Benefits and Consequences of CECL-Adoption During COVID-19

Introduction

In June 2016, the Financial Accounting Standards Board (FASB) introduced the current expected credit loss (CECL) model to replace its predecessor, the incurred credit loss (ICL) model, marking a shift toward a more forward-looking and proactive accounting approach for the banking sector. Since banks following ICL determine potential credit losses only when losses are foreseeable within a short horizon, the ICL model has been criticized by academics and regulators for not providing timely loan loss provisions (LLPs) and undesirably facilitating banks’ procyclical behavior (e.g., Beatty & Liao, 2011; Bushman & Williams, 2015; Nicoletti, 2018). As a result, banks may delay recognizing credit losses or underestimate future credit losses during economic booms when the risk is lower, and then record a sudden increase in the recognition of credit losses during economic downturns when the risk is higher, resulting in reserves being “too little, too late” and arguably exacerbating the effects of economic cycles (Cohen & Edwards, 2017; Gaston & Song, 2014). In contrast, the CECL model requires banks to recognize the lifetime expected credit losses on all loan receivables at the time of loan origination and is thus likely to reduce the procyclical tendencies under ICL by accumulating timely and sufficient reserves before an economic downturn (Chae et al., 2018).

Interestingly, early anecdotal evidence reveals the ramifications of the accounting policy. Among the first banks to adopt CECL from the end of 2019, for instance, JP Morgan Chase booked a $16.3 billion LLP for 2020, more than tripling the amount in 2019. However, this was followed by a sizeable reversal of $9.3 billion LLP in 2021 and another round of LLP recognition ($6.4 billion) in 2022. Such a pattern, which is nevertheless procyclical, also occurred for several other systemically important banks, ¹ clearly conflicting with the primary purpose of CECL, which is for banks to build up timely reserves (during good times) and hold them for a longer period of time to absorb losses at credit events (during bad times). In other words, the subsequent rapid LLP reversals, even though in situations of few major credit events and the earlier-than-anticipated economic recovery from 2021, may not be an expected short-term consequence under CECL because they weaken banks’ ability to absorb future credit losses in a timely manner if economic conditions worsen again after those LLPs have been released. Two plausible reasons may explain this phenomenon. First, banks may have exploited the flexibility in this accounting policy to manage earnings (i.e., management bias). ² Second, banks may have booked “too much, too soon” in LLPs due to the inherent difficulty in predicting future credit losses, particularly during a time of the COVID-19 when an economic inflection point is difficult to predict (i.e., management error). ³ Both explanations reflect underlying managerial short-termism. Specifically, managers may have incentives to prioritize short-term earnings targets due to their compensation schemes or pressure from investors to meet analysts’ forecasts. These pressures are particularly salient during economic recoveries, such as in 2021, when market expectations might rebound faster than banks’ core operations. Consequently, the rapid LLP reversals could serve as a tool for earnings management and short-termism. Even if part of the later reversals could stem from management error, both the timing and size of the reversals relative to the initial LLP buildup remain highly procyclical and questionable, again pointing to either opportunistic or incidental short-termism.

Regardless of the motivations, banks’ short-termism in building up reserves after the CECL-adoption (by releasing them within a short time span), as opposed to the intended CECL purpose, is concerning in that the CECL-adoption may bring unnecessary volatility in the estimation of LLPs that could further magnify procyclicality. Fatouh and Giansante (2020) document that when the economy is volatile, the large differences in lifetime probabilities of default under CECL between booms and busts cause sharp increases in LLPs in deep downturns, also suggesting procyclicality brought by volatile LLPs under CECL. Following these considerations, in this paper, we empirically examine the differences in loan loss provisioning patterns and volatilities between CECL-adopters and non-CECL-adopters, the timeliness and validity of LLPs, and the associated capital market impacts and potential motivations. While FASB initially mandated CECL for all large banks starting from the end of 2019, not all have adopted CECL. ⁴ Also, recognizing economic and operational challenges posed by COVID-19, regulators issued an interim final rule (IFR) for a five-year transition period to delay CECL’s impact on regulatory capital. Therefore, the non-universal CECL-adoption by U.S. banks, together with the unexpected onset of the COVID-19 pandemic in 2020, provides a unique natural experiment setting involving an accounting policy change and an exogenous shock, allowing for a difference-in-difference (DiD) method to investigate the early impact of CECL.

Using a sample of 289 U.S. bank holding companies (BHCs) that are static CECL- or non-CECL-adopters ⁵ between 2018 and 2021, we first examine the CECL impact on LLP and ALL volatilities in a DiD setting to empirically verify the aforementioned ramification. We find that from 2018–19 to 2020–21, the increase in ALL (LLP) volatility is 50% (100%) higher for CECL-adopters than for non-CECL-adopters. As expected, we also find a similar pattern for the change in volatility of earnings before taxes (EBT) but not for the change in volatility of earnings before taxes and provisions (EBTP). These results are consistent with CECL-adopters expectedly increasing LLPs and ALL following the policy adoption in 2020 but unexpectedly booking large reversals in 2021 (i.e., the volatilities might not have increased as much in 2021 had those reversals not been booked), suggesting that the increased LLP and ALL volatilities are a direct consequence of the CECL-adoption. Our results on changes in LLP and ALL volatilities are robust to a range of sensitivity checks.

Next, we test whether CECL-adopters’ LLPs are timelier and more valid (as intended) than non-CECL-adopters’ after the onset of CECL-adoption in 2020. Consistent with prior literature (e.g., Beatty & Liao, 2021; Nicoletti, 2018), we measure LLP timeliness as the marginal correlation coefficient between the current period’s LLPs and future periods’ incremental non-performing loans (NPLs). We measure LLP validity as the marginal correlation coefficient between the current period’s LLPs and future periods’ net loan charge-offs (NCO). We find that LLP timeliness and validity of CECL-adopters are not different from those of non-CECL-adopters in 2018 and 2019 but are significantly higher than those of non-CECL-adopters in 2020 and 2021. As a robustness check, we also employ a DiD design with three-way interactions and find consistent results. These results suggest that, compared to non-CECL-adopters, CECL-adopters’ LLPs are more forward-looking and thus better predictors of future NPL and NCO, which is consistent with the mandate of the CECL policy. Following LLP timeliness and validity, we also investigate whether CECL-adopters are more likely to incorporate more forward-looking credit losses of their current loan portfolios beyond the current and the immediate future periods by examining “abnormal” LLPs (two-stage LLP regression ⁶ ) and directly controlling for the normal LLP level (single-stage LLP regression). We continue to find the expected results.

On capital market implications, we explore whether the stock market credits the CECL model’s ability to incorporate forward-looking information in bank earnings in post-CECL years. Consistent with prior literature, we document that the cumulative abnormal return (CAR) around earnings announcement dates is positively related to earnings surprise. ⁷ Moreover, within a DiD empirical design, we find that such earnings response coefficients are significantly higher for CECL-adopters than for non-CECL-adopters in 2020, but not in 2021. Focusing on the year 2021, we find that banks with LLP reversals show lower earnings response coefficients compared to banks without LLP reversals. Our results imply that, while the stock market credits CECL’s ability to incorporate more information in 2020, it may see the large and hasty LLP reversals in 2021 as a potential opportunistic behavior that nevertheless diminishes the informativeness of earnings.

We further investigate the potential motivations behind these LLP reversals. Market participants may interpret the LLP reversals in 2021 as an indication of banks adopting a short-term orientation with the CECL model. The motivations, as aforementioned, may be opportunistic if such reversals are booked to mask the lower-than-anticipated earnings in 2021. This incentive can be higher when managers of CECL-adopters ex-ante anticipate a higher earnings response coefficient—because failing to beat the market consensus may bring negative market reactions. Considering this, we explore whether earnings-per-share (EPS) shortfall before provisions, defined as the difference between the most recent analysts’ consensus forecasts and the EPS before loan loss provisions, is associated with CECL-adopters’ likelihood of booking LLP reversals—we indeed find the expected results. Hence, our findings suggest that such a relationship between EPS shortfall before provisions and the likelihood to use LLP reversals may indicate some opportunistic incentives behind the large and hasty 2021 LLP reversals for CECL-adopters. More interestingly, from an ex-ante perspective, the incentive to beat shortfall with LLP reversals—as opposed to the actual lower earnings response coefficients associated with LLP reversals documented earlier—may indicate more managerial self-interests than investor-driven concerns, as managers may have recognized that booking large and rapid reversals was unlikely to be interpreted as a positive signal by the market.

Our study contributes to the literature in several important ways. First, this study extends the emerging literature on evaluating the expected credit loss accounting models under both the FASB and IFRS. ⁸ While studies on IFRS 9 have documented increased loan loss allowances upon transition, improved timeliness in loan loss provisioning, and better prediction of credit and equity risk, early evidence of CECL’s impact is sparse. Among the exceptions, Beck and Beck (2022) document an increase in provisions in 2020 quarter 1, but not in quarters 2 and 3; Chen et al. (2022) find that banks that adopted CECL show reduced loan growth during the pandemic, suggesting potential lending procyclicality. However, these studies are constrained by short post-adoption time periods. In contrast, we evaluate post-adoption changes in patterns of loan loss provisions and allowances over a two-year period and explore managers’ short-termism and opportunistic earnings management incentives. Our results of the increased volatility suggest that regardless of managers’ ex-ante uncertainty in predicting loan losses during the onset of COVID-19 and initial policy adoption, the procyclical behavior enabled by the flexibility in the accounting policy itself could also be a contributory factor in post-CECL volatile LLPs and ALL.

Second, different from the existing studies, which are mostly international, we focus on U.S. markets, hence mitigating potential biases from unobservable cross-country heterogeneities. The estimation of credit losses, as well as the ECL adoption, can be inherently driven by a series of domestic factors, including economic conditions and policies, governmental intervention, business cycle, regulations, and national credit levels. Consequently, these cross-country differences are challenging to control simultaneously, potentially rendering the ECL adoption endogenous and making it difficult to capture the true effects of ECL on the estimation of loan loss reserves. Therefore, by focusing on U.S. banks with a DiD design, our study contributes to the literature on the CECL impact on banks by mitigating this concern.

Third, our study benefits from the concurrent occurrence of the CECL-adoption and the pandemic in early 2020, with the earlier-than-anticipated economic recovery in 2021. This particular sample period provides us with a unique setting to examine whether there are opportunistic motivations (which, if any, should be amplified by and thus more observable in the context of COVID-19 and the 2021 recovery) behind the large and hasty LLP reversals in 2021, further explaining what shapes the tension between banks’ short-termism in building up reserves and the intended CECL purpose by policymakers. Since prior findings on earnings management or opportunistic behavior under ECL of IFRS 9 are mixed, ⁹ our study adds another important layer of evidence of earnings management to this line of research, alerting policymakers. Our study is different from Obserson (2021), who argues for evidence of earning management incentives by documenting a higher relationship between ALL and return on assets after the implementation of IFRS 9—We test whether the EPS shortfall before provisions is associated with the likelihood of booking LLP reversals, supported by management’s augmented incentives to avoid missing consensus earnings forecast due to the higher ex-ante expected (and also ex-post tested) stock market responsiveness to earnings surprise.

