How Did CECL Affect Bank Lending?*

Abstract:

1 Introduction

How banks measure possible loan losses affects the size and timing of earnings because loan loss provisions reduce earnings and, therefore, capital. Because capital is a costlier funding source for banks, changes to loan loss accounting could affect lending just as a change in capital requirements might. This paper uses a required change in loan loss accounting methods to explore the relationship between loan loss allowances and lending growth.

On January 1, 2020, most large banks in the United States were required to adopt the current expected credit losses (CECL) methodology for loan loss provisioning, and all remaining banks were required to adopt it three years later. Under CECL, banks need to establish allowances sufficient to cover expected losses over the entire loan contract horizon. In comparison, the previous incurred loss model (ILM) approach required allowances only for probable losses as of the reporting date. Consequently, under CECL, loss provisions are made earlier in a loan’s life cycle, before signs of credit deterioration emerge, and the level of allowances is generally higher, depending on the loan characteristics. In addition, the CECL methodology requires provisions to reflect changes in expected future credit losses associated with the macroeconomy, leading to significant changes to provisions around turning points in the economic cycle. As a result, the effect of CECL is likely to depend on the economic environment as well. Therefore, our empirical approaches and conclusions carefully consider the economic disruption from the coronavirus disease 2019 (COVID-19) pandemic, which emerged shortly after CECL’s adoption by the largest banks.

To analyze the effect of loan loss accounting on lending, we compare 2020 CECL adopters (early adopters) with 2023 adopters (late adopters) using public regulatory data between 2016:Q1 and 2024:Q2. These bank groups are different by construction: Large public companies are early adopters, and private, generally smaller companies, are late adopters. This paper trims the panel to drop the very large and small banks to create a sample of banks that are operationally more similar. We also exclude lending associated with the Small Business Administration’s Paycheck Protection Program, introduced to support the pandemic economy. To further address concerns about potential confounding due to differential exposure to the pandemic disruptions, our approaches control for differences in local area employment conditions and exposure to regional macroeconomic trends.

To identify effects of CECL, we use variation in CECL adoption’s impact over time, across firms, and between loan products. We also use multiple approaches including differences-in-differences, instrumental variables (IV), and synthetic controls to control for various potential sources of confounding. Our conclusions are drawn from consideration of these approaches and robustness tests. While other papers have studied the effect of CECL on lending, our range of empirical approaches is more comprehensive, and our preferred instrumental variable approach ties the causal loan loss allowance effect of CECL more directly to its effect on bank capital. In addition, this preferred approach uses variation in the impact of CECL at adoption on January 1, 2020, which occurred well before the severe disruption of COVID-19 was widely anticipated.¹

Our analysis starts simple with a difference-in-differences approach to show that adoption of the CECL methodology has a different sized effect on allowances across the different loan types, both for early and late adopters. However, the estimated effect on loan growth and allowances do not consistently go in opposite directions, as might be expected given CECL’s impact on capital. Because the evidence at the loan type level is weak, we use approaches that make use of additional sources of variation in CECL’s effect on allowances.

Through use of IV based on the size of the adoption effect, we can directly assess the link between loan allowances and lending. Our preferred IV approach uses firms’ day-one CECL adoption effect as a source of variation. This approach suggests that a one percentage point increase in allowance coverage–the ratio of loan loss allowances to loans–decreases lending by approximately 0.85 percent per quarter. Extrapolating from this effect and considering the 24 basis point effect of CECL adoption on allowance coverage within our sample of banks, CECL adoption resulted in a 77 basis point decline in annual loan growth. This decline is moderately significant relative to aggregate annual loan growth, which averaged about 6.4 percent in our sample of banks. In contrast, we find no statistically significant effect of allowances on banks’ capital distributions in the form of common stock distributions or share repurchases. In combination, these estimates suggest that the shock to allowances is accommodated primarily by reduced lending.

For further evidence that regulatory capital constraints are the mechanism behind the measured effect, we test whether the lending response to allowances varies with the extent to which capital requirements are a constraint for the bank. We find that the effect of changes in allowances on loan growth is significantly stronger for the most capital constrained banks than it is for the least constrained, with the effect size decreasing by about one-third of the average for each standard deviation increase in excess capital.

We further test the robustness of our results to reasonable variations in the testing methodology, including a synthetic instrumental controls approach that is robust to the possibility that early CECL adoption and its impact are correlated with an observed trend. The magnitude of results varies somewhat, with stronger results when weighting by portfolio size and weaker results with synthetic controls, but the overall conclusion remains. Overall, our results highlight the influence that accounting changes can have on bank behavior and, consequently, on the broader economy.

2 Background on CECL

The Financial Accounting Standards Board issued the CECL methodology on June 16, 2016.² CECL went into effect as part of U.S. generally accepted accounting principles for fiscal years starting after December 15, 2019, for Securities and Exchange Commission (SEC) filers, with adoption required for private companies three years thereafter. Consequently, CECL was adopted on January 1, 2020 by most public U.S. banks and January 1, 2023 by the remaining U.S. banks. Given the many factors associated with the decision to go public or private, firms effectively had little ability to choose their adoption date in advance.

CECL replaced the incurred loss model (ILM) method, as ILM proved insufficiently responsive to the deterioration in economic conditions that occurred in the 2007–09 financial crisis. Specifically, allowances remained low under the ILM after the economy had already entered a recession and the market value of bank equity had declined significantly, suggesting that a forward-looking standard might more closely align the accounting value of capital with economic reality.

Under CECL, firms are required to forecast loan balances, charge-off rates, and recovery rates. While expectations for model sophistication vary with firm size, firms are expected to incorporate forward-looking information to a greater extent than under the ILM method. Allowances are intended to cover only “incurred" losses under the ILM but are meant to cover losses over the expected remaining lifetime of the loan under the CECL methodology.

The differing conceptual basis for allowances under CECL means that they tend to be larger in practice, particularly when economic conditions are deteriorating. As allowances grow, accounting and regulatory measures of capital decline. Thus, an increase in the allowances constrains a bank similarly to an increase in the regulatory capital requirement.³ Furthermore, CECL’s requirement to recognize losses over the expected remaining lifetime moves forward the timing of loss recognition in a way which disincentivizes lending when banks are particularly concerned about meeting capital requirements in the short run.

To help illustrate the differences between the CECL methodology and the ILM, some simple examples are provided. Figure 1 shows an example of a three-year loan that originated when the economic outlook is for steady growth over the life of the loan. During such a macroeconomic setting, suppose a bank expects a 5 percent loss. So, on day one (\(t=0\)), the allowance coverage (loan loss allowance/loan amount) is 5 percent. Ultimately, in this example, the loan does not default. The modeled allowance shrinks over time as the probability of a loan defaulting shrinks as the loan matures and bad outcomes seem less likely to materialize. The allowance at the end of year three is zero. The graph on the right in figure 1 shows the same coverage ratio but under the ILM methodology. Because the loan never defaulted or even showed signs of deterioration, there was never a loan loss allowance change.

Figure 1: Allowance Model Comparison: Good Economy, Good Loan
Assume a three-year loan. Economic forecast projects slow, steady growth. The loan does not show any signs of deterioration over the three years.

Accessible version

Figure 2 shows the same loan and economic outlook except now the loan has observable deterioration at the end of year two and ultimately defaults at the end of year three. The red dashed line indicates the actual loss. The allowance under CECL based on expectations for the first year is the same but would adjust higher when evidence of the possible default appears. For the ILM, the allowance would also increase in year two if the evidence made loan default probable. In this example, the same level is shown for CECL and the ILM. However, they could be different, depending on modeling assumptions and design. At default in year three, the same amount of provisions and reserves need to be booked to cover the loss under both methodologies.

Figure 2: Allowance Model Comparison: Good Economy, Bad Loan
Assume a three-year loan. Economic forecast projects slow, steady growth. The loan shows signs of deterioration in year two and ultimately defaults at maturity. The red dashed line indicates the loss at default.

Accessible version

Across these two examples, allowances build earlier under CECL than in the ILM. Relatedly, the average level of allowances under CECL is greater.

We turn now to the implementation of CECL by U.S. banks. Federal banking regulators passed a rule in early 2019 that would spread the adoption effect over three years.⁴ This transition period was designed to give banks additional time for their capital plans to adjust.

