (3) (4) (5) (6) Dep. Var. LLP LLP (w/ Day 1) t t CECL Bank Split. EBLLP Low Medium High Low Medium High Treat×Post×∆CoIndex -0.028*** -0.037*** -0.035*** -0.050*** -0.066*** -0.074*** t+ (0.007) (0.005) (0.008) (0.010) (0.009) (0.015) Treat×Post×∆CoIndex -0.008 -0.008 -0.027** -0.012 -0.011 -0.040** t (0.017) (0.014) (0.013) (0.019) (0.015) (0.017) Treat×Post×∆CoIndex -0.007 -0.021 -0.045** 0.000 -0.013 -0.049** t− (0.021) (0.015) (0.020) (0.027) (0.018) (0.024) Observations 2,951 2,944 2,923 2,951 2,944 2,923 Firm FE Yes Yes Yes Yes Yes Yes Year-Quarter FE Yes Yes Yes Yes Yes Yes Controls Yes Yes Yes Yes Yes Yes Adj. Overall R-squared 0.530 0.535 0.614 0.523 0.554 0.576 Adj. Within R-squared 0.056 0.093 0.098 0.104 0.170 0.133 32

Table OA.10: Reporting Manipulation Incentives, EBLLP, continued Panel C: LLP-related Disclosures (1) (2) (3) (4) (5) (6) (7) (8) (9) Dep. Var. LLP Disc. LLP Disc. - Fwd. LLP Disc. - Quant. CECL Bank Split. EBLLP Low Medium High Low Medium High Low Medium High Treat×Post 0.145** 0.135*** 0.143*** 0.220*** 0.237*** 0.212*** 0.099 0.080 0.124** (0.057) (0.034) (0.033) (0.074) (0.044) (0.046) (0.092) (0.053) (0.050) Observations 330 336 348 330 336 348 330 336 348 Firm FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Year FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Controls Yes Yes Yes Yes Yes Yes Yes Yes Yes Adj. Overall R-squared 0.873 0.930 0.975 0.852 0.895 0.950 0.780 0.892 0.947 Adj. Within R-squared 0.048 0.100 0.095 0.072 0.149 0.112 0.005 0.005 0.023 33

Table OA.10: Reporting Manipulation Incentives, EBLLP, continued Panel D: Loan Default (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) (13) (14) (15) Dep. Var. Default Default Default Default Default Default Default Default Default Default Default Default Default Default Default Split. Vars. All NewBorrower ExistingBorrower HighPD LowPD EBLLP Low Medium High Low Medium High Low Medium High Low Medium High Low Medium High Treat×Post -0.044 -0.259*** -0.282* 0.432 0.137 0.172 -0.046 -0.268*** -0.298** -0.043 -0.281*** -0.294** 0.112 -0.046 -0.077 (0.100) (0.089) (0.139) (0.588) (0.599) (0.619) (0.098) (0.080) (0.131) (0.111) (0.090) (0.141) (0.175) (0.207) (0.209) Observations 323,944 419,391 561,680 99,128 194,575 336,864 309,629 405,076 547,365 295,808 391,255 533,544 112,949 208,396 350,685 Year-QuarterFE Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes BankFE Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes LoanTypeFE Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Controls Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes AdjustedR-squared 0.069 0.062 0.052 0.027 0.023 0.009 0.069 0.061 0.051 0.075 0.066 0.055 0.014 0.015 0.006 AdjustedWithinR-squared 0.066 0.058 0.049 0.023 0.020 0.007 0.066 0.058 0.048 0.071 0.062 0.051 0.012 0.013 0.005 34