While our empirical evidence indicates management bias in those large LLP reversals, we acknowledge that it is challenging to strictly distinguish between management bias and error in explaining the source of the reversals, given the lack of experience with CECL for managers themselves. In other words, management error could also be manifested as overprovisioning in 2020 and large reversals in 2021, even with an earnings shortfall observed. However, controlling for banks’ ex-ante (i.e., pre-COVID) ability to model LLPs, which potentially reflects the propensity of management error, does not significantly change our conclusions, thus somewhat mitigating the concern about management error instead of management bias.

This paper is organized as follows. Section 2 discusses the related literature and provides the hypothesis development; section 3 describes empirical design; section 4 describes the sample and empirical results; and section 5 offers concluding remarks.

Related Literature and Hypotheses Development

Several existing studies have evaluated the change in loan loss accounting policy from ICL to ECL under IFRS 9. ¹⁰ In terms of the effect of ECL on banks’ loan loss allowances, Lopez-Espinosa et al. (2021) use a sample of 293 banks from 74 countries between 2014 and 2019 and exploit the cross-country differences in the requirements and timing of IFRS 9 implementation. Focusing on the first-day impact, they find that the implementation of ECL is associated with an increase in loan loss allowances and that most of such increases are related to non-default loans. Using a sample of Slovenian banks, Groff and Mörec (2021) document that banks without transferring their non-performing assets to state-assisted restructuring programs recognized additional allowances for financial assets after implementation of ECL. In addition, Gomaa et al. (2019), employing an experimental economic approach, find that both the amount and the adequacy of ECL loan loss allowances on trade receivables increase after implementing a simplified approach to ECL. Nevertheless, some of these studies document somewhat unexpected results. For example, based on a sample of public entities in China, Guo et al., 2022 do not find significant changes in loan loss allowances on financial assets on the transition to ECL. ¹¹

On the effect of ECL on the timeliness of LLPs, Oberson (2021) uses a sample of 69 banks from 24 countries over 2014–2019 and documents that banks implementing IFRS 9’s ECL recognize timelier LLPs. Kim et al. (2021) use a sample of banks from 33 countries and find that the transition to ECL significantly improves LLP timeliness, with such an effect stronger for riskier banks and banks that previously recognized smaller loan loss allowances. In addition, based on a sample of 326 banks in China between 2015 and 2019, Hung et al. (2022) documented that the timeliness of LLPs increases for non-state-owned banks but not state-owned banks.

In addition, some existing studies examine the effect of ECL on the predictability of provisions and allowances on future credit risks. Comparing the informativeness of LLPs in credit default swap (CDS) markets between banks implementing ECL and banks applying ICL, Oberson (2021) finds that the positive relation between banks’ LLPs and CDS pricing is stronger for banks applying ECL, especially for banks with weaker pre-IFRS 9 information environments. Also, Lopez-Espinosa et al. (2021) document that the predictive ability of LLPs for future credit and equity risk is higher under the ECL model in countries experiencing deteriorating credit conditions.

Compared to the existing studies on the effect of IFRS 9, relatively fewer studies have examined the early evidence of CECL-adoption under U.S. GAAP. For instance, using a sample of U.S. bank holding companies between 2019Q1 and 2020Q3, Beck and Beck (2022) examine the initial adoption of CECL and document that CECL-adopting banks report larger provisions in the first quarter of 2020 compared to non-CECL-adopting banks. They also document that in the third quarter of 2020, CECL-adopting banks reported significantly smaller provisions than non-CECL-adopting banks. Thus, despite the limitation of data, findings from Beck and Beck (2022) imply a higher volatility in provisions for CECL-adopting banks. Using a sample of U.S. bank holding companies from 2018Q1 to 2021Q1, Chen et al. (2022) examine the relationship between the CECL approach and lending procyclicality and find that CECL-adopting banks reduce loan growth more during recessions compared to non-CECL-adopting banks. They argue that a key reason for such increased procyclicality can be attributed to the difficulties that macroeconomic models face in predicting business cycle turning points (Covas & Nelson, 2018). Similarly, Yang (2025) studies how the anticipation of CECL influences bank lending during the transition period from 2016 to 2018 and argues that since banks must recognize additional expected credit losses for existing loans by debiting retained earnings, CECL is anticipated to lower regulatory capital on the first day of adoption and, in turn, decreases banks’ willingness to lend, especially for capital-constrained banks. Using two ten-quarter windows surrounding the fourth quarter of 2015 (when CECL was initially approved), Yang (2025) applies a DiD design and finds that capital-constrained banks reduce their growth of total loans and residential loans following CECL’s approval. Chae et al. (2018) suggest that while provisions under CECL are generally less procyclical compared to those under ICL, CECL may complicate the comparability of provisions across banks and time, and that market participants will need to disentangle the degree to which variation in provisions across firms is driven by underlying risk versus differences in modeling assumptions. In addition, Grealis et al. (2021) use quarterly Call reports across the implementation window and compare the first quarter of 2020 to prior quarters and years to discern the difference in banks’ discretionary accruals. They find additional discretion opportunities available to management in the post-CECL period, which are contrary to the FASB’s intent. Jacobs Jr. (2020) also notes that CECL may lead to under-prediction of credit losses and that the goal of mitigating procyclicality may fail to materialize.

Moreover, Schroeder (2023) argues that early evidence does not support widely publicized concerns about CECL, including reduced lending, increased capital needs, and investor confusion. Harris et al. (2018) develop a measure of the one-year-ahead expected rate of credit losses using disclosed credit information of U.S. BHCs and find that it substantially outperforms other credit risk metrics and reflects nearly all the credit loss-related information in the charge-offs, providing evidence of the accuracy of CECL. Mahieux et al. (2023) document that more timely information brought by CECL enhances efficiency either when banks are insufficiently capitalized or when regulatory intervention is likely to be effective, but not when banks are moderately capitalized or when regulatory intervention is costly. Handorf (2018) suggests that the change in GAAP will not necessarily require a significant increase in allowances for the average bank predicated upon the model adopted, and that the accounting shift only impacts unimpaired long-term loans, which are of better quality than classified advances. Wheeler (2021) finds that investors are able to obtain information about expected credit losses beyond that recorded in financial statements, and that these unrecognized expected credit losses are negatively associated with bank stock prices. Willi III (2020) documents that since CECL will dramatically alter how expected credit losses are calculated, it poses an enormous data challenge. As a result, many banks have been forced to turn to third-party vendors for assistance, and this change will weigh much more heavily on smaller banks. Interestingly, Granja and Nagel (2023) find that greater reserve requirements following the adoption of CECL induce a statistically significant but economically moderate increase in loan interest rates, potentially coinciding with the reduced loan growth documented by other studies.

Building on these contemporary studies, we extend the sample period to two years before and after the initial CECL-adoption (i.e., between 2018Q1 and 2021Q4) and comprehensively examine the benefits and consequences of the CECL-adoption during COVID-19. Specifically, we study the influence of CECL-adoption on the volatilities of LLPs and ALL, the timeliness and validity of LLPs, market reactions to earnings surprises, and potential earnings management motivations.

Ideally, the CECL model should not introduce higher volatility or procyclicality in banks’ quarterly reported LLPs and ALL. Since CECL considers expected credit losses over the lifetime of a loan at the time of origination or acquisition and incorporates forward-looking information, CECL aims to provide a more accurate reflection of credit risk, thus leading to more predictable and stable provisions over time, rather than passively reacting to short-term changes in credit conditions. However, given the post-adoption evidence of major U.S. banks’ increased LLPs in 2020 followed by large and hasty LLP reversals in 2021, we argue that managers’ opportunistic earnings management incentives, if any, may lead to higher LLP and ALL volatilities. These opportunistic incentives can stem from the accounting policy itself, as CECL-adopters have higher flexibility in loan loss provisioning in that LLPs under the CECL approach are based on extensive management judgments and estimations. Consequently, both over- and under-provisioning due to the managerial bias can construct significant discretionary LLPs (either positive or negative) and thus potentially contribute to the sources of earnings manipulation for managers’ opportunism. We also argue that although the intended CECL purpose is for banks to build up more stable reserves, higher flexibility itself under the CECL approach may lead to higher volatilities in LLPs and ALL, because banks’ short-termism can be more pronounced with higher flexibility in LLP estimation. For example, managers of CECL-adopting banks may continuously adjust current probabilities of defaults to reflect changing economic conditions, especially during fluctuations. However, those fluctuations should have been absorbed and reflected in LLPs at least one period ahead of the present time (i.e., more forward-looking and timelier LLPs), and if they have not (e.g., due to banks’ short-termism), they are likely to lead to higher volatilities in current LLPs and ALL, as a direct consequence of higher flexibility under CECL.

Moreover, conditional on the arguments above, CECL-adopters may also exhibit a higher volatility in earnings before tax than non-CECL-adopters, given that LLP is the largest accrual of banks. ¹² To further demonstrate that the main source of earnings volatility is the volatility in LLPs, we also predict that CECL-adopters do not exhibit higher volatility in earnings before tax and provisions than non-CECL-adopters.

Based on the above discussion, we have the following hypotheses (framed in DiD form):

H1a: CECL-adopters exhibit a higher increase in volatilities of LLPs and ALL compared to non-CECL-adopters, after versus before CECL-adoption.

Moreover, by requiring banks to recognize lifetime expected credit losses in their current LLPs, the main purpose of the CECL approach is to increase the timeliness of LLPs, which is often described by prior research as the ability of the current period’s LLPs to predict future periods’ incremental non-performing loans (e.g., Nicoletti, 2018). Consistent with the literature, we define LLP timeliness as the extent to which current LLPs are related to subsequent periods’ incremental non-performing loans and LLP validity as the extent to which current LLPs are related to subsequent periods’ net loan charge-offs (Beatty et al., 2019; Nicoletti, 2018). Using these two definitions, we evaluate the effectiveness of the CECL approach by examining whether CECL-adopters recognize LLPs that are both timelier and more valid. In addition, if CECL-adopters’ LLPs are timelier and more valid, they should also be more informative about earnings—therefore, we also predict that the earnings announced by CECL-adopters are more informative in the capital market. Last, we focus on the large and hasty LLP reversals made by many CECL-adopters in 2021. We expect that, while the market participants may credit the CECL model for its designed ability to incorporate more future information about future credit losses, they may see through the large and hasty LLP reversals in 2021 and penalize the earnings informativeness. In addition, we further explore the motivations behind the reversals, which may be opportunistic if they are recognized to mask the lower-than-anticipated earnings in 2021 and that such opportunistic incentives can be amplified when managers of CECL-adopters ex-ante anticipate a higher earnings response coefficient—because failing to beat the market consensus may bring severe negative market reactions. Hence, we also aim to explore whether the large LLP reversals in 2021 are associated with earnings shortfall below benchmarks before any provisions.

Based on these discussions, we state the following hypotheses:

H2: CECL-adopters exhibit higher timeliness and validity of LLPs compared to non-CECL-adopters after CECL-adoption, but not before CECL-adoption.

