Sageworks Advisory Services PRACTICAL CECL TRANSITION EXPEDIENTS VERSUS CASH FLOWS

Transcription

1 Sageworks Advisory Services PRACTICAL CECL TRANSITION EXPEDIENTS VERSUS CASH FLOWS

2 Use of this content constitutes acceptance of the license terms incorporated at This content license confers certain rights to you, the user, of this material, and reserves others to Sageworks, Inc. Further, this content license disclaims liabilities for use of the content. It should be carefully reviewed before transmitting or using this content. The following summary is not a substitute for, nor supercedes, the content license agreement referenced above: You may share this content in the exact form it was provided to you, so long as it is shared unaltered and in its entirety, including these terms, and our brands and marks -- any user of the content will be bound by the same Agreement by using the content. While we encourage sharing of this content, you are prohibited from profiting from the use of the content or any derivative works. We make no warranty of any kind for this content -- it is provided in good faith as a courtesy to the industry we are privileged to serve. As this content is provided to an industry, it may have absolutely no relevance or applicability to your particular institution or use case, and your critical thinking is required to successfully use (or not use) the content. We undertake no obligation to keep the content up to date, answer inquiries regarding the content, or support it in any other way. 2

3 EXECUTIVE SUMMARY Not only does the CECL standard grant institutions broad latitude in the data and information used in measurement, the standard is non-prescriptive in methodologies to be used (though does go so far as to enumerate several sensible options). Of the methodologies described in the standard, only the use of discounted cash flow (DCF) techniques gets a special disclaimer, both in guidance and in the ancillary supervisory and Transition Resource Group (TRG) materials: the use of DCF techniques are not required. It is also the only method with specific guidance regarding its application. However, given the fundamental exercise under the standard is the estimation of collectability of future cash flows given periodic, time-specific assumptions such as those for reasonable and supportable forecasts and reversion periods, we consider the use of DCF techniques to be the most conceptually sound and defensible approach for the challenge at hand; moreover, the DCF approach is broadly applicable to a wide variety of financial assets. We consider the other measurement approaches commonly evaluated to be expedients to the cashflow method of estimation, each with unique conceptual challenges APPLICABLE GUIDANCE 1 FASB Accounting Standards Update Topic 326 (CECL): The allowance for credit losses may be determined using various methods. For example, an entity may use discounted cash flow methods, loss-rate methods, roll-rate methods, probability-of-default methods, or methods that utilize an aging schedule. An entity is not required to utilize a discounted cash flow method to estimate expected credit losses. Similarly, an entity is not required to reconcile the estimation technique it uses with a discounted cash flow method p109,

4 INTRODUCTION The standard concludes its introduction of potential methodologies by specifically disclaiming the requirement to use one of its enumerated methodologies. Regulatory bulletins and similar publications from the accounting community have repeated this disclaimer; and for a very good reason. These materials often include sometimes on the same page a statement that reads that new models are not required, or that vendor models will not be required, or that software purchase or build will not be required. As we discuss elsewhere in this guide, this is a conditionally true statement. For example, an entity conducting a cohort migration-to-loss analysis over the loss emergence period could achieve compliance by changing the loss emergence period input to a weighted-average-life input; no new models, no new software, etc. required. Of course, this assumes the entity has available data to support the larger weighted-average-life, that the migration methodology is appropriate for all financial assets over those longer terms, etc. No stakeholder in this accounting transition could credibly make the statement that software is not required while also making the statement that discounted cash flow approaches are preferred. Bottom-up cashflow estimation, tuning, rollforward, etc. without purpose-built software would not be practical for most institutions. However, we assume in our Practical CECL transition methodology that such software is available, and can make these techniques accessible to institutions in broad ranges of complexity. The fundamental nature of the new standard aligns conceptually with a cashflow model, which is a projective, time-sensitive approach driven by individual modeling inputs. Conceptual concerns with other modeling approaches are resolved by the use of a DCF measurement. Further, DCF measurement does not require a long historical run of loan-level details to drive its component inputs. Finally, the componentization of those inputs (and outputs) provide clear avenues to layer peer or industry parameters when an institution s own first-party parameters are either unreliable (due to numerical concerns) or unavailable (due to data issues). For these reasons, we consider DCF analysis to be the most practical approach to the standard s requirements, and we consider other methodologies to be expedients to this form of analysis. Readers of this chapter should be able to: Articulate the conceptual advantages of cashflow modeling Understand the limitations inherent in other expedients Conceptualize how methodologies and models might be blended into a DCF approach Justify the use of a DCF model Principle: Cashflow estimation is the most directly applicable modeling approach for requirements under ASU 326; other approaches are expedients to this method. 4