Table OA.11: Reporting Manipulation Incentives, Capital Ratio Thistablerepeatsthecross-sectionalanalysesTable6, Table7, Table8, andTable9bycomparingCECLbankswithlow, medium, andhighcapitalratioto ILM banks. Treat equals one for banks that adopted CECL on January 1, 2020 and zero for banks that did not adopt CECL as of December 31, 2021. Post equals one for bank-quarters after 2020 and zero otherwise. ∆NPL is the change in non-performing loans divided by the beginning total loans. Standard errorsreportedinparenthesesareclusteredbybank. AllvariablesaredefinedinAppendixA. *,**,and***indicatestatisticalsignificanceofthecoefficients at the 10%, 5%, and 1% levels, respectively. Panel A: Timeliness of Loan Loss Provisioning (1) (2) (3) (4) (5) (6) Dep. Var. LLP LLP (w/ Day 1) t t CECL Bank Split. Capital Ratio Low Medium High Low Medium High Treat×Post×∆NPL 0.292** 0.384* 0.427** 0.533** 0.637** 0.611*** t+ (0.146) (0.215) (0.181) (0.216) (0.271) (0.205) Treat×Post×∆NPL 0.259*** 0.248* 0.281** 0.442** 0.343*** 0.430*** t (0.091) (0.127) (0.115) (0.172) (0.128) (0.144) Treat×Post×∆NPL 0.059 -0.087 0.104 0.037 -0.063 0.084 t− (0.080) (0.144) (0.159) (0.111) (0.160) (0.194) Observations 3,117 3,124 3,088 3,117 3,124 3,088 Firm FE Yes Yes Yes Yes Yes Yes Year-Quarter FE Yes Yes Yes Yes Yes Yes Controls Yes Yes Yes Yes Yes Yes Adj. Overall R-squared 0.587 0.513 0.577 0.542 0.497 0.536 Adj. Within R-squared 0.044 0.059 0.052 0.066 0.085 0.068 35

Table OA.11: Reporting Manipulation Incentives, Capital Ratio, continued Panel B: Reflection of Local Economic Conditions in Provisions (1) (2) (3) (4) (5) (6) Dep. Var. LLP LLP (w/ Day 1) t t CECL Bank Split. Capital Ratio Low Medium High Low Medium High Treat×Post×∆CoIndex -0.027*** -0.041*** -0.037*** -0.046*** -0.072*** -0.085*** t+ (0.005) (0.008) (0.007) (0.009) (0.011) (0.013) Treat×Post×∆CoIndex -0.022** -0.014 -0.017 -0.029** -0.012 -0.019 t (0.010) (0.013) (0.013) (0.011) (0.013) (0.015) Treat×Post×∆CoIndex -0.015 -0.026 -0.023 -0.017 -0.020 -0.016 t− (0.014) (0.018) (0.018) (0.016) (0.018) (0.018) Observations 2,949 2,986 2,916 2,949 2,986 2,916 Firm FE Yes Yes Yes Yes Yes Yes Year-Quarter FE Yes Yes Yes Yes Yes Yes Controls Yes Yes Yes Yes Yes Yes Adj. Overall R-squared 0.583 0.511 0.580 0.538 0.524 0.564 Adj. Within R-squared 0.066 0.094 0.086 0.086 0.160 0.146 36

Table OA.11: Reporting Manipulation Incentives, Capital Ratio, continued Panel C: LLP-related Disclosures (1) (2) (3) (4) (5) (6) (7) (8) (9) Dep. Var. LLP Disc. LLP Disc. - Fwd. LLP Disc. - Quant. CECL Bank Split. Capital Ratio Low Medium High Low Medium High Low Medium High Treat×Post 0.106*** 0.141*** 0.113*** 0.175*** 0.239*** 0.169*** 0.048 0.108* 0.092* (0.035) (0.040) (0.033) (0.043) (0.054) (0.045) (0.053) (0.063) (0.046) Observations 367 368 359 367 368 359 367 368 359 Firm FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Year FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Controls Yes Yes Yes Yes Yes Yes Yes Yes Yes Adj. Overall R-squared 0.970 0.860 0.930 0.955 0.820 0.897 0.931 0.771 0.911 Adj. Within R-squared 0.067 0.067 0.068 0.076 0.096 0.084 0.020 0.015 0.0120 37

