Current expected credit loss procyclicality: it depends on the model

ABSTRACT

The new guidelines for loan loss reserves, current expected credit loss (CECL), were initially proposed so that lenders’ loss reserves would be forward-looking. Some recent studies have suggested that CECL could be procyclical, meaning that loss reserves would peak at the peak of a crisis. Although it is better than seeing failure only after it has happened, being required to raise liquidity at the peak of a crisis could still fail to save the lender from collapse, and it may even facilitate the lat- ter. However, previous procyclicality studies have explained all losses using macro- economic factors, ignoring the changes in credit risk and other portfolio drivers that preceded the recession. The current work looks at a wide range of models to test the degree to which CECL is procyclical for different types of model. The tests were also run using real historical macroeconomic scenarios, flat scenarios or mean-reverting scenarios. All tests were conducted on publicly available data from Fannie Mae and Freddie Mac using publicly disclosed models. Our study found that CECL lifetime loss estimates were only marginally sensitive to the quality of the economic scenario but changed dramatically with different modeling techniques. Some methods pre- dicted increased loss reserve requirements as early as 2006, while others only saw the recession as it happened or even afterward. Therefore, procyclicality under CECL will be strongly influenced by the choices of the lender.

Keywords: current expected credit loss (CECL) approach; procyclicality; vintage models; roll rate models; state transition models; survival models.

1 INTRODUCTION

Following the 2009 US recession and the global financial crisis (GFC), the Financial Accounting Standards Board (FASB) and the International Accounting Standards Board (IASB) sought new loan loss reserve rules that would allow lenders to reserve for losses that could be reasonably anticipated given what is knowable in their port- folio and the economy. These rules were intended to be an improvement on the pre- vious ones, which were inherently backward-looking: they only considered losses after a default was believed to have occurred. In addition, the rules were designed to ensure loss reserves looked further into the future.

In the United States, the new rules adopted in June 2016 – the current expected credit loss (CECL) – set all loss reserves to cover the full lifetime of the loan (Financial Accounting Standards Board 2012). The international standard, Interna- tional Financial Reporting Standard 9 (IFRS 9), uses a twelve-month loss reserve for stage 1 accounts (performing as expected) but also adopts a lifetime loss calcula- tion for stage 2 accounts (increased risk) (International Accounting Standards Board 2014). The primary motivation for the divergence in the standards was the need to accommodate the thousands of smaller lenders in the United States. Therefore, CECL is essentially IFRS 9 stage 2 for everyone.

Given the goal of anticipating and reserving for future loss events, the 2009 US recession and the GFC serve as good test cases of whether loss reserves under these rules will lead, lag or coincide with a crisis. Although provisions that lag serve little purpose, as was true under the old rules, provision requirements that peak coincident with a crisis can also cause significant financial stress because of the difficulty in raising capital during a crisis. When loss provision requirements are coincident with the underlying losses, this is referred to as procylicality.

To prepare for future macroeconomic stresses, we should determine how we can reasonably expect CECL estimates to behave. By applying CECL to the 2009 US recession in the case of mortgages and considering what was knowable at the time, we seek to determine the degree to which CECL is procyclical. Most importantly, prior experience has shown that different modeling approaches can provide differ- ent degrees of foresight, so we tested time series, vintage, roll rate, state transition and survival models, all of which are statistical techniques. In addition, we consid- ered some simple, spreadsheet-based methods that were proposed recently for use by smaller organizations to comply with CECL. These are loss timing, vintage copy forward and weighted average remaining maturity (WARM).

To capture what was knowable at each forecast point, we purchased historic pub- lications of macroeconomic scenarios from Consensus Economics corresponding to the month before each forecast date. Those scenarios were for the following two years. Using this as the foreseeable period, we applied a mean-reverting algorithm to the scenarios in order to forecast the remainder of the term of the thirty-year fixed- rate conforming mortgages from Fannie Mae and Freddie Mac. We also considered a flat macroeconomic extrapolation of two years followed by mean reversion. The third option was immediate mean reversion. Any of these approaches would have been realizable through the economic cycle.

The results of this study show the extreme importance of model choice for CECL. Also, once we abandon perfect foresight, the differences between realistic economic scenarios are minor.