Research Design

Main Empirical Models

To test H1, we estimate the following DiD model, with bank fixed effects ¹⁶ and standard errors of the estimates clustered by banks:

σ A L L ( o r σ L L P , σ E B T , σ E B T P ) = α 0 + α 1 P O S T + α 2 C E C L + α 3 P O S T × C E C L + V + B + ε

(1)

where the dependent variable is σALL or σLLP for H1a, and σEBT or σEBTP for H1b. POST is the after versus before indicator variable which equals 1 if the observation is after or including 2020q1, and 0 if before 2020q1. CECL is the treatment indicator that equals 1 if the observation is a CECL-adopter in 2020–2021, and 0 if the observation is not a CECL-adopter in 2020–2021. V is a vector of control variables including bank-level characteristics measured in 2-year windows, the standard deviation of bank-level characteristics measured in 2-year windows, and the average natural logarithm of quarterly state-level GDP measured in 2-year windows. B is bank fixed effects.

We select bank-level controls that prior studies document are associated with loan loss provisioning and allowances. As suggested by prior studies such as Beck and Narayanamoorthy (2013) and Beatty and Liao (2014), we include the following variables, measured both in 2-year averages and in standard deviations based on 2-year windows—that is, we control for bank size measured as the natural logarithm of lagged total assets (SIZE _{2yr avg} and σSIZE), quarterly earnings before tax and provisions scaled by lagged total assets (EBTP _{2yr avg} and σEBTP), total loans scaled by lagged total assets (LOANS _{2yr avg} and σLOANS), non-performing loans scaled by lagged total loans (NPL _{2yr avg} and σNPL), net loan charge-offs scaled by lagged total loans (NCO _{2yr avg} and σNCO), tier-1 capital ratio (TIER1 _{2yr avg} ), the percentage of commercial loans to total loans (CMLNS _{2yr avg} ), the percentage of consumer loans to total loans (CSLNS _{2yr avg} ), and the percentage of real estate loans to total loans (RELNS _{2yr avg} ). ¹⁷ When σEBT or σEBTP is used as the dependent variable, we also include the 2-year averages and standard deviations of ALL (ALL _{2yr_avg} and σALL) as our bank-level control variables. Appendix A includes detailed definitions of all the variables. H1a hypothesizes that CECL-adopters exhibit a higher increase in volatilities of LLPs and ALL compared to non-CECL-adopters, after versus before CECL-adoption; hence, we expect α ₃ to be positive. H1b hypothesizes that CECL-adopters exhibit a higher increase in volatilities in EBT, but not EBIT, compared to non-CECL-adopters, after versus before CECL-adoption; hence, we expect α ₃ to be positive (not positive) when σEBT (σEBTP) is used as the dependent variable.

To test H2, we estimate the following DiD model and examine the three-way interaction terms that denote the CECL effect on LLP timeliness or validity, with bank fixed effects and standard errors of the estimates clustered by banks:

L L P t = β 0 + β 1 P O S T × C E C L × Δ N P L t + β 2 P O S T × C E C L × Δ N P L t + 1 + β 3 P O S T × C E C L × Δ N P L t + 2 + β 4 P O S T × Δ N P L t + β 5 C E C L × Δ N P L t + β 6 P O S T × Δ N P L t + 1 + β 7 P O S T × C E C L + β 8 C E C L × Δ N P L t + 1 + β 9 P O S T × Δ N P L t + 2 + β 10 C E C L × Δ N P L t + 2 + β 11 P O S T + β 12 Δ N P L t + β 13 C E C L + β 14 Δ N P L t + 1 + β 15 Δ N P L t + 2 + W + M + B + ε

(2)

L L P t = β 0 + β 1 P O S T × C E C L × N C O t + 2 / t + β 2 P O S T × C E C L + β 3 P O S T × N C O t + 2 / t + β 4 C E C L × N C O t + 2 / t + β 5 P O S T + β 6 C E C L + β 7 Δ N C O t + 2 / t + W + M + B + ε

(3)

where the dependent variable is the LLP _t in quarter t. The main explanatory variables of interest are the three-way interaction terms between CECL, POST, and the changes in non-performing loans (ΔNPL _t ) in quarter t, t + 1, and t + 2 in equation (2) and the interaction term between CECL, POST, and future two quarters’ net charge-offs (NCO _t+2/t ) in equation (3). W is an array of bank-quarter-level characteristics based on existing studies such as Nichols et al. (2009) and Nicoletti (2018). These control variables include bank size (SIZE _t-1 ), earnings before tax and provision scaled by lagged total assets (EBTP _t ), quarterly change in total loans scaled by lagged total assets (ΔLOANS _t ), lagged allowances for loan and lease losses scaled by lagged total loans (ALL _t-1 ), tier-1 capital ratio (TIER1 _t ), net charge-offs scaled by lagged total loans (NCO _t ), non-performing loans scaled by lagged total loans (NPL _t ), the percentage of commercial loans to total loans (CMLNS _t ), the percentage of commercial loans to total loans (CSLNS _t ), the percentage of commercial loans to total loans (RELNS _t ), and a dummy variable that equals 1 if the BHC is public (PUBLIC _t ). Moreover, M is an array of macroeconomic variables including the quarterly growth rate of state-level GDP (GDPGROWTH _t ). In addition, B stands for bank fixed effects. H2 predicts that CECL-adopters exhibit higher timeliness and validity of LLPs compared to non-CECL-adopters after CECL-adoption, but not before CECL-adoption; hence, we expect that β ₁ , β ₂ , and β ₃ in equation (2) capturing the CECL effect on LLP timeliness, and β ₁ in equation (3) capturing the CECL effect on LLP validity are positive.

Lastly, the empirical models of the earnings response coefficients tests and earnings management motivation tests will be discussed in detail in sections 4.6 and 4.7.

Empirical Results

Summary Statistics and Univariate Analysis

Appendix B2 presents the summary statistics and correlations of regression variables for the full sample with 4460 observations of 289 unique BHCs. As reported in Appendix B2 Panel A, the mean percentage of BHCs that adopted CECL in the first 2 quarters of 2020 is 57.51%, suggesting that about 58% of banks in our sample are CECL-adopters (i.e., in the treatment group). The mean (median) value of the volatility of LLPs is 0.001 (0.0005) and the mean (median) value of the volatility of ALL is 0.0018 (0.0009). Appendix B2 Panel B reports Pearson’s correlations (Panel B below the diagonal) and Spearman’s correlations (Panel B above the diagonal) between the main variables in our analyses. Because Appendix B2 only shows summary statistics and pairwise univariate correlations for the full sample, which do not provide insights and rationales for the DiD research design, we conduct several univariate analyses by splitting the sample based on CECL-adoption status in 2020–2021.

In Figure 1, we compare patterns of the key variables between CECL-adopters (i.e., banks that adopt CECL in 2020–2021) and non-CECL-adopters (i.e., banks that do not adopt CECL in 2020–2021). Figure 1 Panel A and Panel B show that both the level and trending pattern for ALL and LLP volatilities are similar and parallel for non-CECL-adopters and CECL-adopters before 2020–2021. However, after 2020q1, CECL-adopters demonstrate a much higher increase in the volatility of ALL and LLPs. Figure 1 Panel C breaks down the trending patterns of LLPs for CECL- versus non-CECL-adopters into quarters, showing that starting from 2020q1 and q3, CECL-adopters’ LLPs increase much more dramatically than non-CECL-adopters’ LLPs, and that starting from 2020q4 to 2021q4, CECL-adopters are much more likely to recognize LLP reversals (i.e., negative LLPs) than non-CECL-adopters. In addition, to alleviate the concern that such trending differences may be caused by the fact that public banks are more likely to be CECL-adopters, we limit our sample to public banks only—In Figure 2, we show that even after we limit our sample to public banks, similar results to those in Figure 1 are observed.OPEN IN VIEWER

Figure 1.

Comparison of the key variable changes between CECL and non-CECL banks—including both public and private banks.

OPEN IN VIEWER

Figure 2.

Comparison of key variable changes for CECL and non-CECL banks—public banks only.

Table 1 presents the results of the univariate analyses. Table 1 Panel A compares summary statistics between CECL-adopters and non-CECL-adopters in the years 2020 and 2021. Table 1 Panel B compares summary statistics between CECL-adopters and non-CECL-adopters in the years 2018 and 2019. These two panels demonstrate that in the post-CECL era, CECL-adopters have significantly higher σALL and σLLP and than non-CECL-adopters; while in the pre-CECL era, the differences in σALL and σLLP between CECL-adopters and non-CECL-adopters are insignificant. Moreover, in terms of the absolute value of discretionary LLP (ABSDLLP), ¹⁸ CECL-adopters show a significantly higher ABSDLLP than non-CECL-adopters in the post-CECL era but not in the pre-CECL era. The mean comparisons of all other bank-level characteristics variables are also demonstrated in Table 1.

OPEN IN VIEWER

Table 1.

Univariate Analysis: Mean Differences for CECL-Adopters and Non-CECL-Adopters, Before and After 2020.

	Sample Size	(1)	(2)	(2)–(1)	Test of Differencet StatH0: Diff ≠ 0	p-Value
		Non-CECL-Adopters	CECL-Adopters	Differences
Panel A: Mean differences in 2020 and 2021 (i.e., post-CECL)
σALL 2020–2021	272	0.0015	0.0025	0.0010***	3.54	0.0005
σLLP 2020–2021	272	0.0012	0.0024	0.0012***	5.76	<0.0001
ABSDLLP	2166	0.0010	0.0015	0.0006***	7.73	<0.0001
SIZE	2166	15.54	17.01	1.47***	24.22	<0.0001
EBTP	2227	0.0044	0.0043	−0.0001	−0.93	0.3543
LOANS	2227	0.6293	0.6161	−0.013*	−1.65	0.0992
NPL	2227	0.0091	0.0091	−0.0001*	−0.17	0.8647
NCO	2227	0.0010	0.0019	0.0010***	5.81	<0.0001
TIER1	2084	0.1509	0.1364	−0.0145***	−7.11	<0.0001
ALL	2227	0.0150	0.0180	0.0030***	3.32	0.0009
LLP	2227	0.0009	0.0012	0.0003**	2.51	0.0123
CMLNS	2234	0.1925	0.2192	0.0266***	5.25	<0.0001
CSLNS	2234	0.0778	0.0899	0.0120*	1.70	0.0893
RELNS	2234	0.6608	0.5985	−0.0622***	−6.26	<0.0001
Panel B: Mean differences in 2018 and 2019 (i.e., pre-CECL)
σALL 2018 – 2019	269	0.0008	0.0007	−0.00005	0.50	0.6214
σLLP 2018 – 2019	269	0.0005	0.0004	−0.00007	0.98	0.3280
ABSDLLP	2149	0.0005	0.0005	−0.0000	−0.0630	0.9498
SIZE	2213	15.29	16.75	1.46***	23.45	<0.0001
EBTP	2213	0.0042	0.0045	0.0004***	3.23	0.0013
LOANS	2213	0.6718	0.6556	−0.0162**	2.00	0.0454
NPL	2213	0.0088	0.0088	−0.0000	−0.0087	0.9930
NCO	2213	0.0011	0.0019	0.0008***	4.53	<0.0001
TIER1	2191	0.1488	0.1354	−0.0133***	−6.42	<0.0001
ALL	2213	0.0132	0.0108	−0.0024***	−3.96	0.0001
LLP	2213	0.0006	0.0009	0.0003***	3.38	0.0007
CMLNS	2226	0.1640	0.1900	0.0260***	5.25	<0.0001
CSLNS	2226	0.0811	0.0877	0.0066	0.94	0.3461
RELNS	2226	0.6808	0.6265	−0.0543***	5.32	<0.0001

Table 1 compares the differences in the mean values of for an extended list of bank financial characteristics variables of interest between CECL-adopters and non-CECL-adopters (all static adopters, i.e., no change of status) during the 2 years after 2020 and the 2 years before 2020. Continuous variables are winsorized at top and bottom 1%. All variables are defined in Appendix A.*, **, *** denote significance at the 10%, 5%, and 1% levels, respectively, based on a two-tailed test. Bold fonts means the coefficients of interest.