However, the COVID-19 pandemic disrupted the economy shortly after most public banks adopted CECL. Several emergency rulemakings responding to the pandemic affected the adoption and temporarily muted the effect of CECL on banks’ regulatory capital. The Coronavirus Aid, Relief, and Economic Security (CARES) Act, signed into law on March 27, 2020, provided banking organizations the option of deferring CECL adoption until the end of the COVID-19 national emergency. Simultaneously, banking regulators issued an interim final rule that delayed the estimated effect of CECL on regulatory capital for three years, with a two-year transition period following.⁵ Virtually all banks that chose to transition the capital effect of CECL opted for the more generous version under this emergency rulemaking. However, relatively few opted to delay CECL adoption under the CARES Act. Delaying would not provide firms much if any benefit to their regulatory capital given the more generous transition period put in place. In addition, as CECL had already come into effect, deferment would have imposed its own operational challenges associated with reverting back to old processes. Despite these challenges, other research suggests that those banks that did opt to delay CECL adoption tended to be better informed and more concerned about CECL’s adoption effect.⁶ Thus, by excluding these firms from our sample, as discussed in section 4, we are understating the extent to which banks responded to CECL adoption.

3 Related Literature

The allowance for loan losses is a contra-asset account. Therefore, it is effectively funding the loans that are expected to lose value or otherwise have reduced payments. Within the funding structure, capital is also designed to absorb losses because it is the residual claimant. There is no legal obligation to pay back holders of capital. The appropriate level of capital has been a subject of long-term debate and is outside the scope of this paper.⁷ However, compared with other sources of bank funding, capital is generally more expensive, at least from the banks’ perspective. By changing the requirement to build in an earlier loan loss buffer and possibly increasing funding costs, CECL could change lending behavior.

Previous research has documented the importance of bank capital requirements for credit supply in several different settings. All of these results are consistent with capital being a more expensive funding source. Across studies, the long-run elasticity of lending to capital requirements (cumulative percentage change in lending in response to a one percent change in capital requirements) is between -0.2 and -0.8, with much of the lending adjustment observed over the few years following the change in capital requirements. For example, several papers look at the adoption effects of capital requirements in Europe. Fraisse et al., 2020 analyze bank responses to adopting the Basel II framework in supervisory data. They find that a 1 percentage point (or roughly 12 percent) increase in capital requirements reduces lending by 2.3 to 4.5 percent. Mésonnier and Monks, 2015 estimate that banks forced to raise their core Tier 1 capital ratio by 1 percentage point following the European Banking Authority’s 2011–12 capital exercise have annualized loan growth 1.2 percent lower than unconstrained banks. Gropp et al., 2018 study the same capital exercise and find that banks subject to the exercise increase their capital ratios by 1.9 percentage points and decrease risk-weighted assets by 16 percentage points relative to the control group. Behn et al., 2016 use a quasi-experimental research design to explore how a sudden increase in a capital requirement affects lending. They find that capital increases of 0.5 percentage point are associated with a 2.1 to 3.9 percentage point decline in lending. Additional examples of empirical studies of regulatory capital and lending include Aiyar et al., 2014a; Aiyar et al., 2014b; Bridges et al., 2014; and De Jonghe et al., 2020. The range of our point estimates for the impact of allowances on loan growth is consistent with this literature. For example, our preferred estimate equates to an estimated elasticity of lending to capital requirements of about -0.6.⁸

Other papers explore the relationship between capital and lending in a dynamic model. For example, Berrospide and Edge, 2010 use a variety of methods to explore how bank capital influences lending and generally find modest impact. In particular, an increase in the capital ratio of 1 percentage point is associated with a long-run increase in annualized BHC loan growth that is only between 0.7 and 1.2 percentage points. This is an equilibrium effect: banks with higher capitalization rates tend to lend more in the next period. This is true even after trying to control for things like bank characteristics and the macroeconomic environment.⁹ Carlson et al., 2013 try to identify the relationship at the local level so as to better control for loan demand. They estimate that the elasticity of bank lending with respect to capital ratios is higher in crisis periods when capital ratios are relatively low. This suggests that the effect of capital ratio on bank lending is nonlinear. This is consistent with our results when we interact the adoption of CECL with how capital-constrained an individual firm is. Firms with tighter capital constraints have even less loan growth than other CECL adopters.

The first stage of our work closely relates to Covas and Dionis, 2021, as well as Loudis et al., 2021. These papers explore the effects of CECL on allowances and growth of different types of loans. They find a significant effect of CECL on the allowances and growth of consumer loans, but no effect of CECL adoption on the growth rates of other forms of lending. Covas and Dionis, 2021 argue that this non-result is expected as the emergency rulemakings offset the capital impact for other forms of lending, while Loudis et al., 2021 are hesitant to infer much from the impact on consumer lending. Our work expands materially on these earlier findings by also considering variability in firm-specific CECL effect, employing additional empirical strategies, and incorporating the post-pandemic experience of CECL late adopters.

Chen et al., 2025, Granja and Nagel, 2025, and Yang, 2025 also estimate the effect of CECL adoption on lending, but are more narrow in coverage than this paper. Yang, 2025 focuses on residential mortgage lending, and finds a reduction in balance sheet lending in favor of securitization for banks adopting CECL. Granja and Nagel, 2025 finds an effect of CECL on the pricing of consumer loans. Chen et al., 2025 find a significant effect of CECL on overall lending and possible downstream effect on employment in 2020. By focusing on that stress episode, they argue that the effects demonstrate that CECL has a procyclical effect. Several other studies explore the effect of CECL on lending procyclicality, but most are more hypothetical, using data before actual CECL adoption.¹⁰ In comparison to Chen et al., 2025, our work considers a longer time period and additional empirical methods to better identify the mechanism.

This paper is connected to the broader literature on CECL. For example, other papers study other effects of CECL adoption, including bank information production and financial reporting. Hu, 2024 looks at the effect of CECL on small business lending and finds that banks reduce their lending more on businesses that are farther from physical branches. This result is consistent with theories about information acquisition. Adopting CECL is associated with higher information acquisition or production costs, making distant borrowers more expensive. Relatedly, Kim et al., 2026 show that banks adopting CECL have changes in their financial reporting and operations that are consistent with better information production.

This paper is also related to research that examines the lending effects of a similar accounting standard issued by the International Accounting Standards Board (IASB) on bank lending in other countries. In 2018, under International Financial Reporting Standards (IFRS) 9, firms using IASB standards needed to use an expected credit loss (ECL) model for valuing financial assets and liabilities. Ertan, 2025 finds that banks that adopted IFRS 9 reduced lending to small and medium-sized enterprises by over 10 percent. However, the research explains these results as more due to banks’ financial reporting concerns and implementation difficulties as opposed to simply higher loan loss allowances. Bojar and Olszak, 2022 study publicly traded banks in Poland and do not find a direct effect of IFRS 9 on lending. However, there is a positive and significant relationship between IFRS 9 and capital, suggesting that banks with higher capital ratios were better able to extend more loans during the implementation period.

4 Data

Our data sample begins in 2016:Q1 and ends in 2024:Q2. As discussed in section 2, entities adopted CECL either following January 1, 2020 (“early adopters”) or January 1, 2023 (“late adopters”). This time frame means our sample has four years of data before and after CECL enactment for the early adopters and six quarters of data following enactment by the late adopters.

Because this paper analyzes bank holding companies (BHCs) as well as standalone commercial banks that are not in a BHC, the data are a combination of Call Report (commercial bank) and FR Y-9C (BHC) data. Throughout this paper, we refer to all firms as “banks.”

To ensure that changes in corporate structure, such as mergers, are not misinterpreted as large shocks, we aggregate balance sheets across entities so that they represent consistently consolidated banking businesses over time. Most commonly, one or more depository institutions have always been consolidated into the same holding company over our sample period. In this case, we include only the balance sheet for the holding company.¹¹ However, in the case of mergers that occur during our sample period, this means constructing historical “pseudo-banks" by retroactively consolidating the balance sheets of the entities that later merged.