Table OA.11: Reporting Manipulation Incentives, Capital Ratio, continued Panel D: Loan Default (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) (13) (14) (15) Dep. Var. Default Default Default Default Default Default Default Default Default Default Default Default Default Default Default Split. Vars. All NewBorrower ExistingBorrower HighPD LowPD CapitalRatio Low Medium High Low Medium High Low Medium High Low Medium High Low Medium High Treat×Post -0.336*** -0.052 -0.309* 0.207 0.843 0.221 -0.345*** -0.068 -0.324** -0.347*** -0.049 -0.334* -0.101 0.166 -0.037 (0.081) (0.145) (0.155) (0.660) (0.930) (0.633) (0.076) (0.130) (0.146) (0.082) (0.141) (0.163) (0.234) (0.184) (0.210) Observations 559,127 426,839 359,493 334,311 202,023 134,677 544,812 412,524 345,178 530,991 398,703 331,357 348,132 215,844 148,498 BankFE Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes LoanTypeFE Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Year-QuarterFE Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Controls Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Adj. OverallR-squared 0.057 0.057 0.066 0.019 0.011 0.019 0.057 0.057 0.066 0.061 0.061 0.071 0.014 0.007 0.011 Adj. WithinR-squared 0.054 0.054 0.063 0.016 0.008 0.016 0.053 0.054 0.062 0.057 0.058 0.067 0.012 0.005 0.010 38

Table OA.12: Comparison of New and Existing Borrowers Thistablereportsthedescriptivestatisticscomparingnewandexistingborrowers. Thetablepresentssummarystatisticsoftheadditionalloan-orborrowerlevel characteristics for our loan-level analyses. Columns (1) to (8) provide descriptive statistics for the full sample. Columns (9) to (14) show the mean differences for the samples of new borrowers before and after CECL. All variables are defined in Appendix A. *, **, and *** indicate statistical significance of the mean differences at the 10%, 5%, and 1% levels, respectively. (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) (13) (14) New Borrowers Before CECL After CECL Two-sample t-test Variables N Mean Std. Dev. 10th 25th Median 75th 90th N Mean N Mean Diff. p-value SizeB 21,613 17.141 2.804 14.051 15.203 16.659 18.611 21.356 9,138 17.521 12,475 16.870 0.651*** <0.001 Leverage 21,613 0.373 0.278 0.046 0.150 0.325 0.542 0.758 9,138 0.382 12,475 0.363 0.019*** <0.001 Private 21,613 0.924 0.265 1 1 1 1 1 9,138 0.902 12,475 0.938 -0.036*** <0.001 Default 21,613 0.254 5.038 0 0 0 0 0 9,138 0.263 12,475 0.257 0.006 0.9946 PD 21,613 0.021 0.041 0.002 0.005 0.011 0.019 0.035 9,138 0.016 12,475 0.024 -0.008*** <0.001 Maturity 21,577 6.700 5.426 1.255 3.167 4.964 9.600 14.759 9,131 6.524 12,446 6.841 -0.318*** <0.001 Collateral 21,613 0.953 0.213 1 1 1 1 1 9,138 0.939 12,475 0.964 -0.025*** <0.001 Guaranteed 21,613 0.403 0.491 0 0 0 1 1 9,138 0.487 12,475 0.337 0.15*** <0.001 Syndicated Loan 21,613 0.112 0.315 0 0 0 0 1 9,138 0.121 12,475 0.100 0.02*** <0.001 39