This study was inspired by the results of previous studies. An early study on pro- cyclicality argued on largely theoretical grounds that CECL would reduce procycli- cality (Cohen and Edwards 2017). However, these conclusions made several assump- tions about what lenders would be able to estimate. A more recent analysis showed that CECL estimates for mortgages would have been countercyclical if the future macroeconomic conditions were known (Chae et al 2018). Neither of these studies addressed the question of the real-world timing of CECL provisions given what was knowable at the time.

Most recently, a study by Covas and Nelson (2018) created a vector autoregression model to generate macroeconomic scenarios at each forecast point. These were fed into a time series model similar to what is included in the current work. Their result was that CECL estimates are highly procyclical and would have caused a worsening of the crisis through reduced liquidity and subsequent lending by banks. Their study did not consider any other model types.

In fact, previously published work on the mortgage crisis has shown that losses did not peak purely due to deterioration in the economic environment. Several authors have explored the impact of the moral hazard of securitization in driving risky lend- ing (Elul 2015; Keys et al 2010; Nadauld and Sherlund 2013). Levitin et al (2009) argued that securitization was the sole cause of the crisis, although this is rebutted by Foote et al (2012).

Breeden (2011) used a data set back to 1990 to identify credit cycle peaks in 1991, 1995 and 2001 to argue that although securitization played a deleterious roll in the 2009 crisis, this loss peak also corresponded to a credit risk cycle. A subsequent analysis by Breeden and Canals-Cerda used a richer, more recent data set to show that half of the mortgage crisis was explainable by typical scoring factors. Although some part of the residual could have been attributable to unseen underwriting changes, it again correlates strongly to drivers of consumer risk appetite as measurable from the Federal Reserve Board’s Senior Loan Officer Opinion Survey (SLOOS).

Regardless of the exact mix of effects identified by these authors, the message from all of them is that mortgage losses were the result of more than an economic crisis. One could assume that economic crises are unpredictable, but credit cycles are a different matter. This leaves the possibility that losses are not as unpredictable as some have assumed, if the correct model is used.

2 DATA

Our test was performed on publicly available data from Fannie Mae and Freddie Mac: specifically, thirty-year fixed-rate conforming mortgages were analyzed. Loan- level information from 2001 through 2017 from both sources was normalized and combined. In addition to monthly loan status, the database contains a number of attributes suitable for loan-level credit risk estimation. The full list of data fields for Fannie Mae and Freddie Mac appears in Table 1.

For the models developed in the study, the following definitions were used.

• Default: current loan delinquency status > 3, ie, 90C days past due (DPD).
• Active: nondefault and current actual UPB > 0.
• Attrition: zero balance code D 1 (prepaid).
• Outstanding balance: current actual UPB if status D active.
• Default balance: current actual UPB if status D default.
• Origination balance: current actual UPB if current date D vintage.
• Loss: default balance C accrued interest C total costs total proceeds.
• Accruedinterest:defaultbalance((currentinterestrate/1000.0035)/12) (months between last principal and interest paid date and zero balance date).
• Total costs: foreclosure costs C property preservation and repair costs C asset recovery costs C miscellaneous holding expenses and credits C associated taxes for holding property.
• Total proceeds: net sales proceeds C credit enhancement proceeds C repur- chase make whole proceeds C other foreclosure proceeds.

The data analyzed in this study represents more than US$2 trillion of conform- ing mortgages. All models were segmented by risk grade: subprime is less than 660 FICO, prime is 660 to 780, and superprime is 780 and above. Although data was available by zip code, only nationwide models were created, in alignment with the available macroeconomic scenarios.

2.1 Macroeconomic scenarios

The macroeconomic scenarios in this study were obtained by purchasing reports from Consensus Economics published in the month preceding each quarter’s fore- cast, so these CECL estimates use the real economic assumptions available at the time. Consensus Economics provides quarterly scenarios for the following two-year period. From the factors provided, the following were considered for mortgage mod- eling: real gross domestic product, real disposable personal income, unemployment rate, three-month Treasury bill rate and ten-year Treasury bond yield. These sce- narios represent the average predictions of twenty-four prominent economists or economic forecasting institutions.

Figure 1 shows that the consensus economic scenarios follow a rational pattern. In any quarter, they follow recent trends and then begin to revert to loan-run averages.