In addition, we measure loan balances throughout our sample exclusive of loans issued through the Small Business Administration’s Paycheck Protection Program (PPP).¹² PPP loans were forgivable and would ultimately be repaid by the government. Relatedly, PPP loans did not require allowances for credit losses and had no associated risk-based capital requirements. While the PPP was a significant source of aggregate loan growth in 2020, particularly for CECL late adopters, it would not have a clear connection with firms’ capital management processes—and the channel by which it would have been affected is unclear. By excluding these loans, we remove their potentially distortionary effect from the analysis and match the practices of other researchers (e.g., Chen et al., 2025; Kim et al., 2026).

Our data sample is designed to focus on a set of banks that are relatively comparable in terms of their complexity and regulatory environment and for which a significant but not overwhelming share adopted CECL in January 2020. Toward this aim, we exclude banks with less than $1 billion or more than $100 billion in consolidated assets as of 2019:Q4. The largest banks are excluded because they are subject to heightened supervision and additional regulation, and almost all adopted CECL in 2020. As a result, they are not good controls against which to measure outcomes. Similarly, the smallest banks are largely privately held, and few were required to adopt CECL in 2020. In addition, for the smallest banks, the operational burden of CECL adoption and maintenance may represent a greater burden than the change in allowances. To simplify the analysis, from this set of 667 intermediate-sized banks, we also exclude 139 that adopted CECL in quarters other than 2020:Q1 or 2023:Q1.¹³ From the remaining 528 firms, we exclude 6 for which lending is a relatively unimportant part of the business model.¹⁴ Our remaining data filters are designed to address potential data quality and comparability issues.¹⁵ Our final sample consists of 364 firms, accounting for 9,779 firm-quarters of data. We create a further panel of 45,574 firm-loan type-quarters using the six loan types for which the regulatory reports provide tracking of both allowances and loan balances: construction (CNSTR), commercial real estate (CRE), commercial (COM), residential real estate (RRE), credit cards (CARDS), and other consumer loans (OTHCONS).¹⁶ We also winsorize variables with significant outliers (loan growth, delinquency measures) at the 1st and 99th percentiles.¹⁷

Table 1 summarizes the primary sample used in this paper, comparing the average values of the variables for the early adopters with those for the late adopters. As expected, the early adopters tend to be larger (see log assets). They also tend to have lower allowance coverage before CECL despite having similar credit quality measures (delinquency rates and nonaccrual rate). Reflecting that early adoption is primarily determined by whether the bank is public, early adopters have a much higher share repurchase rate, and somewhat higher average dividend rate–as a share of assets. However, both early and late adopters have similar loan portfolios (bottom six rows), with early adopters having a bit more exposure to commercial loans and a bit less exposure to RRE loans. Few firms within the sample have significant credit card exposure.¹⁸ Notwithstanding this overall similarity in loan portfolios, early CECL adopters’ allowances had a significantly higher adoption effect (0.38 versus 0.10 percent of loan balances). Finally, from the observation counts, we have more late (268) than early (96) adopting banks within the sample.

Table 1: Summary Statistics (Means), by CECL Adopter Type and Time Period

Source: FR Y-9C, Call Reports, FFIEC branch data, and Bureau of Labor Statistics.

Table 2 provides the distribution of the variables within the early and late adopter subgroups: standard deviation, 5th percentile, and 95th percentile. There is significant overlap in the distribution of each variable between the two adoption groups, with the smallest early adopters smaller than the largest late adopters. The last two rows provide some evidence that early and late adopters also faced similar macroeconomic environments. Branch-weighted average unemployment rates and employment growth have similar distributions across the two groups. This is particularly important to check given the relationship between CECL, the economic environment, and general loan demand.

Figure A1 in the appendix shows that our sample is geographically representative of all banks, and that early and late adopter banks’ branches have similar geographic distributions. The top U.S. map subtracts each state’s share of branches across all banks from state’s branch share within our sample. The most notable difference is that our sample has a lower share of branches in New York state, but our sample is broad enough that there is no particular regional concentration. The bottom map compares the average branch share, by state, of the early and late adopters. Most states are light red or blue, indicating that the share of bank branches between the compared groups are at most 2 percentage points different from each other. The differences across all states are not statistically significant; the p-value for the sum of absolute differences across states is 0.28.

Table 2: Summary Statistics (Distributions), by CECL Adopter Type

Source: FR Y-9C, Call Reports, FFIEC branch data, and Bureau of Labor Statistics.

5 Empirical Analysis

As shown in the simple loan example in the introduction, the CECL methodology was designed to build loan allowances earlier than under the ILM methodology, which has important implications for bank balance sheets and capital. As a contra-asset account, loan allowances offset loans on the balance sheet, effectively funding a portion of the loan portfolio. Without the credit loss provisions that build these allowances, more earnings would have been retained, increasing capital. For these reasons, the change to the credit loss accounting is comparable to an increase in required capital. Within our sample, the impact of the increase in allowances at adoption was equivalent to a 2.6 percent increase in capital.¹⁹ As established in the literature (see section 3), higher capital requirements increase a bank’s cost of funding, potentially reducing loan growth.²⁰

In summary, the expected relationships are as follows: (1) CECL will increase a bank’s loan allowance coverage (loan loss allowance / total loans), and (2) the increase in allowance coverage will raise funding costs, which will decrease loan growth.

5.1 CECL Effect, by Loan Type

We start our analysis by applying a well-established difference-in-differences methodology, measuring the change in loan growth and allowance coverage around CECL adoption. We perform this analysis separately for each of the six loan loan types \(p\) for which the regulatory reports provide allowance and loan balance data. We also include separate dummies (\(OnCECL\)) to indicate the bank-time observations where the CECL methodology is in effect for early and late adopting banks. This approach allows the effect of CECL to be different for these two cohorts. We include a set of bank-level variables as additional controls (BankControls). The bank-level control variables include three asset quality controls, three bank characteristic variables, and additional controls for local macroeconomic conditions–which may affect loan demand. Asset quality is measured by delinquency rates at 30-day, 90-day, and nonaccrual frequencies. The three bank characteristic variables are lagged by one period to avoid endogeneity issues and include log assets, leverage ratio, and return on assets to control for size, solvency, and profitability. The macroeconomic controls include branch-weighted average unemployment rates and employment growth over the previous year.²¹ In addition, the shares of each banks’ branches in each of nine regional Census Divisions (\(Region\)) in every quarter is included to control for regional economic trends. Our regression corresponds to equation (1) \[\begin{eqnarray} Y_{p,i,t} &=& \beta^{early}_{p}\,\mbox{OnCECL}_{i,t} + \beta^{late}_{p}\,\mbox{OnCECL}_{i,t} \nonumber \\ && + \Gamma_{p}\,\mbox{BankControls}_{i,t-1} + \Lambda_{p,t}\,\mbox{Region}_{i,t-1} + \delta_{p,t} + \theta_{p,i} + \varepsilon_{p,i,t}\tag{1} \end{eqnarray}\] where \(Y_{p,i,t}\) is allowance coverage or loan growth at the bank-loan type level. The \(\beta\)’s are the coefficients of interest in this specification and capture average differences that emerge between the bank groups from the time of CECL adoption.

Before running this regression, we first test an identifying assumption—that loan growth and allowance follow parallel trends for early and late adopters before early CECL adoption (2020). To test this assumption, we replace the “on CECL” dummies and quarterly fixed effects with an interaction of all quarterly dummies with an early CECL adopter dummy. Next, we run a Wald test of the joint statistical significance of the 16 quarterly interaction terms occurring in the years before early CECL adoption, 2016 to 2019. A low joint significance is a sign that the parallel trends assumption cannot be rejected, and causal claims in the difference-in-difference regression might be valid. The p-values for this test are provided in table 3, and the full set of interaction terms are plotted in the appendix (figures A2 and A3).

Table 3: p-values for Joint Significance of 2016–19 Early Adopter X Quarter Interaction Terms

None of the loan types have pre-adoption loan growth trends that are statistically significantly different for early and late CECL adopters. The pre-adoption allowance trend is only statistically significantly different for CRE loans, but the difference in trends is economically small and appears to result from a gradual decline in the relative CRE allowances of early CECL adopters leading up to 2020. (See appendix figure A3.) The consequence of this deviation from parallel trends could be a slight underestimate of the effect of CECL on CRE allowances in 2020.