The Volatility of ALL and LLPs

In this section, we report the results of the test of H1a, which examines the influence of the CECL-adoption on the volatility of banks’ ALL and LLPs in a DiD design. In Table 2 Columns (1) to (3), we report the results using σALL as the dependent variable with bank fixed effects. In Column (1), we regress σALL on the dummy variable POST and find that, regardless of other control variables and CECL-adoption indicator, banks’ σALL increases from 2018–2019 to 2020–2021. In Column (2), we use the DiD design and regress σALL on the interaction of POST and CECL with no other control variables. We document that the increase in σALL for CECL-adopters is significantly higher than that for non-CECL-adopters. In Column (3), we add all control variables and find similar results as in Column (2). In Columns (4) to (6), we repeat the analyses in Columns (1) to (3) but changing the dependent variable to σLLP. In Column (4), we find that banks’ σLLP increases from 2018–2019 to 2020–2021, without interaction terms and control variables. In Column (5), we document that the increase in σLLP for CECL-adopters is significantly higher than that for non-CECL-adopters, without control variables. In Column (6), we add all control variables and find similar results as in Column (5). The influence of CECL on the increase in σALL (σLLP) from 2018–2019 to 2020–2021 is also economically significant. Using Columns (3) and (6) as an illustration, the increase in σALL (σLLP) for CECL-adopters from 2018–2019 to 2020–2021 is about 50% (100%) higher than the increase in σALL (σLLP) for non-CECL-adopters. ¹⁹ These results empirically document the CECL ramification: while the initial increase in CECL-adopters’ loan loss provisioning (and thus its volatility) is expected in 2020 as they begin to recognize lifetime expected credit losses following the policy adoption, the extended volatility increase in 2021 shows procyclicality because the volatility might not have increased as much in 2021 if those reversals had not been booked. In other words, the continuous increase in CECL-adopters’ ALL and LLP volatilities in 2021 is less likely to be triggered by another round of LLP recognition (unless there are universally major credit events or negative economic anticipation, which might not be the case at the time ²⁰ ), but more likely by LLP reversals. As discussed previously, such procyclical behavior, partly enabled by the flexibility in the accounting policy itself, conflicts with the primary purpose of CECL and weakens banks’ ability to absorb future credit losses in a timely manner if economic conditions worsen again after those LLPs have been released. Overall, our results of ALL and LLP volatilities are consistent with H1a.

OPEN IN VIEWER

Table 2.

ALL and LLP Volatility in Years 2020–2021 (8 Quarters) Compared to Years 2018–2019 (8 Quarters) for CECL versus Non-CECL Bank Holding Companies (DiD Design).

	(1)		(2)		(3)		(4)		(5)		(6)
	σALL		σALL		σALL		σLLP		σLLP		σLLP
POST	0.0013***	(11.04)	0.0008***	(6.31)	0.0004	(1.47)	0.0015***	(14.77)	0.0008***	(6.79)	0.0005**	(2.33)
POST × CECL			0.0009 ***	(4.07)	0.0006 ***	(3.09)			0.0012 ***	(6.45)	0.0010 ***	(6.18)
SIZE 2yr avg					0.0011	(1.19)					0.0008	(1.05)
σSIZE					−0.0013	(-0.59)					−0.0032*	(-1.81)
EBTP 2yr avg					−0.2586**	(-2.10)					−0.0029	(-0.05)
σEBTP					0.0285	(0.50)					0.0479	(1.19)
LOANS 2yr avg					0.0031	(1.15)					−0.0024	(-1.26)
σLOANS					0.0073*	(1.93)					0.0074**	(2.36)
NPL 2yr avg					−0.0034	(-0.77)					−0.0019	(-0.61)
σNPL					0.1064**	(2.25)					0.0854**	(2.57)
NCO 2yr avg					−0.2521***	(-2.90)					−0.1060	(-1.32)
σNCO					0.5924***	(2.78)					0.5324***	(4.76)
TIER1 2yr avg					0.0239**	(2.16)					0.0166**	(2.34)
PUBLIC					−0.0030***	(-3.01)					−0.0016**	(-2.24)
CMLNS 2yr avg					0.0011	(0.35)					−0.0054**	(-2.23)
CSLNS 2yr avg					0.0000	(0.00)					−0.0004	(-0.05)
RELNS 2yr avg					0.0043*	(1.91)					0.0020	(1.12)
GDP					0.0031***	(7.24)					0.0017***	(4.90)
CONSTANT	0.0007***	(12.42)	0.0007***	(12.75)	−0.0640***	(-3.97)	0.0004***	(8.33)	0.0004***	(8.91)	−0.0351***	(-2.86)
Bank FE	Yes		Yes		Yes		Yes		Yes		Yes
SE cluster	Bank		Bank		Bank		Bank		Bank		Bank
R-squared	0.70		0.72		0.82		0.71		0.75		0.84
N	534		534		496		534		534		496

This table reports the results of regressing σALL and σLLP on CECL with a difference-in-differences design. Column 1 reports the differences in σALL in the 2 years after 2020Q1 and 2 years before 2020Q1, without control variables. Column 2 shows the differences in the reported difference in Column 1 for CECL versus non-CECL banks, without control variables. Column 3 shows the results with all the control variables. Columns 4 to 6 repeat the tests in Columns 1 to 3, but with σLLP as the dependent variable. All models include bank fixed effects (with bank fixed effects, the individual effects of CECL are subsumed, and we find similar results with county fixed effects, which are not tabulated). All variables are defined in Appendix A. The coefficients on CECL are omitted due to multicollinearity brought by bank fixed effects. Continuous variables are winsorized at the 1 and 99 percentiles. t statistics are shown in parentheses. *p < .1, **p < .05, ***p < .01. Robust standard errors are clustered on banks. Bold fonts means the coefficients of interest.

The Volatility of Earnings

In this section, we report the results of the test of H1b, which examines the influence of CECL-adoption on the volatility of banks’ EBT and EBTP in a DiD design. Table 3 reports test results for H1b using σEBT and σEBTP as the dependent variables. In Table 3 Columns (1) to (3), we report the results using σEBT as the dependent variable with bank fixed effects. In Column (1), we regress σEBT on the dummy variable POST and find that, regardless of other control variables and CECL-adoption indicator, banks’ σEBT increases from 2018–2019 to 2020–2021. In Column (2), we regress σEBT on the interaction of POST and CECL with no other control variables. We document that the increase in σEBT for CECL-adopters is significantly higher than that for non-CECL-adopters. In Column (3), we add all control variables and find similar results as in Column (2). The influence of CECL on the increase in σEBT from 2018–2019 to 2020–2021 is also economically significant. Using Column (2) as an illustration, the increase in σEBT for CECL-adopters from 2018–2019 to 2020–2021 is about 50% higher than that for non-CECL-adopters. ²¹ H1b also predicts that, when σEBTP is used as the dependent variable, we should not observe a positive difference-in-differences effect of the interaction term POST × CECL because the main volatility contributor, LLP, is taken out of the measurement of earnings. We indeed find the expected results. In Columns (4) to (6), we repeat the analyses in Columns (1) to (3) but change the dependent variable to σEBTP—we do not find a significant positive coefficient on the interaction term of POST × CECL. Our results of EBT and EBTP volatilities suggest that the increase in CECL-adopters’ earnings volatility, as another unexpected CECL consequence, is essentially due to the volatility of LLPs.

OPEN IN VIEWER

Table 3.

Earnings Volatility in Years 2020–2021 (8 Quarters) Compared to Years 2018–2019 (8 Quarters) for CECL versus Non-CECL BHCs (DiD Design).

	(1)		(2)		(3)		(4)		(5)		(6)
	σEBT		σEBT		σEBT		σEBTP		σEBTP		σEBTP
POST	0.0011 ***	(6.30)	0.0006***	(2.62)	0.0001	(0.31)	0.0006 ***	(4.82)	0.0005***	(3.07)	0.0002	(1.30)
POST × CECL			0.0009 ***	(2.68)	0.0006 ***	(3.14)			0.0001	(0.24)	−0.0003 ***	(-3.20)
SIZE 2yr avg					−0.0002	(−0.24)					−0.0004	(−0.74)
σSIZE					−0.0035**	(−2.44)					0.0027**	(2.38)
LOANS 2yr avg					0.0022	(0.95)					−0.0011	(−0.69)
σLOANS					0.0040	(1.12)					−0.0028	(−1.18)
NPL 2yr avg					0.0064	(1.25)					−0.0022	(−0.72)
σNPL					−0.0025	(−0.06)					−0.0210	(−0.59)
NCO 2yr avg					−0.0591	(−0.71)					0.0409	(0.71)
σNCO					0.5298***	(3.26)					−0.2990***	(−2.87)
ALL2yr_avg					0.0001	(0.01)					0.0010	(0.12)
σALL					0.2265**	(2.25)					−0.1302*	(−1.86)
EBTP 2yr avg or EBT 2yr avg					−0.1823	(−1.18)					0.0603	(0.67)
σEBTP or σEBT					1.3285***	(11.99)					0.5998***	(10.78)
TIER1 2yr avg					0.0092	(1.26)					−0.0073	(−1.22)
PUBLIC					−0.0015*	(−1.76)					0.0013**	(2.12)
CMLNS 2yr avg					−0.0028	(−1.06)					0.0011	(0.67)
CSLNS 2yr avg					−0.0003	(−0.06)					0.0013	(0.41)
RELNS 2yr avg					−0.0011	(−0.67)					0.0005	(0.39)
GDP					0.0016***	(3.57)					−0.0007***	(−2.75)
Constant	0.0010***	(10.98)	0.0010***	(11.11)	−0.0178*	(−1.80)	0.0009***	(15.22)	0.0009***	(15.21)	0.0161**	(2.17)
Bank FE	Yes		Yes		Yes		Yes		Yes		Yes
SE cluster	Bank		Bank		Bank		Bank		Bank		Bank
R−squared	0.68		0.69		0.94		0.74		0.74		0.95
N	534		534		496		534		534		496

This table reports the results of regressing σEBT (Columns 1–3) and σEBTP (Columns 4–6) on the adoption of CECL with a difference-in-differences design. Column 1 reports the differences in σEBT in the 2 years after 2020Q1 and 2 years before 2020Q1, without control variables. Column 2 shows the DiD test results without control variables. Column 3 shows the results with all the control variables. Columns 4 to 6 repeat the tests in Columns 1 to 3 but with σEBTP as the dependent variables. All models include bank fixed effects (with bank fixed effects, the individual effects of CECL are subsumed, and we find similar results with county fixed effects, which are not tabulated). All variables are defined in Appendix A. The coefficients on CECL are omitted due to multicollinearity brought by bank fixed effects. Continuous variables are winsorized at the 1 and 99 percentiles. t statistics are shown in parentheses. *p < .1, **p < .05, ***p < .01. Robust standard errors are clustered on banks. Bold fonts means the coefficients of interest.