The models tested here were originally developed as part of a more extensive CECL mortgage study (Breeden 2018). In that study, the selection of macroeconomic factors was chosen from mortgage-related factors available in the government’s Dodd–Frank Act Stress Test (DFAST) scenarios, so as to maintain reproducibility of the results. Among the factors considered in the models, the house price index (HPI) and Dow Jones US Total Stock Market Index were missing from the Con- sensus Economics list. The Dow Jones index rarely appears in the models, but HPI is quite important. To run the models, a simple vector autoregression model was created to predict the values of the missing variables from the available data of all macroeconomic factors. This model was trained on data preceding each fore- cast date. Accuracy was not a primary concern, only that the models could be run out-of-sample.

3 MODELS

CECL is built on models. To study procyclicality, the models developed in the CECL mortgage study by Breeden (2018) were employed, along with some recent addi- tions: WARM and loss timing. The following sections provide summary descriptions of each model. Complete descriptions are available in the original study details.

All models were segmented by risk band (subprime for FICO 6 660, prime for FICO between 660 and 780, superprime for FICO > 780) and by US states and territories (fifty-two in total). Therefore, we actually built 156 independent models within each of the model types listed below. The online appendixes to this paper attempt to summarize those models since the full details cannot be provided here.

3.1 Time series

The simplest forward-looking model in this study requires us to create macro- economic time series models of the balance default and pay-down rates. Lifetime losses can then be simulated by projecting forward under a mean-reverting base macroeconomic scenario until all currently outstanding balances are either paid or charged off. Transformation of the macroeconomic data and model estimation are primary considerations (Enders 2004; Hastie et al 2009; Jolliffe 2002; Judge et al 1985; Wei 1990).

The time series model used macroeconomic factors to predict the balance loss rate and balance payment rate, segmented into subprime, prime and superprime. These models use lagged transforms of the economic factors described in Section 2.1:

Pure macroeconomic time series models necessarily assume that all portfolio dynamics are explainable by the economy, with no regard for changes in underwrit- ing or other policies. As mentioned in the introduction, many other factors may have contributed to the mortgage crisis. In this time series model, those factors will either be absorbed indirectly via possibly spurious correlations to macroeconomic factors or be missed entirely.

3.2 Roll rate

For the last forty years, the two most common kinds of model for retail lending portfolios have been credit scores and roll rates. Roll rate models are similar in spirit to state transition models, but they are estimated on aggregate monthly balance flows from one delinquency bucket to the next (Federal Deposit Insurance Corp 2007):

Historically, roll rate models have used moving averages of past rolls. For CECL estimation, Ri .t / is modeled with macroeconomic factors.

In addition, the balance pay-down rate for nondelinquent accounts is modeled with macroeconomic data so that both charge-off and pay-off end states are included. Thus, the roll rate model is like the time series model, but with intermediate delin- quency transitions added. The final lifetime loss is calculated by summing the monthly losses until all existing loans reach zero balance.

3.3 Age-period-cohort (vintage)

Vintage models naturally capture the timing of losses and attrition versus the age of the loan; therefore, they are an obvious choice for lifetime loss calculations. An age– period–cohort (APC) approach is commonly used to estimate such models (Breeden 2007; Breeden and Thomas 2008; Glenn 2005; Holford 2005). Using rates for prob- ability of default (PD), exposure at default (EAD), loss given default (LGD) and probability of attrition (PA), monthly loss forecasts are created and aggregated to a lifetime loss estimate.

Here, “age” is the age of the vintage; “vintage” is the origination date; “life cycle” is also known as the hazard function or the loss-timing function; “credit risk” mea- sures the relative risk of each vintage; and “environment” captures macroeconomic impacts and account management changes through time.

The life cycle, environment and vintage functions can be represented either with splines or nonparametrically. The coefficients for those functions are typically esti- mated with logistic regression or with a Bayesian estimator (Schmid and Held 2007). To obtain a solution, a constraint must be set on the linear trends, as described in the APC literature (Holford 1983).

Macroeconomic scenarios are used to project the future value of the environment function, which is then combined with the vintage and life cycle functions to produce monthly forecasts for each vintage. The lifetime loss forecast sums across both vin- tages and calendar date to the end of the loan term or until the outstanding balance reaches zero.

3.4 State transition

State transition models are the loan-level equivalent of roll rate models. Rather than modeling aggregate movements between delinquency states, the probability of tran- sition is computed for each account. The states considered are current, delinquent up to a maximum of six months, charge-off and pay-off.

In the roll rate model, account transition probabilities rather than dollar transitions are modeled. These derive from Markov models, although in practice they may not satisfy the Markov criteria that no memory of a previous state be used in the model. They were used first and most heavily for corporate ratings and commercial lending, where most of the literature is still to be found (Israel et al 2000; Tru ̈ck 2014; Wei 2003). However, they are often used for retail lending, most often for mortgages (Bangia et al 2002; Thomas et al 2001). The method used here is most like that of Berteloot et al (2013).