Another identifying assumption for the difference-in-differences methodology is that the treatment is applied quasi-exogenously, or orthogonal to other reasons that would predict a different change in allowance or loan growth between the early and late CECL adopters. Treatment, or early adoption of CECL, was primarily based on the definition of an SEC filer. Firms choose to be filers, so the treatment is not random. However, SEC filers and non-filers appear broadly similar apart from average size differences (see tables 1 and 2), and there are no other proximate changes related to SEC filing status that would affect the outcomes we measure. Thus, it may be reasonable to consider SEC filing status, and thus the timing of CECL adoption as as quasi-exogenous, or as good as random for the purposes of identifying CECL’s effect.

Of course, the early adoption of CECL coincides with the onset of the COVID-19 pandemic–which undoubtedly had an effect on allowances and loan growth. Our identification approach only requires that the pandemic’s effect on individual banks was orthogonal to whether or not they were early adopters. However, the presence of the pandemic does change the interpretation of the CECL effect we measure. We estimate the effect of CECL on allowances and loan growth under conditions characterized by significant economic uncertainty and expected, though largely unrealized, economic stress.

As the identification considerations are generally satisfied, we proceed to estimate regression equation (1) by ordinary least squares (OLS), using standard errors robust to clustering within banks. The coefficients representing CECL’s effect on each loan type are presented as data points in figure 3, with estimates for the early adopters (i.e., \(\beta^{early}_{p}\)) in the top plot and estimates for the late adopters (i.e., \(\beta^{late}_{p}\)) in the bottom plot. The x-axis is the estimate from the allowance coverage regression, and the y-axis is the estimate from the loan growth regression. The dashed gray lines represent statistical 95 percent confidence intervals for these estimates.

Figure 3: Difference-in-Differences Estimates
These figures plot the results of running the following regression: \[\begin{eqnarray} Y_{p,i,t} &=& \beta^{early}_{p}\,\mbox{OnCECL}_{i,t} + \beta^{late}_{p}\,\mbox{OnCECL}_{i,t} + \Gamma_{p}\,\mbox{BankControls}_{i,t-1} + \Lambda_{p,t}\,\mbox{Region}_{i,t-1} + \,\delta_{p,t} + \theta_{p,i} + \varepsilon_{p,i,t}, \nonumber \end{eqnarray}\] where \(Y\) is loan growth or allowance coverage, and BankControls\(_{i,t-1}\) include three asset quality controls, three bank characteristic variables, and macroeconomic controls. In the top figure, the \(\beta^{early}_{p}\) values are plotted with the dots for each loan type \(p\). The bottom figure plots \(\beta^{late}_{p}\). The dashed gray lines represent statistical 95 percent confidence intervals for these coefficient estimates. cards = credit card loans, com = commercial and industrial loans, cre = commercial real estate loans, cnstr = construction loans, and othcons = other consumer loans.
Source: Call Report, FR Y-9C, and author calculations.

Accessible version

For the 2020 adoption plot, all six dots are to the right of the zero line for the x-axis, suggesting that, on average, allowances increased for CECL adopters after the rule adoption. The result is statistically significant for three loan types: RRE, CRE, and construction. Five of the dots are below the zero line for the y-axis, suggesting that, on average, loan growth was negative after CECL for adopters. However, this result is not statistically significant for any of the loan types.

For the 2023 adoption plot, four dots are to the right of the zero line for the x-axis, suggesting that, on average, allowances increased for those loan types after the rule adoption. But the result is only statistically significant for construction loans. Three of the dots are below the zero line for the y-axis, and none are statistically significant. This result suggests that the later CECL adoption had minimal effect on average loan growth.

This loan type-level analysis does not present a perfectly consistent story of CECL affecting loan growth through an effect on allowances.²² For example, in the 2020 adoption, construction loans have a large allowance effect for early and late adopters but a statistically zero effect on loan growth. RRE loans have a positive loan growth rate for late adopters but statistically zero effect on allowance coverage. Only commercial lending for late adopters is consistent with a loan allowance mechanism and in the opposite direction. Allowances decreased, and lending increased. However, the evidence across loan types is suggestive: Five of the coefficient estimates are in the bottom-right quadrant for 2020 adoption. The dots in the 2023 adoption plot are near the top-left and bottom-right quadrants. The weakness of the difference-in-differences results is primarily because loan composition is not strongly correlated with effect of CECL on allowances. This is why our preferred specification will use information about the firm-specific effect of CECL adoption on allowances.

5.2 Allowance-Driven Lending Effect

We turn next to specifications that leverage heterogeneity in the CECL methodology’s effect on allowances to test for evidence of an allowance-driven effect of CECL on loan growth. By restricting the impact of CECL to operate through its effect on allowances, identification is cleaner as we restrict the potential sources of confounding to those differences between banks that also relate to the effect of CECL on allowances. In addition, this approach also allows for a direct comparison to studies of the effect of changes in capital requirements on lending.

In our specifications, we regress loan growth on allowance coverage, and use instruments that capture the effect of CECL on allowance coverage. We use two instrumental variable (IV) strategies: one based on variation in CECL effect across banks and one based on variation in CECL effect across loan types. In each, we still expect an increase in allowance coverage to have a negative effect on loan growth.

IV Strategy 1: CECL’s bank-specific adoption effect

The first instrumentation strategy identifies effects from heterogeneity in the firm-specific effect of CECL adoption on allowances.²³ Specifically, a dummy “OnCECL,” indicating when the given firms’ allowances are calculated under CECL, interacts with the bank-specific effect of CECL adoption on allowances (“AdoptionEffect”). The adoption effect is calculated as the increase in allowances due to CECL on day one divided by total loans.

The bank level testing specification is \[\begin{eqnarray} \mbox{LoanGr}_{i,t} &=& \beta\,\mbox{AllowanceCov}_{i,t} + \Gamma_{p}\,\mbox{BankControls}_{i,t-1} + \Lambda_{p,t}\,\mbox{Region}_{i,t-1} \nonumber \\ &&+ \Phi_{t}\mbox{LoanShare}_{i,t}+ \theta_{i} + \varepsilon_{i,t} \qquad \quad\tag{2} \\ \mbox{AllowanceCov}_{i,t} &=& \alpha\mbox{AdoptionEffect}_{i}\times\mbox{OnCECL}_{i,t} + \Gamma_{p}\,\mbox{BankControls}_{i,t-1} + \Lambda_{p,t}\,\mbox{Region}_{i,t-1} \nonumber \\ && + \Phi^{*}_{t}\mbox{LoanShare}_{i,t} + \theta^{*}_{i} + \varepsilon^{*}_{i,t},\tag{3} \end{eqnarray}\] where the bank-level control variables are defined as before and include three asset quality controls, three bank characteristic variables, and the macroeconomic controls. The additional control \(\mbox{LoanShare}_{i,t}\) is a vector of five variables that are equal to loan shares of five of the six loan types. We allow the effect of loan composition on bank level allowances and loan growth to be time-varying to account for differences in trends across the different loan types.

IV Strategy 2: CECL’s effect on allowances by loan type

The second instrument uses the \(OnCECL_{i,t}\) dummy directly. Because CECL is a forward-looking methodology and, importantly, reflects expected credit losses through loan contracts’ maturity, the average increase of allowance coverage when CECL is adopted will likely vary by loan type. This testing strategy treats the heterogeneity in CECL’s effect across loan types along with the differences in CECL adoption dates as quasi-exogenous changes in allowances with which to assess the effect on loan growth.

The testing specification is \[\begin{eqnarray} \mbox{LoanGr}_{p,i,t} &=& \beta\,\mbox{AllowanceCov}_{p,i,t} + \Gamma_{p}\,\mbox{BankControls}_{i,t-1} + \Lambda_{p,t}\,\mbox{Region}_{i,t-1} \nonumber \\ &&+ \delta_{p,t} + \theta_{i} + \varepsilon_{p,i,t}\tag{4} \\ \mbox{AllowanceCov}_{p,i,t} &=& \alpha_{p}\,\mbox{OnCECL}_{i,t} + \Gamma_{p}\,\mbox{BankControls}_{i,t-1} + \Lambda_{p,t}\,\mbox{Region}_{i,t-1} \nonumber \\ &&+ \delta^{*}_{p,t} + \theta^{*}_{i} + \varepsilon^{*}_{p,i,t},\tag{5} \end{eqnarray}\] where the control variables are defined as before and \(p\) represents the loan type.