Robustness Checks for the Main Empirical Results

Correction for Selection Bias

Since some smaller banks adopted CECL earlier and some public banks delayed their adoption, our baseline DiD model may be subject to a selection bias. To control for this bias, we use the Heckman (1979) two-stage approach and propensity score matching.

First, using the Heckman approach, we model the selection of CECL-adoption by estimating a probit model with observable bank characteristics related to the choice of CECL-adoption. We then compute an inverse Mills ratio (LAMBDA) for each sample bank and include LAMBDA in our baseline DiD model. The results are reported in Table 4 Panel A. Column (1) shows that the first-stage regression explains about 45% of the variation in CECL-adoption, and that most bank characteristics, such as bank size, earnings, total loans, non-performing loans, tier-1 capital ratios, public versus private status, loan composites, and GDP are strong predictors. The second-stage results are reported in Columns (2) and (3), with LAMBDA as an additional control variable – in both regressions, we find that even after controlling for the selection bias, the baseline DiD effects remain positive and significant.

OPEN IN VIEWER

Table 4.

Correction for Selection Bias.

Panel A: Heckman 2-Stage Selection Model
	(1) First Stage		(2) Second Stage, Controlling for LAMBDA		(3) Second Stage, Controlling for LAMBDA
	CECL		σALL		σLLP
POST			0.0003	(1.08)	0.0005**	(2.21)
POST × CECL			0.0007 ***	(3.09)	0.0010 ***	(5.25)
LAMBDA			−0.0010	(−0.79)	0.0001	(0.17)
SIZE 2yr avg	0.5834***	(22.24)	0.0010	(1.05)	0.0008	(1.07)
σSIZE	1.4287**	(2.20)	−0.0019	(−0.83)	−0.0031*	(−1.71)
EBTP 2yr avg	98.5474***	(5.72)	−0.3340*	(−1.84)	0.0061	(0.08)
σEBTP	25.0112	(1.58)	0.0016	(0.02)	0.0511	(1.13)
LOANS 2yr avg	−1.2485***	(−4.59)	0.0045	(1.32)	−0.0026	(−1.20)
σLOANS	3.1146**	(2.51)	0.0056	(1.19)	0.0076**	(2.32)
NPL 2yr avg	−8.7491***	(−2.77)	0.0030	(0.30)	−0.0027	(−0.54)
σNPL	−28.8473*	(−1.76)	0.1282**	(2.51)	0.0828**	(2.30)
NCO 2yr avg	4.3423	(0.51)	−0.2507***	(−2.98)	−0.1061	(−1.31)
σNCO	150.6771***	(4.73)	0.5261**	(2.52)	0.5403***	(4.32)
TIER1 2yr avg	−4.8914***	(−5.46)	0.0293**	(2.18)	0.0160*	(1.85)
PUBLIC	1.5861***	(28.48)	−0.0040**	(−2.26)	−0.0014	(−1.36)
CMLNS 2yr avg	0.9529***	(2.74)	0.0010	(0.32)	−0.0054**	(−2.22)
CSLNS 2yr avg	−1.9868***	(−5.76)	0.0020	(0.26)	−0.0006	(−0.08)
RELNS 2yr avg	1.2096***	(5.11)	0.0043*	(1.86)	0.0020	(1.13)
GDP	−0.0664***	(−2.71)	0.0033***	(5.82)	0.0017***	(4.33)
CONSTANT	−8.9652***	(−12.87)	−0.0656***	(−3.97)	−0.0350***	(−2.79)
R-squared	0.45		0.82		0.84
Bank FE	No		Yes		Yes
N	4073		478		478

Panel B: Propensity Score Matching
Model	(1)Nearest-Neighbor Matching (N = 1)	(2)Nearest-Neighbor Matching (N = 1)	(3)Nearest-Neighbor Matching (N = 1)	(4)Nearest-Neighbor Matching (N = 1)
Outcome Var.	σALL	σALL	σLLP	σLLP
Subsample	Years 2020–2021	Years 2018–2019	Years 2020–2021	Years 2018–2019
Average Treatment Effect for the Treated (ATT):
CECL	0.00067*** (t = 2.88)	−0.00006 (t = −0.38)	0.00066*** (t = 2.88)	−0.00002 (−0.18)
Control Var.	Yes	Yes	Yes	Yes
Bank FE	No	No	No	No
Year FE	No	No	No	No
Standard error	Abadie & Imbens (2006)	Abadie & Imbens (2006)	Abadie & Imbens (2006)	Abadie & Imbens (2006)
Pseudo R-squared	0.3041	0.3035	0.3041	0.3035
N	251	265	251	265

Model	(5)Epanechnikov Kernel	(6)Epanechnikov Kernel	(7)Epanechnikov Kernel	(8)Epanechnikov Kernel
Outcome Var.	σALL	σALL	σLLP	σLLP
Subsample	Years 2020–2021	Years 2018–2019	Years 2020–2021	Years 2018–2019
Average Treatment Effect for the Treated (ATT):
CECL	0.00063** (t = 2.02)	−0.00002 (t = −0.20)	0.00067** (t = 2.54)	−0.00000 (−0.05)
Control Var.	Yes	Yes	Yes	Yes
Bank FE	No	No	No	No
Year FE	No	No	No	No
Standard error	Bootstrapped	Bootstrapped	Bootstrapped	Bootstrapped
Pseudo R-squared	0.3041	0.3035	0.3041	0.3035
N	251	265	251	265

Table 4 Panel A reports the results of Heckman (1979) two-stage approach. Column 1 estimates a probit selection model using observable bank characteristics potentially related to CECL-adoption choices. The coefficients on CECL are omitted due to multicollinearity brought by bank fixed effects. The parameters from the first-stage model are then used to compute an inverse Mills ratio for each sample bank holding company (the inverse Mills ratio is denoted as LAMBDA). Column 2 and 3 reports the second-stage test results, which include LAMBDA in the baseline DiD model. Column 2 uses σALL as the dependent variable and Column 3 uses σLLP as the dependent variable. All variables are defined in the Appendix A. Continuous variables are winsorized at the 1 and 99 percentiles. t statistics are shown in parentheses. *p < .1, **p < .05, ***p < .01. Robust standard errors are clustered on banks. Bold fonts means the coefficients of interest.

Panel B reports results using various setups of propensity score matching. CECL is the status of receiving “treatment.” It equals 1 if a bank is a CECL-adopter during 2020 and 2021, and zero otherwise. The outcome variables are σALL in Columns 1, 2, 5, and 6, and σLLP in columns 3, 4, 7, and 8. In Columns 1 to 4, CECL-adopters and non-CECL-adopters are 1-on-1 matched based on bank-level characteristics (i.e., the bank-level control variables in the baseline regression model). In Columns 1 to 4, standard errors are Abadie and Imbens (2006) robust. In Columns 5–8, we repeat the test in the first four columns by employing the Epanechnikov kernel matching for CECL-adopters and non-CECL-adopters. In Columns 1, 3, 5, and 7, the subsample of years 2020–2021 is used, while in Columns 2, 4, 6, and 8, the subsample of years 2018–2019 is used. The average treatment effects for the treated (ATT) are reported in this table. For all models, we impose common support by dropping the 10% of the treatment observations at which the p-score density of the control observations is the lowest. Propensity scores are calculated using the logit function. We use a caliper of 0.25 for the nearest-neighbor matchings. All bank-level continuous variables are winsorized at 1% and 99%. *p < .1, **p < .05, ***p < .01.

Second, we perform propensity score matching (PSM) to test the treatment effect of CECL-adoption. To preserve the merit of a DiD model, we performed PSM analysis for the subsamples containing observations in the years 2020–2021 and the years 2018–2019 separately. Our prediction is that the treatment effect of CECL-adoption shall be significant and positive only for the subsample containing observations in years 2020–2021, but not observations in years 2018–2019. Table 4 Panel B reports the PSM analysis results. In Columns (1) to (4), CECL-adopters and non-CECL-adopters are matched based on bank-level characteristics using Abadie and Imbens (2006) robust standard errors. Column 1 shows a positive and significant treatment effect for the 2020–2021 subsample with the volatility of loan loss allowances (ΔALL) as the outcome variable. Column 2 shows no significant treatment effect for the 2018–2019 subsample. In Columns (3) and (4), we repeat the analyses in Columns (1) and (2) by replacing the outcome variable with the volatility of loan loss provisions (ΔLLP), and we find similar results. In Columns (5) to (8), we repeat the PSM analysis using Epanechnikov kernel matching with a bootstrapped standard error – we find consistent results. In sum, the PSM analysis confirms that the treatment effect of CECL-adoption is significant and positive only in 2020–2021, aligning with our expectations.

Alternative Model Specifications and Other Selection Issues

We further conduct an array of robustness checks and address endogeneity concerns for the main empirical results, including alternative model specifications, placebo tests to validate the parallel trends assumption, results without banks that are subsidiaries of a foreign bank from IFRS countries, and the incorporation of the moderating effect of acquired loans and unused commitments. As discussed in Appendix OA1 and reported in Table OA2, Table OA3, and Table OA4 of the Online Appendix, our main results remain robust under all these model specifications.

LLP Timeliness and Validity Tests

Existing studies have well documented that under the traditional incurred credit loss model, LLPs should be explained by subsequent periods’ changes in non-performing loans and net charge-offs. Beatty and Liao (2014) examine nine LLP models in a factor analysis and document that LLPs can be explained by variables such as the current, subsequent, prior periods’ changes in non-performing loans, lagged bank size, lagged ALL, current period’s net charge-offs, change in loans, and change in macroeconomic conditions. ²² Moreover, other studies such as Beatty and Liao (2011) and Bushman and Williams (2015) use the incremental explanatory power of subsequent periods’ changes in non-performing loans as a measurement of increased timeliness of LLPs, measured as the difference in R-squared between an LLP model with, and an LLP model without, subsequent periods’ changes in non-performing loans in a 12-quarter rolling window. However, the measurement in Beatty and Liao (2011) and Bushman and Williams (2015) requires banks to have data for at least 12 quarters. Considering this limitation, existing studies such as Nicoletti (2018) directly use the coefficient of subsequent periods’ changes in non-performing loans as the proxy for the timeliness of LLPs. In addition, Altamuro and Beatty (2010) argue that LLP validity can be measured as the association between the current period’s LLPs and subsequent periods’ net loan charge-offs. Based on these existing studies, we test whether CECL-adopters show more timeliness and validity in their LLPs after the CECL rule became effective in 2020q1.