For modeling, multinomial logistic regression is a preferable approach to estimat- ing a set of logistic regression models for the transition from a given state to all other possible states. However, multinomial logistic regression required so much memory (estimating many transitions simultaneously) that the data set had to be reduced to the point where some of the transitions became unstable for modeling. The above approach is therefore a compromise for available resources. The regression models consider external macroeconomic drivers as well as internal factors for the accounts, such as FICO, LTV, etc. Functions of age were also considered in order to capture life cycle effects.

To make forecasts, if the input variables to the transition probability models sat- isfy the Markov condition of having no memory or prior state, then the forecasts may be created via a series of matrix multiplications, as in the Markov chain approach. However, if the input factors do have memory, such as number of times delinquent in the previous six months (a common predictive factor), then a Monte Carlo approach must be applied to a sample of the accounts to simulate possible portfolio perfor- mance. At each time step, each account is assigned a specific state based on the probability of that transition and a drawn random number.

In all cases, the probabilities are functions of time because the macroeconomic scenarios will change with time using the same mean-reverting scenarios described earlier. The accounts will be simulated until they reach a terminal state, such as charge-off or pay-off, or they reach the end of the loan term.

EAD and LGD are modeled separately as functions of the age of the loan in order to capture balance pay-down with time.

3.5 Multihorizon discrete time survival model

Conceptually, discrete time survival models are a loan-level enhancement of vin- tage models, usually with the implication of creating loan-level models with scoring attributes. They did not evolve from survival models, but they have obvious similari- ties. Because lending performance data is generally recorded in monthly increments, the discrete time approach is appropriate, in which case a discrete time survival model is identical to logistic regression with a hazard function as a fixed input.

For the present study, the life cycle and macroeconomic correlations from the APC vintage model estimation are used as fixed inputs for a logistic regression panel data model with scoring attributes. This two-step process is undertaken to avoid multi- colinearity problems when trying to estimate everything simultaneously. PD, PA and EAD are estimated using this process.

Separate origination and behavioral models are built, with the former using only factors available at origination and the latter using both origination factors and behav- ioral factors such as recent delinquency. The multihorizon aspect comes from the fact that a separate regression is estimated for each forecast horizon. This is done because the coefficients for delinquency are highly nonlinear with forecast horizon. However, by horizon 12, the coefficients stabilize and can be used for all future val- ues. Because any delinquent account will have either cured or charged off within six to twelve months, the remainder of the forecast is dominated by persistent factors like FICO score and LTV.F

The final lifetime loss forecasts are created by aggregating the loan-level monthly loss estimates.

3.6 Weighted average remaining maturity (WARM)

Popularized in FASB webinars on simple CECL approaches (Federal Deposit Insur- ance Corp 2018), WARM is just the multiplication of the recent average loss rate and the average expected life of the loan. Average expected life can be defined as the age at which half the loans have been paid off or charged off, or it can be defined as the age at which the average outstanding balance will be half of the initial balance. If one were to multiply the loss rate by the outstanding balance in each month of the loan, adjusted for prepayment risk, the result would be equivalent to the age at which the outstanding balance is half of the initial balance. That is the approach used here.

Most notably with WARM, the loss rate used in the calculation is not dependent upon the age of the account, economic conditions, current delinquency or any credit risk factors. This model assumes a steady-state portfolio, where any such changes are introduced manually via quantitative adjustments (Q-factors). No Q-factors were used in these forecasts, because we cannot guess what Q-factors portfolio man- agers would have made historically. However, Q-factors existed as part of FAS 5. If these manual adjustments had been effective through the last crisis, one assumes the creation of CECL would not have been necessary.

3.7 Loss timing

Lastly, the loss-timing approach is conceptually the same as a balance-based hazard function (Cox and Oakes 1984), but it is computed simply in a spreadsheet as the average monthly loss rate versus the age of the vintage. The vintages are aligned by age, and an average balance loss rate is computed relative to the origination balance. To create forecasts, the loss-timing function is applied for all forecast horizons for active vintages.

Because this is a simple spreadsheet average rather than a statistical estimate of a hazard function (Aalen 1978; Kaplan and Meier 1958), it will be subject to a number of biases. Again, this is being proposed as a CECL solution but would not normally be considered a suitable model for forecasting.