Table OA.13: Timeliness of Loan Loss Provisioning, Matched Sample ThistablerepeatsthetestsoutlinedinEquation1(PanelA)andEquation2(PanelB)ofthepaperusingcoarsened exact matching (CEM). With CEM, we coarsen the data by dividing observations into five evenly spaced bins of controlvariables(Size,EBLLP,Deposit,andCapRatio )sothatCECLadoptingandILMbankshavesimilarly t−1 weighted histograms of these variables. Then, the weights are applied in a weighted least squares regression. All variables are defined in Appendix A of the paper. *, **, and *** indicate statistical significance of the coefficients at the 10%, 5%, and 1% levels, respectively. Panel A: Timeliness of Loan Loss Provisioning (1) (2) (3) (4) Dep. Var. LLP LLP (w/ Day 1) LLP - Homog. LLP - Hetero. t t t t Treat×Post×∆NPL 0.432*** 0.617*** 0.191 0.402** t+ (0.159) (0.169) (0.369) (0.185) Treat×Post×∆NPL 0.112 0.231** 0.527*** 0.089 t (0.081) (0.111) (0.171) (0.152) Treat×Post×∆NPL -0.012 -0.058 0.497* -0.018 t− (0.100) (0.126) (0.284) (0.163) Treat×∆NPL 0.035 0.037 0.127 -0.059 t+ (0.046) (0.046) (0.130) (0.046) Treat×∆NPL 0.024 0.030 -0.070 0.007 t (0.037) (0.041) (0.101) (0.041) Treat×∆NPL -0.075 -0.060 -0.289 -0.024 t− (0.061) (0.061) (0.188) (0.070) Post×∆NPL 0.016 -0.031 0.126 -0.183 t+ (0.083) (0.084) (0.273) (0.127) Post×∆NPL 0.054 0.054 -0.108 0.103 t (0.049) (0.050) (0.080) (0.110) Post×∆NPL 0.093* 0.110** -0.102 0.106 t− (0.050) (0.056) (0.081) (0.121) ∆NPL -0.005 -0.002 0.077*** 0.040 t+ (0.018) (0.018) (0.024) (0.032) ∆NPL 0.018 0.019 0.110* 0.061** t (0.025) (0.025) (0.065) (0.031) ∆NPL 0.052 0.053 0.127** 0.067 t− (0.046) (0.046) (0.064) (0.066) Treat×Post 0.000* 0.001*** 0.000** 0.000** (0.000) (0.000) (0.000) (0.000) Observations 4,022 4,022 3,310 3,314 Bank FE Yes Yes Yes Yes Year-Quarter FE Yes Yes Yes Yes Controls Yes Yes Yes Yes Adj. Overall R-squared 0.535 0.519 0.605 0.385 Adj. Within R-squared 0.063 0.090 0.048 0.051 40

Table OA.13: Timeliness of Loan Loss Provisioning, Matched Sample, continued Panel B: Reflection of Local Economic Conditions in Provisions (1) (2) (3) (4) Dep. Var. LLP LLP (w/ Day 1) LLP - Homog. LLP - Hetero. t t t t Treat×Post×∆CoIndex -0.032*** -0.057*** -0.011*** -0.028*** t+ (0.005) (0.007) (0.004) (0.008) Treat×Post×∆CoIndex -0.019* -0.017 -0.015** -0.019 t (0.011) (0.012) (0.006) (0.013) Treat×Post×∆CoIndex -0.020 -0.011 -0.011 -0.027 t− (0.014) (0.015) (0.009) (0.023) Treat×∆CoIndex 0.002 0.002* -0.002 -0.003 t+ (0.001) (0.001) (0.001) (0.003) Treat×∆CoIndex 0.006 -0.001 0.010 0.008 t (0.010) (0.011) (0.006) (0.012) Treat×∆CoIndex 0.012 -0.001 0.008 0.019 t− (0.014) (0.015) (0.008) (0.023) Post×∆CoIndex 0.023*** 0.042*** 0.010* 0.019* t+ (0.007) (0.011) (0.005) (0.010) Post×∆CoIndex 0.018* 0.013 0.003 0.016 t (0.010) (0.010) (0.006) (0.012) Post×∆CoIndex 0.015 0.006 0.007 0.033 t− (0.010) (0.011) (0.009) (0.024) ∆CoIndex -0.002 -0.003** 0.001 -0.001 t+ (0.002) (0.002) (0.002) (0.003) ∆CoIndex -0.008 -0.000 -0.000 -0.006 t (0.010) (0.011) (0.007) (0.011) ∆CoIndex -0.008 0.003 -0.005 -0.026 t− (0.011) (0.012) (0.009) (0.024) Treat×Post 0.001*** 0.001*** 0.000*** 0.001*** (0.000) (0.000) (0.000) (0.000) Observations 3,864 3,864 3,098 3,102 Bank FE Yes Yes Yes Yes Year-Quarter FE Yes Yes Yes Yes Controls Yes Yes Yes Yes Adj. Overall R-squared 0.517 0.524 0.521 0.421 Adj. Within R-squared 0.080 0.132 0.042 0.075 41