We present the results of timeliness and validity tests of bank LLPs using a DiD design in Table 5. In Column (1), we find that the coefficient on the three-way interaction (POST × CECL × ΔNPL _t ) is positive and significant at times t and t+2. This supports the hypothesis that the difference in LLP timeliness after versus before the beginning of 2020 is significantly greater for CECL-adopters than for non-CECL-adopters. Similarly, in Column (2), the coefficient on the three-way interaction (POST × CECL × NCO _t+2/t ) is also positive and significant, validating our prediction that the change in LLP validity after versus before the start of 2020 is significantly higher for CECL-adopters than for non-CECL-adopters. Therefore, while CECL-adopters exhibit increased volatility in their loan loss provisioning, our empirical evidence suggests that CECL-adopters’ loan loss provisioning is both more timely and more valid than non-CECL-adopters after the CECL rule became effective. In Appendix OA2, we continue to perform subsample analyses of LLP timeliness and validity by dividing the full sample into the 2018–2019 and 2020–2021 bank-quarters, to further examine banks’ pre- and post-CECL loan loss provisioning patterns. The results in Table OA5 demonstrate that those “would-be” CECL-adopters do not show significantly higher loan loss provisioning timeliness or validity for the subsample with bank-quarters in 2018–2019, whereas actual CECL-adopters exhibit significantly higher timeliness or validity for the subsample with bank-quarters in 2020–2021, consistent with the DiD inferences.

OPEN IN VIEWER

Table 5.

LLP Timeliness and Validity Tests—Full Sample Tests With DiD Specification (3-Way Interactions).

	(1)		(2)
	LLP		LLP
POST × CECL × ΔNPL t	0.133 **	(2.10)
POST × CECL × ΔNPL t+1	0.079	(1.42)
POST × CECL × ΔNPL t+2	0.089 *	(1.72)
POST × CECL × NCO t+2/t			0.094 ***	(2.98)
POST × ΔNPL	0.034	(0.74)
CECL × ΔNPL	0.004	(0.17)
POST × ΔNPL t+1	0.037	(1.21)
POST × CECL	0.001***	(4.12)	0.000**	(2.19)
CECL × ΔNPL t+1	0.021	(0.87)
POST × ΔNPL t+2	0.017	(0.55)
CECL × ΔNPL t+2	0.017	(0.98)
POST × NCO t+2/t			0.048***	(3.82)
CECL × NCO t+2/t			0.030	(0.64)
POST	0.001***	(6.17)	0.001***	(5.41)
ΔNPL	−0.021	(−1.24)
ΔNPL t+1	−0.013	(−1.23)
ΔNPL t+2	−0.004	(−0.36)
NCO t+2/t			0.116***	(2.95)
SIZE	−0.002***	(−6.15)	−0.001***	(−5.12)
EBTP	0.039	(1.63)	0.050**	(2.22)
ΔLOANS	0.001	(0.94)	0.005***	(3.45)
ALL t−1	−0.124***	(−8.52)	−0.148***	(−5.77)
TIER1	−0.018***	(−4.43)	−0.013***	(−3.59)
NCO	0.196***	(6.41)	0.203***	(5.90)
NPL	0.036**	(2.19)	0.021	(1.50)
CMLNS	0.007***	(3.24)	0.007**	(2.57)
CSLNS	0.005*	(1.97)	0.005**	(2.10)
RELNS	0.003*	(1.89)	0.003**	(2.06)
PUBLIC	0.001***	(2.72)	0.001***	(3.01)
GDPGROWTH	−0.012***	(−8.71)	−0.009***	(−6.72)
CONSTANT	0.027***	(5.66)	0.020***	(4.66)
Bank FE	Yes		Yes
SE cluster	Bank		Bank
R2	0.65		0.72
N	4139		4156

This table reports the results of testing the influence of CECL on the timeliness and validity of LLP with a difference-in-differences design. The timeliness of LLP is measured as the relationship between the current quarter LLP and future changes in NPL. The validity of LLP is measured as the relationship between the current quarter LLP and future two quarters’ NCO. Column 1 presents the LLP timeliness tests and Column 2 reports the LLP validity tests. All variables are defined in Appendix A. Continuous variables are winsorized at the 1 and 99 percentiles. The coefficients on CECL are omitted due to multicollinearity brought by bank fixed effects. t statistics are shown in parentheses. *p < .1, **p < .05, ***p < .01. Robust standard errors are clustered on banks. Bold fonts means the coefficients of interest.

Do LLPs under CECL Contain More Information than LLPs under ICL?

Following LLP timeliness and validity tests, we investigate whether CECL-adopters are more likely to incorporate more information about their current loan portfolios beyond the current and the immediate future periods by examining “abnormal” LLPs through a two-stage LLP regression model. Appendix OA3.1 and Table OA6 of the Online Appendix show that CECL-adopters recognize significantly more abnormal LLPs than non-CECL-adopters after 2020. In addition, we examine the influence of CECL-adoption on the information content of LLPs by controlling for determinants of the normal LLP level directly through a single-stage LLP regression model. ²³ Given that the discretionary portion of CECL-adopters’ LLPs can be positive in 2020 and negative in 2021, we have utilized two-year dummies to interact with the CECL-adoption variable. The related empirical results are shown in Table 6. We use an LLP model following Kanagaretnam et al. (2010) and Beatty and Liao (2014) and add two interaction terms, namely YEAR2020 × CECL and YEAR2021 × CECL, into the LLP model. YEAR2020 (YEAR2021) is a dummy variable equal to one if the observation is in year 2020 or 2021, and equal to zero if the observation is in year 2018 or 2019. Column (1) reports results without bank fixed effects and Column (2) reports results with bank fixed effects. We find that in both models, the interaction term YEAR2020 × CECL is positive and significant, and the interaction term YEAR2021 × CECL is negative and significant. These results are consistent with the two-stage LLP regression results, and further indicate that CECL-adopters incorporate significantly more positive (negative) discretionary LLP in year 2020 (2021) compared to in the years 2018–2019. We also provide a subsample analysis of discretionary LLPs (with a single-stage model) based on loan portfolio composition in Appendix OA3.2 and Table OA7 to increase the robustness of this finding.

OPEN IN VIEWER

Table 6.

Discretionary LLP Tests (Single-Stage Model).

	(1)		(2)
	LLPQ		LLPQ
YEAR2020	0.0007***	(7.40)	0.0006***	(5.32)
YEAR2021	−0.0001	(−1.31)	−0.0004***	(−3.50)
CECL	0.0001	(1.04)
YEAR2020 × CECL	0.0006 ***	(4.80)	0.0007 ***	(4.92)
YEAR2021 × CECL	−0.0007 ***	(−5.69)	−0.0007 ***	(−4.89)
ΔLOANS	0.0040***	(5.61)	0.0046***	(3.71)
NCO	0.3590***	(43.64)	0.0993***	(2.77)
ΔNPL	0.0555***	(5.56)	0.0422***	(2.71)
ΔNPL t+1	0.0375***	(3.81)	0.0262*	(1.68)
ΔNPL t−1	0.0473***	(4.78)	0.0361***	(3.41)
ΔNPL t−2	0.0234**	(2.33)	0.0333*	(1.87)
SIZE	−0.0001***	(−3.77)	0.0005	(1.65)
CMLNS	0.0006**	(2.27)	0.0019**	(1.98)
CSLNS	0.0020***	(8.48)	0.0040	(1.43)
RELNS	0.0002	(0.93)	0.0011	(1.16)
GDPGROWTH	−0.0125***	(−15.36)	−0.0117***	(−9.13)
CONSTANT	0.0010***	(2.60)	−0.0092*	(−1.77)
R-squared	0.58		0.65
Bank FE	No		Yes
SE cluster	Bank		Bank
N	4315		4311

This table reports the single-stage LLP model regression results. The model follows Kanagaretnam et al. (2010) and Beatty and Liao (2014). YEAR2020 is a dummy variable equal to one if the observation is in year 2020, and equal to zero if the observation is in year 2018 or 2019. YEAR2021 is a dummy variable equal to one if the observation is in year 2021, and equal to zero if the observation is in year 2018 or 2019. Column 1 reports results without bank fixed effects and Column 2 reports results with bank fixed effects. In Column 2, the coefficients on CECL are omitted due to multicollinearity brought by bank fixed effects. All variables are defined in Appendix A. Continuous variables are winsorized at the 1 and 99 percentiles. t statistics are shown in parentheses. *p < .1, **p < .05, ***p < .01. Robust standard errors are clustered on banks. Bold fonts means the coefficients of interest.

Earnings Response Coefficients (Market Perceptions) Tests

In this section, we investigate the stock market perceptions of earnings information of CECL-adopters after versus before CECL-adoption, compared to non-CECL-adopters. We measure the market perceptions of earnings informativeness with the earnings response coefficients (ERC), which has been used extensively in prior studies as a proxy for the perceived credibility of earnings information (e.g., Collins & Kothari, 1989; DeFond & Zhang, 2014; Teoh & Wong, 1993). ERC is constructed as the coefficient of unexpected earnings (UE) and the three-day cumulative abnormal returns (CAR) around the earnings announcement dates, where UE is calculated as the difference between actual EPS and I/B/E/S most recent consensus EPS forecast, and CAR is calculated as the cumulative three-day returns on top of the CRSP value-weighted market portfolio returns (CAR _VW ). ²⁴

Because banks following the new CECL rule are mandated to incorporate more forward-looking information in estimating LLPs, which are the largest accruals on banks’ income statement, we expect that ERC is more positive (or less negative) for CECL-adopters than non-CECL-adopters, after versus before 2020q1. However, due to the large and hasty LLP reversals in 2021, it is also plausible that market may see through the reversals and, if any, the potentially opportunistic behavior. Therefore, we predict that the stock market may credit the CECL rule for enhancing earnings informativeness in the first year of CECL-adoption (i.e., year 2020), but less so in the second year of CECL-adoption (i.e., year 2021).

We conduct two sets of capital market tests. In the first set of tests, we examine whether and how the change in ERC for CECL-adopters from 2019 to 2020 differs from the corresponding change in ERC for non-CECL-adopters. The results are reported in Table 7 Panel A, Columns (1) to (4). In Column (1), we report the regression results with UE proxied as the difference between actual EPS and the mean of the most recent EPS forecasts (i.e., UE_Mean) and without the treatment effect of CECL-adoption and control variables—we document that on average, earnings surprise (UE) is positively related to cumulative abnormal returns in the three-day window around earnings announcement dates (CAR _VW ) and that public banks’ ERC do not change significantly from 2019 to 2020. In Column (2), we add the treatment effect of CECL-adoption and find that the change in ERC for CECL-adopters is significantly more positive than the change in ERC for non-CECL-adopters. ²⁵ In Columns (3), we add all control variables and bank fixed effects based on Column (2), and we find that the treatment effect of CECL-adoption remains positive and significant. In Column (4), we repeat the analysis in Column (3) but changing the proxy for UE to the difference between actual EPS and the median of the most recent EPS forecasts (i.e., UE_Median), and we find comparable results. In summary, the first set of tests suggests that the market tends to perceive the earnings reported by CECL-adopters as more informative than those reported by non-CECL-adopters in the year 2020.