4 RESULTS

The goal of the study is to simulate how these models might perform out-of-sample. To achieve this, one would ideally estimate all of the coefficients solely on data prior to the start of each quarter’s forecast. The challenge is that this data does not have enough history prior to the recession to fully estimate these models. Further, when we apply these models to the next recession, we will have this past recession to model against. Although no future recession is expected to be a replay of the previous recession, having one to train on is better than using no historic data at all.

With this in mind, it seemed overly harsh to run the models with no history, and yet we did not want to offer perfect foresight. As a compromise, we used the full history to estimate economic sensitivity and product life cycles, but not to estimate scoring coefficients. It has been the authors’ personal experience that life cycles for a specific product, such as a thirty-year mortgage, are stable through time and across lenders.

Macroeconomic sensitivities, when the correlations are restricted to variables close to the consumers’ finances, are also reasonably stable. Unemployment and changes in house prices are always the dominant effects on mortgages. The biggest change between the 2001 recession and the 2009 recession was the lengthening of unemployment benefits to 99 weeks. This was significantly greater than previous recessions and appears to have caused the optimal lag between unemployment and mortgage default to have increased during the recession.

For the time series and roll rate models, estimating macroeconomic factors across the full data set means they are fully in-sample. For all other models, the macro- economic sensitivities were taken from a model over the full data; other coefficients were reestimated each quarter using only the preceding data. As will be seen later in the scenario comparison results, using partially out-of-sample models probably makes no difference in the comparison.

4.1 Comparing models

The models described above were run each quarter using the corresponding macro- economic scenario illustrated in Figure 1 for the first twenty-four months. Then, a mean-reverting algorithm was applied to the optimally transformed macroeconomic factors using a second-order Ornstein–Uhlenbeck algorithm (Uhlenbeck and Orn- stein 1930; Breeden 2018, Chapter 10). Each point along the time series in Figure 2 represents the CECL lifetime loss estimate at that point for the Fannie Mae/Freddie Mac mortgage portfolio. A CECL estimate was run for each quarter from 2005 Q1 through 2015 Q4. The black line in both parts of Figure 2 is the actual future life- time loss for loans outstanding at that forecast point. The actual performance data runs through 2017, so the tail losses beyond 2017 were filled in with a vintage model, because that was one of the most accurate in our previous study. The actual forward-looking lifetime losses rise steadily between 2005 and mid-2008, because new higher-risk loans were being originated at a rapid pace. The “actual” line is an unobtainable ideal, but at the far opposite extreme is the dotted line for WARM. The graph shows that WARM is really no different from the moving average loss rate typ- ically used previously for ALLL calculations, just with a lifetime multiplier. As such, it peaks well after the crisis is over at a time that all other models correctly predict decreasing loss reserves. Regulatory expectation is that users would apply Q-factors to manually adjust the WARM baseline to expectations concerning the economic and credit cycles. However, because WARM is so out-of-phase with actual reserve needs, lenders would be better off using a completely flat (through-the-cycle) average loss rate than trying to back out the post-peak behavior of WARM. In short, WARM should not be used for CECL.

Between these extremes of perfect foresight and pure hindsight, things get much more interesting. The most important question is not “when will reserves peak?”. In any recession scenario, reserves will peak when the macroeconomic peak is known. However, not all losses are driven by macroeconomic factors. When large volumes of new loans are booked, the hazard function or APC life cycle will predict when those losses should occur in the future. Similarly, there exists in mortgages a strong credit cycle, which can also be incorporated into the forecast. Therefore, the question

is whether any of the models give a warning or simply jump to peak levels at the last moment. Figure 3 shows the change in quarterly CECL estimates as a percentage of the outstanding loan balance.

The time series model that other studies suggest would be procyclical is, in fact, procyclical. Loss reserves do not begin to rise until late 2008. In the period between 2005 and late 2008, when large volumes of risky loans were being booked, the time series model shows no increase in reserves. If this were the only basis upon which CECL was judged, it would be considered a failure. However, that is a failure of the model, not the guidance.