Table OA.14: Loan Loss Provisioning around the Financial Crisis This table compares the loan loss provisioning of hypothetical groups of banks that would have been subject to CECL vs. banks that would have been exempt from CECL around the financial crisis period (2005–2010) had CECLbeenimplementedthen. Treatequalsoneforpublicbanksexceptforsmallerreportingcompaniesasof2007 Q4. Post equals one for bank-quarters after 2008 and zero otherwise. Panel A reports the results of estimating the timeliness of LLPs using Equation 1 of the paper. Panel B reports the results of estimating the incorporation of local economic conditions in LLPs using Equation 2 of the paper. All variables are defined in Appendix A of the paper. *, **, and *** indicate statistical significance of the coefficients at the 10%, 5%, and 1% levels, respectively. Panel A: Timeliness of Loan Loss Provisioning (1) (2) Dep. Var. LLP LLP t t Subsample 2006 - 2009 2005 - 2010 Treat×Post×∆NPL -0.026 -0.035 t+ (0.040) (0.036) Treat×Post×∆NPL -0.025 -0.011 t (0.042) (0.038) Treat×Post×∆NPL 0.017 0.015 t− (0.053) (0.045) Treat×∆NPL 0.038 0.026 t+ (0.033) (0.030) Treat×∆NPL 0.055 0.043 t (0.036) (0.034) Treat×∆NPL 0.059 0.033 t− (0.040) (0.039) Post×∆NPL 0.051** 0.049** t+ (0.024) (0.021) Post×∆NPL 0.037 0.019 t (0.026) (0.022) Post×∆NPL 0.084*** 0.064*** t− (0.030) (0.025) ∆NPL -0.023 -0.021 t+ (0.020) (0.018) ∆NPL 0.053** 0.054*** t (0.021) (0.019) ∆NPL 0.050** 0.059*** t− (0.021) (0.020) Treat×Post 0.000** 0.001*** (0.000) (0.000) Observations 14,173 18,263 Bank FE Yes Yes Year-Quarter FE Yes Yes Controls Yes Yes Adj. Overall R-squared 0.483 0.451 Adj. Within R-squared 0.143 0.119 42

Table OA.14: Loan Loss Provisioning around the Financial Crisis, continued Panel B: Reflection of Local Economic Conditions in Provisions (1) (2) Dep. Var. LLP LLP t t Subsample 2006 - 2009 2005 - 2010 Treat×Post×∆CoIndex 0.000 0.000 t+ (0.020) (0.016) Treat×Post×∆CoIndex 0.026* 0.050*** t (0.016) (0.014) Treat×Post×∆CoIndex 0.000 -0.002 t− (0.019) (0.017) Treat×∆CoIndex -0.003 0.004 t+ (0.019) (0.014) Treat×∆CoIndex -0.016 -0.041*** t (0.011) (0.010) Treat×∆CoIndex -0.031* -0.019 t− (0.016) (0.015) Post×∆CoIndex -0.026 -0.028* t+ (0.017) (0.015) Post×∆CoIndex -0.018 -0.011 t (0.016) (0.012) Post×∆CoIndex 0.030** 0.035** t− (0.015) (0.015) ∆CoIndex 0.038** 0.031** t+ (0.017) (0.013) ∆CoIndex 0.004 -0.001 t (0.011) (0.005) ∆CoIndex -0.011 -0.032*** t− (0.011) (0.011) Treat×Post 0.000 0.001** (0.000) (0.000) Observations 13,310 17,976 Bank FE Yes Yes Year-Quarter FE Yes Yes Controls Yes Yes Adj. Overall R-squared 0.461 0.429 Adj. Within R-squared 0.123 0.099 43