OPEN IN VIEWER

Table 7.

Earnings Response Coefficients Tests (DiD Design).

Panel A: Subsample Tests
	(1)	(2)	(3)	(4)	(5)	(6)
Proxy for CAR	CAR VW	CAR VW	CAR VW	CAR VW	CAR VW	CAR VW
Post-CECL Indicator:	POST_2020 versus 2019				POST_2021 versus 2019
Proxy for UE	UE_Mean	UE_Mean	UE_Mean	UE_Median	UE_Mean	UE_Median
UE	2.546**	13.025***	12.327***	12.371***	17.097***	17.500***
	(2.02)	(3.69)	(3.40)	(3.36)	(2.90)	(2.97)
POST-CECL	−0.019***	0.011	−0.002	0.001	0.031	0.036
	(−5.15)	(0.69)	(−0.19)	(0.11)	(1.31)	(1.54)
UE × POST-CECL	−1.032	−12.614***	−11.924***	−11.862***	−11.460*	−12.156*
	(−0.80)	(−3.55)	(−3.54)	(−3.49)	(−1.72)	(−1.71)
UE × POST-CECL × CECL		12.518 ***	11.088 ***	11.574 ***	9.046	9.948
		(3.44)	(2.68)	(2.79)	(1.33)	(1.37)
CECL		0.007
		(0.57)
POST-CECL × CECL		−0.031*	−0.027	−0.026	−0.044**	−0.049**
		(−1.78)	(−1.47)	(−1.39)	(−2.35)	(−2.58)
UE × CECL		−11.285***	−9.084**	−9.807**	−11.898**	−12.362**
		(−3.16)	(−2.16)	(−2.33)	(−2.04)	(−2.12)
SIZE			−0.042	−0.050	−0.097***	−0.098***
			(−1.11)	(−1.34)	(−2.66)	(−2.67)
ΔLOANS			−0.009	−0.015	0.058	0.059
			(−0.17)	(−0.27)	(1.21)	(1.24)
ΔNPL			−3.849**	−3.473*	2.829	2.788
			(−2.26)	(−1.93)	(1.54)	(1.49)
LLP_YR			0.605	0.299	0.134	0.060
			(0.31)	(0.14)	(0.07)	(0.03)
ALL t−1			−0.368	−0.265	1.103	1.155
			(−0.37)	(−0.25)	(0.86)	(0.89)
TIER1			−0.259	−0.252	0.311	0.299
			(−0.67)	(−0.67)	(0.69)	(0.66)
NCO			−2.251	−2.064	0.833	1.132
			(−0.92)	(−0.79)	(0.20)	(0.27)
NPL			2.256**	2.162**	−0.991	−0.810
			(2.25)	(2.13)	(−0.53)	(−0.43)
CMLNS			0.083	0.042	−0.202	−0.211
			(0.72)	(0.34)	(−1.16)	(−1.18)
CSLNS			0.199	0.171	0.301	0.280
			(1.00)	(0.87)	(1.16)	(1.08)
RELNS			−0.188	−0.173	0.149	0.150
			(−1.63)	(−1.51)	(1.29)	(1.27)
Bank FE	No	No	Yes	Yes	Yes	Yes
R-squared	0.20	0.25	0.75	0.74	0.65	0.65
N	240	240	220	220	210	210

Panel B: Full Sample Tests
	(1)		(2)		(3)		(4)
Proxy for CAR	CAR VW		CAR VW		CAR VW		CAR VW
Proxy for UE	UE_Mean		UE_Mean		UE_Mean		UE_Median
UE	2.546**	(2.02)	13.025***	(3.68)	14.101***	(3.27)	14.074***	(3.26)
POST_2020 versus 2019	−0.019***	(−5.15)	0.011	(0.69)	0.008	(0.44)	0.014	(0.80)
POST_2021 versus 2019	−0.017***	(−2.74)	0.012	(0.64)	0.027	(1.53)	0.033**	(1.99)
POST_2020 versus 2019 × UE	−1.032	(−0.80)	−12.614***	(−3.54)	−12.840***	(−3.08)	−13.097***	(−3.18)
POST_2021 versus 2019 × UE	−0.413	(−0.29)	−9.711**	(−2.17)	−6.664	(−1.25)	−7.142	(−1.22)
CECL			0.007	(0.50)
CECL × UE			−11.285***	(−3.02)	−10.111**	(−2.23)	−10.581**	(−2.34)
CECL × POST_2020 versus 2019			−0.031*	(−1.92)	−0.022	(−1.13)	−0.026	(−1.34)
CECL × POST_2021 versus 2019			−0.030	(−1.51)	−0.041**	(−2.43)	−0.044***	(−2.76)
CECL × POST_2020 versus 2019 × UE			12.518 ***	(3.33)	11.107 **	(2.53)	11.896 ***	(2.75)
CECL × POST_2021 versus 2019 × UE			10.106 **	(2.16)	4.102	(0.74)	5.051	(0.84)
SIZE					−0.053	(−1.62)	−0.060*	(−1.83)
ΔLOANS					0.027	(0.48)	0.022	(0.39)
ΔNPL					1.076	(0.93)	1.126	(0.96)
LLPY					−0.467	(−0.62)	−0.401	(−0.52)
ALL t−1					−0.111	(−0.18)	−0.196	(−0.31)
TIER1					0.109	(0.40)	0.155	(0.55)
NCO					−3.087	(−1.55)	−3.251	(−1.62)
NPL					0.351	(0.31)	0.509	(0.44)
CMLNS					−0.013	(−0.10)	−0.044	(−0.34)
CSLNS					0.141	(0.80)	0.121	(0.70)
RELNS					−0.004	(−0.04)	−0.006	(−0.06)
GDPGROWTH					0.854	(1.34)	0.856	(1.33)
CONSTANT	−0.009***	(−2.94)	−0.017	(−1.22)	0.856	(1.48)	0.976*	(1.69)
Bank FE	No		No		Yes		Yes
R-squared	0.13		0.17		0.54		0.54
N	359		359		331		331

This table reports the results of regressing the cumulative abnormal returns around earnings announcement dates on unexpected earnings (UE) for CECL- versus non-CECL-adopters, before and after 2020Q1 in a difference-in-differences setting. Columns 1 to 4 report results with POST_2020 versus 2019 as the POST-CECL indicator. Column 5 and 6 report results with POST_2021 versus 2019 as the POST-CECL indicator. Columns 1 reports only results with two-way interactions and shows the change in earnings responsiveness without considering CECL/non-CECL classifications. Column 2 reports the DiD regression without control variables. Columns 3 and 4 report the DiD regression results with control variables and bank fixed effects. All variables are defined in Appendix A. In Columns 3 to 6, the coefficients on CECL are omitted due to multicollinearity brought by bank fixed effects. Continuous variables are winsorized at the 1 and 99 percentiles. t statistics are shown in parentheses. *p < .1, **p < .05, ***p < .01. Robust standard errors are clustered on banks for all models. Constants and macroeconomic controls are not reported to preserve space.

This table reports the results of regressing the cumulative abnormal returns around earnings announcement dates on unexpected earnings (UE) for CECL- versus non-CECL-adopters, before and after 2020Q1 in a difference-in-differences setting. POST_2020 versus 2019 (POST_2021 versus 2019) is a dummy variable equal to 1 if the observation is in year 2020 (2021), and 0 if in 2019. In Columns 3 and 4, the coefficients on CECL are omitted due to multicollinearity brought by bank fixed effects. All variables are defined in Appendix A. Continuous variables are winsorized at the 1 and 99 percentiles. t statistics are shown in parentheses. *p < .1, **p < .05, ***p < .01. Robust standard errors are clustered on banks for all models. Constants and macroeconomic controls are not reported to preserve space. Bold fonts means the coefficients of interest.

In the second set of tests, we examine whether and how the change in ERC for CECL-adopters from 2019 to 2021 differs from that for non-CECL-adopters. The results, reported in Columns (5) and (6) of Table 7 Panel A, do not show a significant DiD coefficient (CECL × POST2021vs2019 × UE), suggesting that the CECL impact on earnings informativeness is not significant in 2021. In summary, the market perception tests support our prediction that the market initially credited the CECL rule for enhancing earnings informativeness, but the large reversals in 2021 seem to have reduced the perceived merit of CECL.

To provide a more integrated vision of the CECL effect on banks’ ERC, we also directly examine the 3-way interactions between CECL-adoption, UE, and years 2020 and 2021 (relative to year 2019) in one regression using the full sample. Table 7 Panel B reports the results. While the coefficient on (CECL × YEAR2021 × UE) is significantly positive when control variables are not included in Column (2), the results in Columns (3) and (4) with control variables included are consistent with those in the above subsample tests: the coefficient on (CECL × YEAR2020 × UE) is significant and positive, whereas the coefficient on (CECL × YEAR2021 × UE) is statistically insignificant, again indicating that CECL-adopters’ earnings are perceived as more informative in 2020 but less credited by the market in 2021. Overall, our results on multi-year interactions using the full sample and separate single-year interactions are similar and thus subject to minimal bias.

Earning Management Motivations of LLP Reversals

As aforementioned, banks’ incentive to use LLP reversals to mask earnings shortfall may be amplified if managers ex-ante anticipate a higher ERC. In this section, we report the results of the empirical tests for H4, which states that CECL-adopters with larger EPS shortfalls below consensus EPS forecasts before provisions are more likely to recognize LLP reversals in 2021. We also examine whether such LLP reversals reduce the informativeness of bank earnings.

We use two models to test H4. In the first model, we use a logit regression to estimate the following equation with standard errors clustered on banks:

R E V E R S A L = β 0 + β 1 E P S _ S H O R T F A L L + β 2 Δ L O A N S + β 3 N C O t + β 4 Δ N P L t + β 5 Δ N P L t + 1 + β 6 Δ N P L t − 1 + Β 7 Δ NPL t − 2 + β 8 SIZE t − 1 + β 9 CMLNS t + β 10 CSLNS t + β 11 RELNS t + β 12 Δ GDP + ε

(4)

where the dependent variable, REVERSAL, is a dummy variable that equals 1 if the CECL-adopter recognizes an LLP reversal in 2021, and zero otherwise. EPS_SHORTFALL is calculated as the most recent mean consensus EPS forecast minus the banks’ actual EPS before LLPs. Hence, EPS_SHORTFALL measures the distance to benchmark earnings before booking any LLP reversals. All other variables are defined in Appendix A. We limit our sample period to year 2021 and limit the banks to only CECL-adopters. We expect that β ₁ is positive, indicating that the larger the EPS shortfall below benchmarks, the more likely the recognition of LLP reversals—hence, a positive β ₁ implies some opportunistic incentives. Table 8 Column (1) reports the results of estimating this logit model. As expected, we find a positive and significant coefficient on EPS_SHORTFALL, indicating that a $1 increase in EPS shortfall increases the log odds of recognizing LLP reversals by 5.53.

OPEN IN VIEWER

Table 8.

LLP Reversal Tests for CECL-Adopters in 2021.