5 BLURRED LINES We commonly encounter a conceptual hurdle when we initially work with our clients, auditors, regulators, and other audiences concerned with the accounting transition. Due in no small part to the bewildering array of seminars, webinars, whitepapers, conferences, and other content produced by a cottage industry of consultants, vendors, and other experts the industry has been trained to think of methodology as a binary proposition. In other words, a PD and LGD analysis is seen as distinct from a cashflow analysis, which is seen as distinct from a vintage analysis. While each of these forms of analysis can exist distinctly, they can also be used as inputs to one another. For example, our cashflow model is formally a PD/LGD/EAD analysis, which incorporates credit inputs and calculations (Probability of Default and Loss Given Default) and timing inputs (contract payments and prepayments) to determine exposure at default. One could further use vintagebased inputs to these elements, taking into account the different prepayment (or default) tendency of seasoned versus new loans. APPLICABILITY OF CASHFLOW CONCEPTS A cashflow model is time-sensitive, and projective. We seek to produce, fundamentally, a simulation of loan behavior through a maximum of each asset s contract life. This projection requires, at minimum, the following inputs on a periodic basis: Expected Principal (Payment) Expected Interest Likelihood of Default (PD) Unrecovered Principal in the Event of Default (LGD) Time to recovery of defaulted principal Providing a cashflow model with a static PD and LGD assumption would produce a sensible current credit loss measurement. If we use our reasonable and supportable forecasts to describe the future time series of PD (and/or LGD) we produce a current expected credit loss measurement. That future, forecastable PD or LGD may be reflective of current conditions or different conditions. PD and LGD beyond our forecastable period should be reflective of longer-run historical conditions. The exercise, then, reduces to producing a future time-series of PD and LGD assumptions to use, the technical specifics of which are discussed in the chapter Loss Driver Analysis. This is not a trivial exercise, but it is a policy-formation exercise; automation of the application of this policy is straightforward. Cashflow estimation can produce the following outputs on a periodic basis: Anticipated Defaulted Principal Contract Principal Collected Additional Principal Collected Anticipated Interest Collected Anticipated Recovered Principal Anticipated Lost Principal 5

6 The time-bound nature of this modeling approach can then lend itself to defensibility and simple but powerful backtesting; the roll-forward of a DCF measurement can be compared with observed portfolio behavior to determine the model s appropriateness. COHORT MIGRATION EXPEDIENTS In a closed-cohort migration, a selection of loans is considered as of a specific date, and some behavior is measured for those assets and only those assets over a period of time. The specific behavior, and the specific period of time, determine the meaning of the analysis. It is a common practice for measurement under the current accounting standard to conduct a closed-cohort analysis for a loss emergence period to produce an estimation of credit losses over the incurred period. Such analysis could be performed at a major pool level, e.g. commercial real estate, or a sub-pool level based on some ordinal risk characteristic (risk rating, classification, etc.). When performed on an ordinal sub-pool level, this method more defensibly prices the risk premium for e.g. Substandard credits vis a vis Pass credits. The mechanics of cohort migration are covered in a dedicated chapter. Cohort migration requires loan-level details and transaction records on those loans, and a cohort migration analysis for purposes of a CECL measurement requires those details for a period of time at least equivalent to the weighted-average-life of the asset. Even then, only a single data point would be produced and we would recommend users of this analysis for either the present or future standard to include several data points in order to reduce measurement volatility. The consideration of cohort migration as a measurement expedient is the likely root of the loan-leveldetail for an economic cycle misconception. We can see the underlying theory more clearly if we consider a portfolio of 1-year assets, for example, annually renewed commercial lines. 6