Table OA.15: Loan-level Default, Robustness Thistablereportstheresultsofestimatingthedecreaseinloan-leveldefaultusingEquation4. Treatequalsonefor FRY-14QreportingbanksthatadoptedCECLonJanuary1, 2020andzeroforFRY-14Qreportingforeignbanks that adopted IFRS 9 in 2018. Post equals one for bank-quarters after 2020 and zero otherwise. Observations start in 2018 to incorporate IFRS adoption of ECL. Panel A presents the results using industry-level fixed effects based on the industry of the borrower, and Panel B presents the results using county-level fixed effects. Panel C shows the results with the dependent variable being time-varying PD expressed in percent. Standard errors are reported in parentheses and are clustered by bank. All variables are defined in Appendix A of the paper. *, **, and *** indicate statistical significance of the coefficients at the 10%, 5%, and 1% levels, respectively. Panel A: Loan-level Default (industry fixed effects) (1) (2) (3) (4) (5) (6) (7) Dep. Var. Default Default Default Default Default Default Default Split. Vars. All Existingvs. New Privatevs. Public Highvs. Low Borrowers Borrowers PD Treat×Post -0.485*** -0.486*** -0.615*** -0.519*** -0.083 -0.568*** 0.003 (0.092) (0.094) (0.211) (0.096) (0.292) (0.114) (0.152) Size -0.152 -0.193 -0.446 -0.006 -0.591 -0.084 -0.240 t (0.265) (0.251) (0.328) (0.262) (0.515) (0.296) (0.252) EBLLP 23.862** 25.194** 10.727 24.121** 31.487 33.528*** 1.491 t (10.562) (10.795) (28.834) (10.733) (30.836) (10.543) (16.381) Deposit -0.124 -0.221 1.970 -0.098 -1.734 -0.765 1.308 t (0.603) (0.625) (1.866) (0.677) (1.989) (0.804) (0.853) CapRatio -4.230 -4.020 -10.813 -3.290 -5.411 -5.196 -2.366 t−1 (3.220) (3.110) (6.402) (2.848) (5.213) (3.393) (2.453) SizeB 0.038 0.037 0.047* 0.011 0.114** 0.008 0.076*** t (0.024) (0.024) (0.026) (0.013) (0.051) (0.019) (0.027) Leverage 0.088* 0.071 0.385*** 0.032 0.271 0.030 0.271* t (0.048) (0.049) (0.129) (0.047) (0.225) (0.048) (0.143) Private -0.308*** -0.302*** -0.706* -0.279*** -0.316*** (0.070) (0.075) (0.364) (0.079) (0.094) PD 10.269*** 10.416*** 5.519* 10.366*** 9.497* 10.474*** -10.095 t (2.067) (2.064) (2.827) (2.069) (5.485) (2.102) (12.287) Maturity 0.000 0.000 -0.000 -0.000** 0.000 0.000 0.000 t (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) Collateral 0.036 0.032 0.140 -0.013 -0.011 -0.088 0.194*** (0.044) (0.042) (0.208) (0.056) (0.082) (0.099) (0.068) Guaranteed -0.049 -0.050 -0.010 -0.007 -0.256 0.001 -0.134* (0.047) (0.048) (0.047) (0.025) (0.184) (0.035) (0.075) SyndicatedLoan -0.371*** -0.367*** -0.421** -0.136*** -1.107*** -0.244*** -0.516*** (0.122) (0.124) (0.156) (0.042) (0.321) (0.078) (0.167) NewBorrower 0.038 -0.016 0.395 -0.002 0.137 (0.059) (0.037) (0.351) (0.042) (0.129) Observations 700,356 678,065 21,976 593,195 107,123 476,641 223,660 BankFE Yes Yes Yes Yes Yes Yes Yes IndustryFE Yes Yes Yes Yes Yes Yes Yes LoanTypeFE Yes Yes Yes Yes Yes Yes Yes Year-QuarterFE Yes Yes Yes Yes Yes Yes Yes Adj. OverallR-squared 0.014 0.014 0.038 0.017 0.041 0.019 0.017 Adj. WithinR-squared 0.006 0.006 0.003 0.007 0.004 0.007 0.002 44