The best two models from the perspective of anticipating the crisis are the APC (vintage) model and the multihorizon discrete time survival (DTS) model. Both the APC and survival models start increasing reserve estimates in 2005 and accelerate provisioning through 2009. Notably, these models are adding reserves when the con- sensus economic scenario suggests that nothing is wrong economically. The models are responding to shifts in credit quality and applying the loss timing to the new originations. Although they are only halfway to a perfect foresight result, they would have provided an early warning to lenders as early as 2006 that risks were increasing significantly. This should be the stated goal of CECL: not to predict the economic cycle, which is unlikely, but to accurately forecast the risk already in the portfolio. The roll rate model is better than nothing, but it is only half as good as the vintage and survival models. This roll rate model used time series models of the net roll rates to incorporate the economic cycle. A more simplistic model that only uses moving averages of the rolls, as is common practice, would give much less warning.

Also shown is the state transition model. As seen in Figure 2, the state transition model showed a subdued response to the recession. Unlike the vintage and survival models, which estimate coefficients relative to the forecast horizon, the state tran- sition predicts only one step ahead (ie, one month). Therefore, its coefficients are optimized for short-term accuracy, not accuracy throughout the economic cycle. This issue of forecast accuracy with horizon is discussed further in Section 6.4.3 of the online appendix. Note that state transition models offer a very flexible approach that could be hybridized with the vintage model to include aspects of the credit cycle and life cycle, thus providing an earlier warning and long-term accuracy. That was not done here in order to highlight the differences between model elements.

Table 2 summarizes the change in CECL reserves as estimated by each model. The values are shown as a fraction of the ideal reserve level that would have been maintained with perfect foresight. The results are ranked by which models gave the most advanced warning during the critical period of 2005 through 2007.

The results in Figure 3 and Table 2 are not the actual quarterly provisions required, because provisions must also include the replacement of charge-off expenses in the

Figure 4 shows the quarterly provisions that would be required under each model as a percentage of outstanding balances in that quarter. Comparing provisions makes the models look more similar, because they all add in the replacement of charge-off balances, even when the CECL calculation is unchanged.

4.2 Comparing scenarios

In the preceding analysis, we used real macroeconomic scenarios from the forecast periods, but how important is it to obtain the best possible scenarios? Figure 5 shows the sensitivity of the APC CECL estimates to macroeconomic scenarios. The solid line represents the forecast with perfect foresight regarding the full life of the loans. The short-dashed line uses perfect foresight for the first two years, but then applies a mean-reverting algorithm for the remainder of the forecast. Therefore, this line is the best possible CECL estimate, because it complies with the CECL rules about reverting to long-run averages beyond a reasonable and supportable period. The two- year reasonable and supportable period used here appears to be the most common value chosen in the industry.

The other lines use either consensus, flat or mean reverting for the first two years, and then continue with mean reverting for the remaining life of the loan. Other than the “actual” and “actual, 2Y” scenarios, all other scenarios are plausible CECL approaches.

Until January 2007, the “actual, 2Y” and all realistic scenarios are equivalent. All of the realistic scenarios diverge from perfect foresight between January 2007 and January 2009. However, all of the realistic scenarios are roughly equivalent dur- ing this period and provide the early warning shown in Figure 3, because that early warning is not dependent on the economic cycle. The consensus economic scenario

is more accurate and one quarter earlier than the flat scenario, but the difference is minor. This result suggests that creating a good forecasting model is more important than finding an optimal macroeconomic scenario.

5 CONCLUSION

During the initial adoption phase for CECL, the biggest step for lenders will be to establish appropriate risk management practices for loss reserves, to gather necessary data and to create the necessary systems. During this time, any model will probably be acceptable so long as overall progress is shown. However, the purpose of CECL is to help lenders survive crises. Simple models appear to offer little help in surviving crises and may actually be harmful. To be useful in the period leading up to a crisis, the model needs to be effective.

The current results demonstrate that the model is more important than the eco- nomic scenario. We accept that perfect macroeconomic foresight would be useful, but it is unobtainable. After that, the difference between the best and the worst real- istic macroeconomic scenarios is slight, but the difference between the best and worst models is significant when considering the rate of change of loss reserves.

For the thirty-year mortgages tested here, the average lifetime for the loans is seven to eight years. For such long-lived assets, most of the CECL estimate will be determined by the period beyond the reasonable and supportable (R&S) period when some form of through-the-cycle average is employed. One might assume that this would cause CECL estimates for such long-lived assets to be less procyclical than FAS 5. However, we find that the combination of a strong credit cycle and a two-year R&S period is enough to create significant volatility throughout the economic cycle.

DECLARATION OF INTEREST

The authors report no conflicts of interest. The authors alone are responsible for the content and writing of the paper.