The two instrumental variable strategies are complementary. The bank-specific effect of CECL adoption is a significant source of variability in allowances beyond differences in the composition of loan portfolios across banks. Specifically, the distribution of banks’ loans across the six types explains only a small part of the variability of the bank-level adoption effect.²⁴ While our data do not provide the adoption effect of CECL on allowances at the bank-loan type level, we similarly note that the change in bank-loan type allowances around CECL adoption is statistically significantly predicted by the loan type even conditional on the bank-level adoption effect.

By representing different sources of variation, the two strategies have different vulnerabilities to potential sources of endogeneity. For example, if it were the case that a type of bank adopted CECL in a particularly conservative way and also grew more slowly for unrelated reasons, this would challenge the exogeneity of the instruments in our first strategy but would not affect our second strategy, which does not use this bank-specific variation.

Table 4 summarizes the results of the OLS and IV regressions specified above in equations (2) through (5). All coefficients of interest (\(\beta\)) have the expected negative sign; higher allowances are associated with lower subsequent loan growth. Also, the OLS coefficients are larger in magnitude than the IV coefficients, consistent with the expected endogeneity bias. Specifically, the financial health of a bank’s customer base is likely to drive allowances for credit losses and credit extension in opposite directions.

Table 4: OLS and IV Regression Results

This table shows the results of using the following specifications. The coefficient on allowance coverage (\(\beta\)) for the loan growth specification is reported. Results from the first set of equations are reported in columns (1) and (2) and the second set in columns (3) and (4). \[\begin{eqnarray} \mbox{LoanGr}_{i,t} &=& \beta\,\mbox{AllowanceCov}_{i,t} + \Gamma_{p}\,\mbox{BankControls}_{i,t-1} + \Lambda_{p,t}\,\mbox{Region}_{i,t-1} + \Phi_{t}\mbox{LoanShare}_{i,t} + \theta_{i} + \varepsilon_{i,t} \qquad \quad \nonumber \mbox{AllowanceCov}_{i,t} &=& \alpha\mbox{AdoptionEffect}_{i}\times\mbox{OnCECL}_{i,t} + \Gamma_{p}\,\mbox{BankControls}_{i,t-1} + \Lambda_{p,t}\,\mbox{Region}_{i,t-1} \nonumber && + \Phi^{*}_{t}\mbox{LoanShare}_{i,t} + \theta^{*}_{i} + \varepsilon^{*}_{i,t}, \nonumber \end{eqnarray}\] \[\begin{eqnarray} \mbox{LoanGr}_{p,i,t} &=& \beta\,\mbox{AllowanceCov}_{p,i,t} + \Gamma_{p}\,\mbox{BankControls}_{i,t-1} + \Lambda_{p,t}\,\mbox{Region}_{i,t-1} + \delta_{p,t} + \theta_{i} + \varepsilon_{p,i,t} \nonumber \mbox{AllowanceCov}_{p,i,t} &=& \alpha_{p}\,\mbox{OnCECL}_{i,t} + \Gamma_{p}\,\mbox{BankControls}_{i,t-1} + \Lambda_{p,t}\,\mbox{Region}_{i,t-1} + \delta^{*}_{p,t} + \theta^{*}_{i} + \varepsilon^{*}_{p,i,t}, \nonumber \end{eqnarray}\] BankControls\(_{i,t-1}\) include three asset quality controls, three bank characteristic variables, and macroeconomic controls. Asset quality is measured by delinquency rates at 30-day, 90-day, and nonaccrual frequencies. The three bank characteristic variables are lagged by one period to avoid endogeneity issues and include log assets, leverage ratio, and return on assets to control for size, solvency, and profitability. \(\Phi^{*}_{t}\mbox{LoanShare}_{i,t}\) is a vector of five variables that are equal to loan shares of five of the six loan types, and OnCECL\(_{i,t}\) is a dummy variable set to one after CECL adoption by bank \(i\).

Standard errors are in parentheses.

\(^{*}\) \(p<0.10\), \(^{**}\) \(p<0.05\), \(^{***}\) \(p<0.01\).

The bank-level instrumental variables regression finds a statistically and economically significant effect of allowances on loan growth. With respect to economic significance, a coefficient of -0.85 suggests that CECL adoption decreased annual loan growth by approximately 77 basis points for the average bank in our sample.²⁵ This is moderately significant relative to aggregate annual loan growth of banks in our sample of about 6.4 percent. At the same time, as noted above, CECL adoption reduced capital by about 2.6 percent across the aggregate of banks in our sample. Our estimates therefore suggest an impact on annual lending growth that is about 0.3 times as large as the percentage impact on capital requirements, which is within the broad range reported in the literature summarized in section 3.

In contrast to estimates using the bank-level effect of CECL, the coefficient from the loan type effect strategy is not statistically different from zero. However, we would caution about drawing strong conclusions from the loan type level estimates given the apparent weakness of the instruments.²⁶ The overall weakness of the instrument likely results from the relatively limited number of firms with the types of loans (e.g., cards and other retail) for which the average impact is very different from the average across all loans. In addition, depending on the composition and expertise in types of loans, individual banks may choose to meet stricter overall requirements in part by reducing credit availability even for loan types where the changes in treatment are minimal.²⁷ If so, this would also help to explain the smaller magnitude of coefficients at the loan type level.

5.3 Extensions

In this section, we extend our analysis in search of a richer set of results to better understand the mechanism. We look for evidence of an anticipatory effect, and test the persistence of the response to CECL adoption. We test whether more capital constrained banks behave differently. Lastly, we test whether CECL adoption affects banks’ capital distributions in addition to their lending.

5.3.1 Timing

First, we relax our assumption around when firms respond to the effect of CECL on allowances. Thus far, our specifications assume that the effect of CECL on loan growth occurs only after adoption and is constant (per unit of effect on allowances) over the observed post-adoption period. There is some reason for skepticism regarding these assumptions. Given that the CECL methodology and adoption requirements were known well beforehand, it is reasonable to expect that most banks knew the likely effect of CECL before adoption. As it may be costly for banks to quickly change their loan exposures, we might expect that banks’ lending decisions prior to CECL adoption might reflect the anticipated effect of CECL on allowances. Relatedly, as the credit supply effect of CECL becomes fully reflected across the bank’s business and loan types over time, we might expect the effect of CECL adoption on loan growth to fade to zero.

To assess the timing of CECL’s effect on lending, we run a panel regression of loan growth on each banks’ adoption effect, interacted with a dummy for the time-to-adoption \(\tau\). As shown in equation (6), the regression specification includes bank and quarter fixed effects as well. The \(\beta_{-1}\) term is removed from the equation, so that the impact of the adoption effect on lending is measured relative to the year prior to CECL’s adoption by the bank. As this regression is run over data from 2016 through 2023, effects more than four years prior to adoption are identified from only late CECL adopters, and effects following the year of adoption are identified only from early CECL adopters. \[\begin{eqnarray} \mbox{LoanGr}_{i,t} &=& \beta_{\tau,\tau\neq -1}\,\mbox{AdoptionEffect}_{i} + \delta_{t} + \theta_{i} + \varepsilon_{i,t}\tag{6} \end{eqnarray}\] The coefficients \(\beta_{\tau}\) from this regression are plotted in figure 4, with \(\beta_{-1}=0\) by construction. The lighter lines bounding the coefficients represent the 95 percent confidence interval.

Figure 4: Relationship of Loan Growth and CECL Adoption Effect, by Time to Adoption
The coefficients \(\beta_{\tau}\) from the following regression are plotted: \[\begin{eqnarray} \mbox{LoanGr}_{i,t} &=& \beta_{\tau,\tau\neq -1}\,\mbox{AdoptionEffect}_{i} + \delta_{t} + \theta_{i} + \varepsilon_{i,t} \nonumber. \end{eqnarray}\] Data sample covers the years 2016 through 2023.

Accessible version

In Figure 4, the relationship between CECL adoption effect and loan growth is stable in the years prior to CECL adoption. There is no evidence of the downward trend that one might expect to result from an anticipatory effect. Following CECL adoption, this relationship remains stable and negative. While this could suggest a longer lasting effect of CECL on loan growth, limited statistical power precludes ruling out a moderate or asymptotic decline in CECL’s effect on growth over time.