Table OA.15: Loan-level Default, Robustness, continued Panel B: Loan-level Default (county fixed effects) (1) (2) (3) (4) (5) (6) (7) Dep. Var. Default Default Default Default Default Default Default Split. Vars. All Existingvs. New Privatevs. Public Highvs. Low Borrowers Borrowers PD Treat×Post -0.496*** -0.499*** -0.773*** -0.535*** -0.043 -0.586*** -0.029 (0.091) (0.093) (0.194) (0.099) (0.272) (0.119) (0.143) Size -0.185 -0.231 -0.425 -0.063 -0.494 -0.128 -0.247 t (0.263) (0.250) (0.304) (0.242) (0.465) (0.282) (0.274) EBLLP 24.936** 26.897** -12.412 24.504** 38.446 33.954*** 2.950 t (11.040) (11.416) (36.189) (10.984) (32.350) (11.129) (17.329) Deposit -0.061 -0.155 0.377 -0.028 -0.547 -0.580 1.176 t (0.568) (0.592) (1.684) (0.639) (1.707) (0.761) (0.885) CapRatio -3.957 -3.891 -7.898 -3.441 -2.861 -4.984 -2.298 t−1 (3.337) (3.224) (7.668) (2.868) (6.497) (3.517) (3.028) SizeB 0.036 0.034 0.087*** 0.014 0.098* 0.011 0.057** t (0.023) (0.024) (0.028) (0.012) (0.056) (0.018) (0.022) Leverage 0.141*** 0.131*** 0.295*** 0.116** 0.048 0.095 0.274** t (0.043) (0.043) (0.107) (0.053) (0.318) (0.062) (0.130) Private -0.306*** -0.301*** -0.602 -0.267*** -0.344*** (0.073) (0.076) (0.387) (0.078) (0.112) PD 10.258*** 10.415*** 5.311* 10.197*** 9.779* 10.477*** -6.836 t (2.248) (2.252) (2.938) (2.284) (5.675) (2.265) (15.103) Maturity 0.000 0.000 -0.000** -0.000*** 0.000 0.000 0.000 t (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) Collateral 0.071 0.071 0.057 0.000 0.138 -0.010 0.200*** (0.044) (0.044) (0.191) (0.054) (0.094) (0.079) (0.066) Guaranteed -0.030 -0.029 -0.016 0.015 -0.218 0.010 -0.131 (0.051) (0.052) (0.064) (0.028) (0.216) (0.036) (0.091) SyndicatedLoan -0.379*** -0.377*** -0.476** -0.172*** -1.022*** -0.244*** -0.550*** (0.119) (0.121) (0.181) (0.043) (0.347) (0.073) (0.158) NewBorrower 0.043 -0.015 0.423 -0.009 0.199 (0.062) (0.037) (0.362) (0.040) (0.137) Observations 696,217 674,309 21,413 590,433 105,751 474,381 221,743 BankFE Yes Yes Yes Yes Yes Yes Yes CountyFE Yes Yes Yes Yes Yes Yes Yes LoanTypeFE Yes Yes Yes Yes Yes Yes Yes Year-QuarterFE Yes Yes Yes Yes Yes Yes Yes Adj. OverallR-squared 0.012 0.012 0.009 0.013 0.025 0.016 0.004 Adj. WithinR-squared 0.006 0.006 0.005 0.007 0.005 0.008 0.002 45