	(1)	(2)	(3)
Dependent Var	REVERSAL = 1	LLP	LLP
Model	Logit	OLS	OLS
Subsample	Year = 2021 and CECL = 1	Year = 2021 and CECL = 1	Year = 2021 and CECL = 1 and LLP <0
EPS_SHORTFALL	5.528 ***	−0.002 ***	−0.002 ***
	(4.76)	(−3.47)	(−2.82)
ΔLOANS	−56.223***	0.003	−0.019
	(−2.97)	(0.58)	(−1.42)
NCO	−20.153**	0.936***	−0.021
	(−2.52)	(226.22)	(−0.14)
ΔNPL	−128.946	0.075	0.148
	(−0.46)	(0.69)	(0.91)
ΔNPL t+1	1180.992***	−0.092*	−0.119***
	(3.87)	(−1.74)	(−2.91)
ΔNPL t−1	−327.404	0.183	0.028
	(−0.84)	(0.81)	(0.10)
ΔNPL t−2	638.450*	0.245**	0.261***
	(1.79)	(2.09)	(2.70)
SIZE	0.327	−0.000	−0.000
	(0.81)	(−1.29)	(−1.32)
CMLNS	−1.972	−0.005*	−0.000
	(−0.82)	(−1.84)	(−0.03)
CSLNS	0.512	−0.013***	−0.000
	(0.12)	(−3.74)	(−0.10)
RELNS	4.937**	−0.002	−0.001
	(1.97)	(−0.90)	(−0.49)
GDPGROWTH	68.541	−0.064	−0.068**
	(0.88)	(−1.64)	(−2.04)
Constant	−6.879	0.008	0.004
	(−0.83)	(1.39)	(1.06)
AUC	0.95	n/a	n/a
R2	n/a	0.90	0.56
SE cluster	Bank	Bank	Bank
N	100	100	76

This table reports the results of regressions that examine the relationship between the 2021 LLP reversals and EPSBP (EPS before loan loss provisions). Column 1 reports the regression results of a logit model, where the dependent variable is a dummy variable (REVERSAL). Column 2 reports the regression results of an OLS model with LLP as the dependent variable. Column 3 reports the results of an OLS model with LLP as the dependent variable and limiting to only those observations with LLP <0. All models use a subsample of observations of only CECL-adopters with year = 2021. Fixed effects are not included due to small sample size (i.e., bank fixed effects would lead to insufficient observations problems). All variables are defined in Appendix A. Continuous variables are winsorized at the 1 and 99 percentiles. t statistics are shown in parentheses. *p < .1, **p < .05, ***p < .01. Robust standard errors are clustered on banks for all models. Bold fonts means the coefficients of interest.

In the second model, we change the dependent variable in equation (4) to LLP and use OLS regressions. The results are reported in Table 8 Columns (2) and (3). In Column (2), we limit our sample to year 2021 and CECL-adopters. We document that EPS shortfall before provisions is negatively related to LLP (coef. = −0.002, t statistic = −3.47), indicating that a higher level of ESP shortfall is associated with a more negative LLP. In Column (3), we further limit our sample to banks with a negative LLP (i.e., all banks in this sample recognize LLP reversals in 2021)—we find similar results. Overall, the results in the second model suggest that banks with more EPS shortfalls before provisions may have the incentive to recognize a higher level of LLP reversals. To rule out an alternative explanation that the analysts may have been inexperienced with CECL or guided by bank managers in certain directions, we run a new DiD test to examine whether the CECL-adoption is a significant driving force of analysts’ forecasting changes. We find that CECL-adoption drives analysts’ forecasts in a rather limited way, which alleviates the concern that EPS forecasts is biased with the CECL rule. We discuss different scenarios under this explanation in detail in Appendix OA4 and report the results in Table OA8 of the Online Appendix.

Another alternative explanation stems from the inexperience of mangers themselves with CECL. Although our results point to management bias as a source of banks’ overprovisioning in 2020 and reversals in 2021, management error, on the other hand, could lead to a similar pattern, rendering the finding on managers’ opportunism less convincing. We mitigate this concern by controlling for banks’ ex-ante ability (i.e., pre-COVID) to model LLPs to proxy for the potential variation in managers’ experience with CECL and ensure that the effect of management error on banks’ reversals is limited. We discuss it in detail in Appendix OA5 and report the results in Table OA9. Overall, our conclusion about managers’ opportunism in booking reversals is unchanged.

In addition, focusing on year 2021, we examine more closely whether LLP reversals reduce the informativeness of bank earnings. Table 9 presents the related findings: Column (1) reports a positive and significant coefficient on UE, and a negative and significant coefficient on REVERSAL × UE, indicating that banks reporting LLP reversals have significantly weaker earnings response coefficients in 2021. In Column (2), we replace the interaction term with CECL × UE and find that CECL-adopters also show much weaker earnings response coefficients. These results suggest that the market might indeed see through the reversals and opportunistic behavior of CECL-adopters, and thus place less reliance on the information conveyed in their earnings.

OPEN IN VIEWER

Table 9.

The Influence of LLP Reversals on Earnings Response Coefficients Tests in 2021.

	(1)		(2)		(1)		(2)
	CAR VW		CAR VW		CAR VW		CAR VW
Subsample Year:	2021		2021		2021		2021
REVERSAL × UE	−3.759 **	(−2.53)			−6.064 ***	(−2.88)
CECL × UE			−4.660 ***	(−2.82)			−5.302 *	(−1.96)
UE	4.472***	(3.81)	5.456***	(3.56)	6.216***	(3.87)	5.798**	(2.46)
CECL	−0.022	(−1.43)	−0.018	(−1.41)	−0.013	(−0.74)	−0.006	(−0.39)
REVERSAL	−0.012	(−1.15)	−0.015	(−1.51)	−0.004	(−0.32)	−0.010	(−0.77)
SIZE	−0.003	(−0.55)	−0.004	(−0.85)	−0.003	(−0.43)	−0.005	(−0.67)
ΔLOANS	−0.043	(−0.42)	0.021	(0.23)	0.054	(0.36)	0.098	(0.65)
ΔNPL	5.702**	(2.35)	4.474**	(2.10)	7.503***	(3.14)	6.069**	(2.50)
LLP_YR	−5.955***	(−5.30)	−5.469***	(−5.08)	−6.638***	(−4.48)	−5.987***	(−4.22)
ALL t−1	1.153	(1.06)	1.328	(1.31)	1.426	(1.04)	1.701	(1.30)
TIER1	0.366	(1.60)	0.319	(1.50)	0.564*	(1.95)	0.474*	(1.71)
NCO	−3.123	(−1.17)	−3.479	(−1.31)	−3.823	(−1.14)	−4.215	(−1.24)
NPL	−1.743**	(−2.01)	−1.670**	(−2.22)	−1.566*	(−1.70)	−1.547*	(−1.67)
CMLNS	0.053	(0.93)	0.045	(0.80)	0.098	(1.18)	0.081	(1.00)
CSLNS	0.031	(0.48)	0.026	(0.42)	0.015	(0.16)	0.007	(0.08)
RELNS	0.040	(1.02)	0.041	(1.09)	0.047	(0.76)	0.049	(0.78)
GDPGROWTH	0.337	(0.34)	0.394	(0.41)	0.000	(.)	0.000	(.)
CONSTANT	−0.047	(−0.44)	−0.019	(−0.17)	−0.085	(−0.53)	−0.042	(−0.27)
Fixed effects	No		No		State		State
R2	0.39		0.42		0.55		0.54
N	101		101		101		101

This table reports results of capital market tests examining the influence of LLP reversals on earnings response coefficients, which focus on observations in year 2021. Columns 1 and 2 show results with no fixed effects and Columns 3 and 4 show results with state fixed effects. County or bank fixed effects are not included because doing so will result in insufficient observations due to small sample size. All variables are defined in Appendix A. Continuous variables are winsorized at the 1 and 99 percentiles. t statistics are shown in parentheses. *p < .1, **p < .05, ***p < .01. Robust standard errors are clustered on banks. Bold fonts means the coefficients of interest.

Finally, we test whether post-adoption analyst forecast accuracy improves, to reinforce the information quality implications of CECL. In Appendix OA7 and Table OA11, we document that CECL-adopters exhibit significantly smaller earnings forecast errors in 2020 relative to 2019, but such improvement does not persist into 2021, consistent with our findings on diminished earnings informativeness associated with LLP reversals.

Concluding Remarks

The CECL model was introduced by the FASB to provide more timely loan loss provisions and reduce banks’ procyclical behavior under the ICL model. The unforeseen onset of COVID-19, coinciding with the partial adoption of CECL, presents a natural experimental setting for examining the effectiveness of CECL during highly uncertain economic conditions, by comparing CECL-adopters and non-CECL-adopters.

We examine the differences in loan loss provisioning patterns between CECL-adopters and non-CECL-adopters, assess the timeliness and validity of LLPs under the CECL model, and explore the capital market impacts and potential motivations for the LLP reversals booked by CECL-adopters in 2021. Our results indicate that the increase in LLP and ALL volatilities is significantly higher for CECL-adopters from 2018–2019 to 2020–2021, and similar findings on EBT volatility but not on EBTP volatility also imply that banks’ increase in earnings volatility mainly originates from LLP volatility. Furthermore, we document that, compared to non-CECL-adopters’ LLPs, CECL-adopters’ LLPs are indeed more forward-looking and accurate predictors of future periods’ change in NPL and NCO, consistent with the mandate of CECL. Moreover, from a capital market standpoint, our findings show that the stock market acknowledges the more forward-looking earnings under CECL in 2020, but not in 2021, when large LLP reversals are observed. Finally, banks’ large and quick LLP reversals in 2021 associated with earnings shortfall suggest that they may have manipulated the LLP reversals to meet earnings benchmarks, which could be a sign of opportunistic behavior.

While our empirical evidence indicates management bias in those large LLP reversals, we acknowledge that it is challenging to strictly distinguish between management bias and error in explaining the source of the reversals, given the lack of experience with CECL for managers themselves. In other words, management error could also be manifested as overprovisioning in 2020 and large reversals in 2021, even with an earnings shortfall observed. However, controlling for banks’ ex-ante (i.e., pre-COVID) ability to model LLPs, which potentially reflects the propensity of management error, does not significantly change our conclusions, thus somewhat mitigating the concern about management error rather than management bias.

Our study contributes to the existing literature by examining the effects of the CECL model in the U.S. market, against the backdrop of the COVID-19 pandemic, and extending the emerging literature on expected credit loss accounting models over a longer post-policy period with more comprehensive policy consequences. Most importantly, while the intended CECL objective is building up more long-term and timely provisions and reducing banks’ procyclical behavior under ICL, our study finds that CECL-adopters may have exploited the flexibility in LLP estimation brought by the accounting policy itself for managerial opportunism, thus exhibiting procyclicality. In essence, while CECL delivers on its promise of providing more forward-looking and accurate credit loss predictions, it may inadvertently introduce unwanted volatility in banks’ loan loss provisions and enable banks’ procyclical behavior, which is ironically what it primarily aimed to address.

Supplemental Material

Supplemental Material