5.3.2 Interaction with Capital Constraints

This paper generally tests the effect of CECL based on its effect on bank capital. In this section, we test whether the effect of CECL on loan growth is stronger at banks where the level of capital is closer to requirements and potentially a greater constraint on growth. We measure the bindingness of the constraint in two ways. As a simple approach, we use the minimum tier 1 leverage ratio of each bank over the period 2016 through 2019, prior to any CECL adoption.²⁸ As a more refined but complex alternative, we measure the ratio of each firms’ common equity tier 1 capital relative to requirements, and do the same for tier 1 and total capital measures. We then take the minimum of these ratios across the three definitions of capital, and across the period 2016 through 2019 and refer to this as “excess capital."

Then, we interact leverage and excess capital-based bindingness within our more-robust bank level IV specification (equation 2) in two ways. First, we find the median of each measure, and define dummies \(\mbox{MoreConstrained}\) (and \(\mbox{LessConstrained}\)) which equal one for banks that are more (and less) constrained than the median bank. The estimation equations with this approach are as follows, where the controls are defined as previously: \[\begin{eqnarray} \mbox{LoanGr}_{i,t} &=& \beta_1\,\mbox{AllowanceCov}_{i,t}\times\mbox{MoreConstrained}_{i} + \beta_2\,\mbox{AllowanceCov}_{i,t}\times \mbox{LessConstrained}_i \nonumber \\ && + \Gamma_{p}\,\mbox{BankControls}_{i,t-1} + \Lambda_{p,t}\,\mbox{Region}_{i,t-1} + \Phi_{t}\mbox{LoanShare}_{i,t} + \theta_{i} + \varepsilon_{i,t} \qquad \quad \nonumber\\ \mbox{AllowanceCov}_{i,t} &=& (\alpha_{0}\mbox{MoreConstrained}_{i} + \alpha_{1}\mbox{LessConstrained}_{i})\times\mbox{AdoptionEffect}_{i}\times\mbox{OnCECL}_{i,t} \nonumber \\ && + \Gamma_{p}\,\mbox{BankControls}_{i,t-1} + \Lambda_{p,t}\,\mbox{Region}_{i,t-1} + \Phi^{*}_{t}\mbox{LoanShare}_{i,t} + \theta^{*}_{i} + \varepsilon^{*}_{i,t}, \nonumber \end{eqnarray}\] Second, we normalize each ratio measure so that it has mean zero and standard deviation of one, and then interact this variable, \(\mbox{Capital}_i\), with allowance coverage within the regression equation. The coefficient on allowance coverage is then interpreted as the effect of allowances on loan growth for a firm with average capital constraints, while the coefficient on the interaction indicates how the strength of that effect changes with each standard deviation in the capital constraint. The estimation equations under this second approach are: \[\begin{eqnarray} \mbox{LoanGr}_{i,t} &=& \beta_1\,\mbox{AllowanceCov}_{i,t} + \beta_2\,\mbox{AllowanceCov}_{i,t}\times \mbox{Capital}_i \nonumber \\ && + \Gamma_{p}\,\mbox{BankControls}_{i,t-1} + \Lambda_{p,t}\,\mbox{Region}_{i,t-1} + \Phi_{t}\mbox{LoanShare}_{i,t} + \theta_{i} + \varepsilon_{i,t} \qquad \quad \nonumber\\ \mbox{AllowanceCov}_{it} &=& \alpha_{0}\mbox{Capital}_{i} \times\mbox{AdoptionEffect}_{i}\times\mbox{OnCECL}_{i,t} \nonumber \\ && + \Gamma_{p}\,\mbox{BankControls}_{i,t-1} + \Lambda_{p,t}\,\mbox{Region}_{i,t-1} + \Phi^{*}_{t}\mbox{LoanShare}_{i,t} + \theta^{*}_{i} + \varepsilon^{*}_{i,t} \nonumber \end{eqnarray}\]

Table 5 reports the coefficients of interest (\(\beta_{1}\) and \(\beta_{2}\)) from the second-stage regressions above. Columns (1) and (2) show the effect of allowances on loan growth for the more and less constrained banks using the two measures of capital constraint bindingness. Under either measure, the CECL effect on allowances appears to slow loan growth more for firms with tighter capital constraints, although the results for more and less constrained banks are only statistically significantly different from each other when using the more complex capital constraint measure.

Table 5: IV Regression Results for More and Less Capital Constrained Banks

The table below presents the \(\beta\) coefficients from the first (columns 1 and 2) and second (columns 3 and 4) set of loan growth regression equations below. We measure capital constraints (\(\mbox{MoreConstrained}\), \(\mbox{LessConstrained}\), and \(\mbox{Capital}\)) with leverage ratios (columns 1 and 3) and excess capital (columns 2 and 4). \(\mbox{Capital}\) in the second set of equations is normalized to mean 0 and standard deviation of 1. See table 4 for a description of the other control variables. \[\begin{eqnarray} \mbox{LoanGr}_{i,t} &=& \beta_1\,\mbox{AllowanceCov}_{i,t}\times\mbox{MoreConstrained}_{i} + \beta_2\,\mbox{AllowanceCov}_{i,t}\times \mbox{LessConstrained}_i \nonumber && + \Gamma_{p}\,\mbox{BankControls}_{i,t-1} + \Lambda_{p,t}\,\mbox{Region}_{i,t-1} + \Phi_{t}\mbox{LoanShare}_{i,t} + \theta_{i} + \varepsilon_{i,t} \qquad \quad \nonumber \mbox{AllowanceCov}_{i,t} &=& (\alpha_{0}\mbox{MoreConstrained}_{i} + \alpha_{1}\mbox{LessConstrained}_{i})\times\mbox{AdoptionEffect}_{i}\times\mbox{OnCECL}_{i,t} \nonumber && + \Gamma_{p}\,\mbox{BankControls}_{i,t-1} + \Lambda_{p,t}\,\mbox{Region}_{i,t-1} + \Phi^{*}_{t}\mbox{LoanShare}_{i,t} + \theta^{*}_{i} + \varepsilon^{*}_{i,t} \nonumber \end{eqnarray}\] \[\begin{eqnarray} \mbox{LoanGr}_{i,t} &=& \beta_1\,\mbox{AllowanceCov}_{i,t} + \beta_2\,\mbox{AllowanceCov}_{i,t}\times \mbox{Capital}_i \nonumber && + \Gamma_{p}\,\mbox{BankControls}_{i,t-1} + \Lambda_{p,t}\,\mbox{Region}_{i,t-1} + \Phi_{t}\mbox{LoanShare}_{i,t} + \theta_{i} + \varepsilon_{i,t} \qquad \quad \nonumber \mbox{AllowanceCov}_{it} &=& \alpha_{0}\mbox{Capital}_{i} \times\mbox{AdoptionEffect}_{i}\times\mbox{OnCECL}_{i,t} \nonumber && + \Gamma_{p}\,\mbox{BankControls}_{i,t-1} + \Lambda_{p,t}\,\mbox{Region}_{i,t-1} + \Phi^{*}_{t}\mbox{LoanShare}_{i,t} + \theta^{*}_{i} + \varepsilon^{*}_{i,t} \nonumber \end{eqnarray}\]

Standard errors are in parentheses.

\(^{*}\) \(p<0.10\), \(^{**}\) \(p<0.05\), \(^{***}\) \(p<0.01\).

Columns (3) and (4) provide results using the normalized measure of capital constraint bindingness. As expected, the coefficient on allowance coverage is negative, and the coefficient on the interaction term is positive; both are statistically significant. Given the normalization, this means that a bank would need a lot of excess capital for CECL to not have an effect on lending. It takes about 3 standard deviations above the mean capital measure to offset the allowance coverage effect on loan growth.²⁹

5.3.3 Capital Distributions

Banks face strong incentives to meet regulatory capital requirements, and possibly market expectations around maintaining a safe buffer of capital above them. To the extent that increases in allowances have little affect on lending quantities, which could theoretically be true,³⁰ then the bank would need to reduce its capital distributions to maintain regulatory capital ratios. Conversely, absence of evidence that capital distributions are affected by changes in allowances provides indirect evidence that the effect on lending is economically significant.