Table OA.15: Loan-level Default, Robustness, continued Panel C: Loan-level PD (1) (2) (3) (4) (5) (6) (7) Dep. Var. PD % PD % PD % PD % PD % PD % PD % Split. Vars. All Existingvs. New Privatevs. Public Highvs. Low Borrowers Borrowers PD Treat×Post 0.361 0.392* 0.141 0.404* 0.141 0.426 -0.060*** (0.223) (0.220) (0.320) (0.233) (0.213) (0.272) (0.015) Size -0.570 -0.313 -0.154 -0.504 -0.802 -0.759 -0.020 t (0.835) (0.748) (1.237) (0.929) (0.784) (1.070) (0.045) EBLLP -3.273 -5.860 -84.660 -7.535 35.469 -0.263 0.030 t (13.685) (12.823) (55.869) (13.503) (40.053) (21.762) (1.120) Deposit -0.563 -0.618 0.641 -0.391 -1.141 -1.477 -0.077 t (1.327) (1.344) (2.393) (1.342) (3.300) (1.483) (0.135) CapRatio 7.439 6.997 14.250 8.667 2.845 7.938 0.546 i,t−1 (7.802) (7.500) (12.554) (8.512) (5.123) (9.098) (0.474) SizeB -0.251*** -0.254*** -0.150*** -0.264*** -0.277*** -0.168*** -0.014*** t (0.044) (0.044) (0.033) (0.053) (0.038) (0.052) (0.003) Leverage 1.701*** 1.688*** 1.932*** 1.642*** 2.165*** 1.619*** 0.030* t (0.143) (0.144) (0.364) (0.149) (0.360) (0.181) (0.016) Private 0.035 0.026 0.216* -0.047 -0.004 (0.142) (0.144) (0.111) (0.194) (0.011) Maturity -0.000 -0.000 -0.000*** -0.000 0.000 -0.000 -0.000** t (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) Collateral 0.594*** 0.598*** 0.384* 0.512*** 0.506*** 0.290** 0.036*** (0.138) (0.138) (0.223) (0.153) (0.121) (0.137) (0.010) Guaranteed -0.247 -0.234 -0.703 -0.271 -0.035 -0.236 0.004 (0.233) (0.223) (0.528) (0.266) (0.092) (0.192) (0.003) Syndicated Loan 0.528*** 0.537*** 0.328 0.698*** -0.282 0.430** 0.021*** (0.148) (0.150) (0.198) (0.177) (0.200) (0.195) (0.007) New Borrower -0.410*** -0.444*** -0.225 -0.616*** -0.010** (0.129) (0.133) (0.159) (0.176) (0.004) Observations 700,410 678,116 22,293 593,242 107,168 476,687 223,723 BankFE Yes Yes Yes Yes Yes Yes Yes LoanTypeFE Yes Yes Yes Yes Yes Yes Yes Year-QuarterFE Yes Yes Yes Yes Yes Yes Yes Adj. OverallR-squared 0.060 0.060 0.087 0.055 0.081 0.043 0.248 Adj. WithinR-squared 0.037 0.037 0.034 0.030 0.050 0.015 0.098 46

Table OA.16: Loan-Level Default: Focusing on Loans issued prior to 2020 This table repeats the cross-sectional analyses Table 5 and Table 9 by limiting the sample to loans issued prior to 2020. Treat equals one for FR Y-14Q reporting banks that adopted CECL on January 1, 2020 and zero for FR Y-14Q reporting foreign banks that adopted IFRS 9 in 2018. Post equals one for bank-quarters after 2020 and zero otherwise. Observations start in 2018 to incorporate IFRS adoption of ECL. Panel A shows the results fixing loans to those issued before 2020 for the full sample. Panel B shows the results for cross-sectional tests when fixing loans to those issued before 2020. Standard errors reported in parentheses are clustered by bank. All variables are defined in Appendix A. *, **, and *** indicate statistical significance of the coefficients at the 10%, 5%, and 1% levels, respectively. Panel A: Loan-Level Default: Fixing Loans before 2020 (1) (2) (3) (4) (5) (6) (7) Dep. Var. Default Default Default Default Default Default Default Split. Vars. All Existingvs. New Privatevs. Public Highvs. Low Borrowers Borrowers PD Treat×Post -0.479*** 0.000 -0.482*** -0.497*** -0.124 -0.547*** -0.122 (0.092) (0.000) (0.097) (0.085) (0.236) (0.112) (0.141) Observations 618,884 27,291 591,593 524,818 94,066 422,843 196,041 Year-QuarterFE Yes Yes Yes Yes Yes Yes Yes BankFE Yes Yes Yes Yes Yes Yes Yes LoanTypeFE Yes Yes Yes Yes Yes Yes Yes Controls Yes Yes Yes Yes Yes Yes Yes Adj. OverallR-squared 0.008 0.006 0.009 0.009 0.011 0.011 0.004 Adj. WithinR-squared 0.007 0.005 0.007 0.008 0.005 0.009 0.002 47

Table OA.16: Loan-Level Default: Loans issued prior to 2020, continued Panel B: Loan-Level Default: Fixing Loans before 2020 (1) (2) (3) Dep. Var. Default Default Default CECL Bank Split. CECL Jobs Low High High Bank Size All All Large Treat×Post -0.397*** -0.492*** -0.458*** (0.084) (0.080) (0.108) Observations 223,880 480,089 382,666 Bank FE Yes Yes Yes Loan Type FE Yes Yes Yes Year-Quarter FE Yes Yes Yes Controls Yes Yes Yes Adj. Overall R-squared 0.012 0.007 0.005 Adj. Within R-squared 0.010 0.006 0.004 48