In this section, we test the impact of CECL on capital distributions, considering common stock dividends and stock repurchases. We focus on these measures of capital distributions as they reflect the least costly margin of adjustment.³¹ Floyd et al., 2015 show that banks have a higher and more stable propensity to pay dividends than other industrial firms. Fahlenbrach et al., 2024 find that banks have lower payouts in times of high regulation uncertainty. As a result, we expect that if there is any capital distribution reaction to CECL, it will most likely come in the form of lower repurchases.

Figure 5 plots average distributions scaled by assets by adoption cohort. Repurchases are only plotted for the early adopters because late CECL adopters are overwhelming privately owned banks with very limited repurchase activity. All distributions dip in 2020, likely driven by the pandemic. The increase in economic uncertainty could make capital distributions less appealing for banks.³²

Figure 5: Capital Distributions from Early and Late CECL Adopters

Accessible version

To establish causality with CECL adoption, we turn to our preferred IV specification for the analysis: \[\begin{eqnarray} \mbox{Distribution}_{i,t} &=& \beta\,\mbox{AllowanceCov}_{i,t} + \Gamma_{p}\,\mbox{BankControls}_{i,t-1} + \Lambda_{p,t}\,\mbox{Region}_{i,t-1} \nonumber \\ &&+ \Phi_{t}\mbox{LoanShare}_{i,t}+ \theta_{i} + \varepsilon_{i,t} \qquad \quad\tag{7} \\ \mbox{AllowanceCov}_{i,t} &=& \alpha\mbox{AdoptionEffect}_{i}\times\mbox{OnCECL}_{i,t} + \Gamma_{p}\,\mbox{BankControls}_{i,t-1} + \Lambda_{p,t}\,\mbox{Region}_{i,t-1} \nonumber \\ && + \Phi^{*}_{t}\mbox{LoanShare}_{i,t} + \theta^{*}_{i} + \varepsilon^{*}_{i,t},\tag{8} \end{eqnarray}\] where the control variables are defined as before and include asset quality controls, bank and loan characteristic variables, and the macroeconomic controls. However, these regressions are run on annual data as distributions data are very lumpy, with strong seasonal patterns for some firms. For example, some banks have quarterly dividend policies while others have annual ones. Similarly, the timing of repurchases within the calendar year varies by firm. \(Distribution\) is defined as either common dividends, measured both in levels and changes, or share repurchases. We scale these distributions by total assets, but results are very similar if we scale by loans, as is done for our measures of allowances and loan growth. See table A1 in the appendix for details about the variable construction.

Table 6 reports the second stage regression results using the asset-scaled measures of capital distributions. The first two columns consider levels and changes in common dividends. The last column looks at share repurchases–considering early adopters with holding companies only–hence the smaller observation count. The coefficients on the variable of interest for the dividend tests are positive, opposite the expected sign, but economically small and not statistically significant. There is not a significant change in dividend practices following the adoption of CECL. The coefficient for the repurchases test is negative but also not statistically different from zero.

Table 6: Capital Distribution Regression Results

This table shows the results of using the following specifications. The coefficient on allowance coverage (\(\beta\)) for the capital distribution specification is reported. \[\begin{eqnarray} \mbox{Distribution}_{i,t} &=& \beta\,\mbox{AllowanceCov}_{i,t} + \Gamma_{p}\,\mbox{BankControls}_{i,t-1} + \Lambda_{p,t}\,\mbox{Region}_{i,t-1} + \Phi_{t}\mbox{LoanShare}_{i,t} + \theta_{i} + \varepsilon_{i,t} \qquad \quad \nonumber \mbox{AllowanceCov}_{i,t} &=& \alpha\mbox{AdoptionEffect}_{i}\times\mbox{OnCECL}_{i,t} + \Gamma_{p}\,\mbox{BankControls}_{i,t-1} + \Lambda_{p,t}\,\mbox{Region}_{i,t-1} \nonumber && + \Phi^{*}_{t}\mbox{LoanShare}_{i,t} + \theta^{*}_{i} + \varepsilon^{*}_{i,t}, \nonumber \end{eqnarray}\] where the control variables are defined as before and include asset quality controls, bank and loan characteristic variables, and the macroeconomic controls. Asset quality is measured by delinquency rates at 30-day, 90-day, and nonaccrual frequencies. bank characteristic variables are lagged by one period to avoid endogeneity issues and include log assets, leverage ratio, and return on assets to control for size, solvency, and profitability. \(\Phi^{*}_{t}\mbox{LoanShare}_{i,t}\) is a vector of five variables that are equal to loan shares of five of the six loan types, and OnCECL\(_{i,t}\) is a dummy variable set to one after CECL adoption by bank \(i\).

Standard errors in parentheses

\(^{*}\) \(p<0.10\), \(^{**}\) \(p<0.05\), \(^{***}\) \(p<0.01\)

CECL adoption does not appear to change capital distribution behavior. However, the standard error is large enough that the coefficients would need to be large to find evidence of an effect. Thus, our evidence against an economically meaningful effect of allowances on distributions is limited.

5.4 Robustness Tests

Overall, our analysis shows that the effect of CECL adoption on allowances decreased loan growth moderately, and that this effect was stronger for more capital constrained banks. In this section, we test the robustness of these conclusions to our specific methodological approach.

First, we address concerns that two-way fixed effect estimators like ours can yield biased or counter-intuitive estimates of treatment effects when the treatment is introduced in a staggered manner and the treatment effect is possibly heterogeneous (e.g., de Chaisémartin and D'Haultfoeuille, 2020 and Baker et al., 2022). As a diagnostic for the magnitude of the potential issue, we estimate the implicit weighting of early and late adopters over time in the estimated average treatment effect. We find that while our estimate of average treatment effect does effectively weight some banks negatively late in the sample, the weights are small in magnitude. Thus, heterogeneity in the treatment effect would have to be fairly extreme to significantly bias our estimates.

To further address the concerns above with two-way fixed effects regressions, in table 7 column (2) we estimate our benchmark IV specification using only observations through 2022 so that late CECL adopters are effectively never-adopters within the estimation sample. As compared our bank-level IV estimate from table 4 (in column (1)), the response of loan growth to changes in allowances appears a bit stronger based on only the cohort of 2020 adopters.

Table 7: IV Regression Alternatives

This table shows the results of variations on regression equation 2, shown in column (1). Only the coefficient on allowance coverage for the loan growth (\(\beta\)) is reported. Column (2) includes only observations through 2022. Column (3) weights observations by dollars. Column (4) uses as instruments the CECL adoption effect interacted with an early adopter dummy and every post-adoption quarter. Column (5) reports results using synthetic IV, with a control group constructed to minimize differences in quarterly pre-CECL loan growth.

Standard errors in parentheses.

\(^{*}\) \(p<0.10\), \(^{**}\) \(p<0.05\), \(^{***}\) \(p<0.01\).

The next alternative regression in column (3) is provided to test whether our conclusions hold when considering dollar-weighted observations, or the aggregate of the banks in our sample. The results are larger, meaning that larger banks tended to have larger decreases in loan growth following CECL adoption. As early adopting banks tend to be larger, the similarity of results in columns (2) and (3) is not unexpected.

The regression specification tested in column (4) revisits our choice of instrumental variable. Noting that the impact of CECL on allowances may change over time, generally reverting to the mean from the adoption impact as the firm loan portfolio and CECL models adjust, but also possibly changing in response to macroeconomic conditions. We check results when interacting the adoption effect of CECL with dummies for each post-adoption quarter in the first-stage regression as shown in equation 9. Leveraging time variation in the strength of the adoption impact instrument results in a stronger instrument overall and consequent slight reduction in standard errors, but the point estimate for the effect of allowances on loan growth is almost unchanged. \[\begin{eqnarray} \mbox{AllowanceCov}_{it} &=& \alpha^{early}_{t}\mbox{AdoptionEffect}_{i}\times\mbox{OnCECL}_{i,t} +\alpha^{late}_{t}\mbox{AdoptionEffect}_{i}\times\mbox{OnCECL}_{i,t} \nonumber \\ && + \Gamma_{p}\,\mbox{BankControls}_{i,t-1} + \Lambda_{p,t}\,\mbox{Region}_{i,t-1} + \Phi^{*}_{t}\mbox{LoanShare}_{i,t} + \theta^{*}_{i} + \varepsilon^{*}_{i,t}\tag{9} \end{eqnarray}\]

Lastly, we run a synthetic instrumental variable (SIV) regression, as described in Gulek and Vives-i-Bastida, 2025. The SIV approach produces unbiased estimates in settings where the regressor is both endogenous and the instruments are potentially correlated with an unobserved trend. For example, given the average size difference in early and late CECL adopters in our sample, some economic shocks may affect these two groups of banks differently—even if there is no good evidence of differences in the average growth trends prior to CECL adoption. The trade off for this additional robustness is usually less statistical power.

In our application of the SIV approach, we first find synthetic control groups for each bank. We take the conventional approach of a establishing a control group for each bank that minimizes differences in quarterly pre-CECL loan growth with the bank.³³ We then ensure that the synthetic controls appear reasonable and robust.³⁴ Next, we use the synthetic control for each bank to remove the influence of unobserved trends from the dependent, independent, and instrumental variables for each bank. Finally, an IV regression is run using this trend de-biased data spanning 2020 through 2024:Q2, following equation 10, where \(db\) indicates that the variable has been debiased via differencing against its synthetic control. Standard errors are constructed by bootstrapping methods where the units are banks. \[\begin{eqnarray} \mbox{LoanGr}^{db}_{i,t} &=& \beta\,\mbox{AllowanceCov}^{db}_{i,t} + \varepsilon_{i,t} \qquad \quad\tag{10} \\ \mbox{AllowanceCov}^{db}_{i,t} &=& \alpha(\mbox{AdoptionEffect}_{i}\times\mbox{OnCECL}_{i,t})^{db} + \varepsilon^{*}_{i,t} \nonumber \end{eqnarray}\]

Our synthetic instrumental variables estimate is reported in column (5) of table 7. Our estimate of the effect of changes in allowances on loan growth is smaller in magnitude when using SIV; it is possible that economic conditions were less favorable to business models that led to larger CECL effect on allowances. However, the standard error is large enough that this difference is not statistically significant.

6 Conclusion

While CECL was designed to improve the timeliness of loss recognition, our results suggest that the economic effects extend beyond the accounting to real bank lending behavior. We find evidence that the general increase in loan allowances under CECL reduced lending, consistent with the literature on the effect of bank capital requirements. This lending effect was economically significant, representing about 12 percent of the average annual loan growth in our sample. Even so, the effect of CECL on lending may have been moderated by the ample regulatory capital buffers that banks had on hand during this period, and the transition rules put in place to spread the adoption effect on regulatory capital over several years.

The CECL methodology is designed to make loan loss allowances more sensitive to macroeconomic conditions. Thus, a more complete assessment of the macroeconomic impact of CECL would consider the impact of CECL on the cyclicality of loan loss allowances over the economic cycle. While CECL adopters in particular had large increases in allowances when the economic outlook changed abruptly in 2020, we will not know how well allowances will build under CECL to counter any credit risk building in the banking industry until the economy has a more standard credit cycle. Future research could explore the long-term effects of CECL on the volatility of economic growth, as well as potential policy adjustments to mitigate negative impacts on credit availability.

A Appendix

A1 Variable Definitions

Table A1: Variable Definitions

Constructed Variables

Figure A1: Geographical Distribution of Branches
These figures plot differences in each state’s share of bank branches as of the end of 2019. The top plot provides the share from the aggregate of banks in our sample minus the aggregate of all banks, while the bottom plot provides the average difference between the early and late CECL adopting banks within our sample. Branches include only main office and full service branches; banks without branches are excluded. Source: FFIEC branch data and author calculations.

Accessible version

A2 Parallel Trends Tests

Figure A2: Loan Growth Effects for Early Adopters
This chart shows the coefficients \(\beta_{p,t}\) from the regression equation below, where each panel represents the time series of \(\beta\) for a different loan portfolio \(p\) (cards = credit card loans, com = commercial and industrial loans, cre = commercial real estate loans, cnstr = construction loans, and othcons = other consumer loans), and the dependent variable \(Y\) is either loan growth (this figure) or allowance coverage (Figure A3 below). The parallel trends assumption is tested by checking whether the \(\beta_{p,t}\) are jointly statistically significant from each other for those quarters \(t\) occuring prior to early CECL adoption in 2020. The Wald test p-values are indicated on the plots. Source: Call Report, FR Y-9C, and author calculations.
\[\begin{eqnarray} Y_{p,i,t} &=& \beta_{p,t}\,\mbox{EarlyAdopter}_{i} + \Gamma_{p}\,\mbox{BankControls}_{i,t-1} + \Lambda_{p,t}\,\mbox{Region}_{i,t-1} + \delta_{p,t} + \theta_{p,i} + \varepsilon_{p,i,t} \nonumber \end{eqnarray}\]

Accessible version

Figure A3: Allowance Effects for Early Adopters
See Figure A2 for a description.

Accessible version

A3 Difference-in-Differences Test Results

Table A2: Allowance Coverage Regression Results

This table reports the allowance coverage regression results summarized by figure 3. The regression specification is: \[\begin{eqnarray} Y_{p,i,t} &=& \beta^{early}_{p}\,\mbox{OnCECL}_{i,t} + \beta^{late}_{p}\,\mbox{OnCECL}_{i,t} + \Gamma_{p}\,\mbox{BankControls}_{i,t-1} + \Lambda_{p,t}\,\mbox{Region}_{i,t-1} + \,\delta_{p,t} + \theta_{p,i} + \varepsilon_{p,i,t}, \nonumber \end{eqnarray}\] where \(Y\) is allowance coverage. cards = credit card loans, com = commercial and industrial loans, cre = commercial real estate loans, cnstr = construction loans, and othcons = other consumer loans. Source: Call Report, FR Y-9C, and author calculations.

Standard errors are in parentheses.

\(^{*}\) \(p<0.10\), \(^{**}\) \(p<0.05\), \(^{***}\) \(p<0.01\).

Table A3: Loan Growth Regression Estimates

This table reports the loan growth regression results summarized by figure 3. The regression specification is: \[\begin{eqnarray} Y_{p,i,t} &=& \beta^{early}_{p}\,\mbox{OnCECL}_{i,t} + \beta^{late}_{p}\,\mbox{OnCECL}_{i,t} + \Gamma_{p}\,\mbox{BankControls}_{i,t-1} + \Lambda_{p,t}\,\mbox{Region}_{i,t-1} + \,\delta_{p,t} + \theta_{p,i} + \varepsilon_{p,i,t}, \nonumber \end{eqnarray}\] where \(Y\) is loan growth. cards = credit card loans, com = commercial and industrial loans, cre = commercial real estate loans, cnstr = construction loans, and othcons = other consumer loans. Source: Call Report, FR Y-9C, and author calculations.

Standard errors are in parentheses.

\(^{*}\) \(p<0.10\), \(^{**}\) \(p<0.05\), \(^{***}\) \(p<0.01\).

Table A4: First-Stage Coefficients

This table shows the coefficients \(\alpha\) from the first-stage regression equations 3 (first column) and 5 (second column), as duplicated below. Source: Call Report, FR Y-9C, and author calculations. \[\begin{eqnarray} \mbox{AllowanceCov}_{i,t} &=& \alpha\mbox{AdoptionEffect}_{i}\times\mbox{OnCECL}_{i,t} + \Gamma_{p}\,\mbox{BankControls}_{i,t-1} + \Lambda_{p,t}\,\mbox{Region}_{i,t-1} + \Phi^{*}_{t}\mbox{LoanShare}_{i,t} + \theta^{*}_{i} + \varepsilon^{*}_{i,t} \nonumber \mbox{AllowanceCov}_{p,i,t} &=& \alpha_{p}\,\mbox{OnCECL}_{i,t} + \Gamma_{p}\,\mbox{BankControls}_{i,t-1} + \Lambda_{p,t}\,\mbox{Region}_{i,t-1} + \delta^{*}_{p,t} + \theta^{*}_{i} + \varepsilon^{*}_{p,i,t} \nonumber \end{eqnarray}\]

Standard errors are in parentheses.

\(^{*}\) \(p<0.10\), \(^{**}\) \(p<0.05\), \(^{***}\) \(p<0.01\).

Footnotes