7 In this case, an institution has loan-level details for an economic cycle, which they use to compute a series of 1-year loss rates. The institution makes a forecast for 2 years, which indicate conditions similar to the recent past and other performing environments; the institution then uses an average of 1-year loss rates from similar environments to produce their reserve estimation. If the institution were to forecast a deteriorated environment, the application would be similar: This concept choosing which data points are applicable, and which are not represents a broad, conceptual change from the current accounting standard, but we do not see clear to a method to fulfilling the guidance s requirement without doing so. Setting aside the data requirement, this approach is conceptually sound for a short-lived asset, but longer termed assets introduce modeling complexities that would provide practitioners reason to question the definition of expedient. Consider a 5-year asset for the same institution: 7

8 Our questions and problems become apparent. For data points that span a wide variety of economic environments, which ones do we choose? Do we consider the 2011 datapoint a recovery or a recession data point? Further, the reversion approach becomes awkward. Conceptually, one could weight the performing rate by 2/5 (given a 2-year performing forecast on a 5 year asset) and then the cycle average historical rate by 3/5. Yet that approach only solves for an immediate reversion strategy, so how would we change our model if instructed to revert on a straightline or other basis during an examination or audit exercise? We frame these concerns as questions not out of a Socratic instinct; we frame them as questions because we have not encountered practices that satisfy these concerns. Further, we haven t even addressed what we would do if our loss rates were negative or zero in the performing environment. The level of effort and data required to perform this analysis does not seem practical, especially if we find ourselves relying on qualitative adjustment for 70% of our allocation. For some institutions and for certain asset classes the cohort-migration expedient will be serviceable, but in our experience practical, conceptual, numerical, and data considerations produce inherent weaknesses that prove difficult to justify given the limited benefits. VINTAGE EXPEDIENT The use of vintage analysis in measurement is mentioned in the guidance and attendant bulletins from the supervisory and audit communities. It is a theoretically satisfying modeling approach, and can yield actionable insights for certain asset classes that might find deployment beyond the Allowance scope. A full description of the vintage approach to measurement is provided in a dedicated chapter. Inherent in any vintage approach is the concept of remaining loss. Essentially, we analyze loan-level details to compute a table that articulates observed loss experience, organized by year of origination and time since origination. Commonly, a vintage analysis will produce one-half of a table like so: 8

9 This table is a numerical representation of the concept of a Loss Curve for homogeneous assets: What we typically find is that newly originated loans tend to perform initially (for most asset classes) with losses occurring later, before seasoning effects and LTVs tend to reduce losses. Often we find a small bump past maturity, essentially reflecting that loans that have not been paid off past that point have specific problems that can lead to credit loss. We may also find a net recovery trough past a certain point. Computation and disclosure of this curve (or table) is a disclosure requirement under the standard, but it is important to note that computing the vintage losses does not on its own produce an estimate of losses under CECL. Further analysis is required to transform the observed vintage-based loss experience into a vintage-based estimation of remaining expected loss. Critically, institutions electing to employ this analysis as an expedient will find difficulty rigorously justifying or quantifying their adjustments to observed loss rates for each vintage year. Observing a downward trend in year-3 loss rates is easy, extrapolating that trend to remaining year-3 balances is a far more complicated matter. This sets aside the question of how an institution would justify adjustments to vintage year expected losses when considering reasonable and supportable forecasts. Given sufficient data, regression analysis might produce an appropriate and defensible model, but we have consistently found conceptual issues extending observed vintage loss experience to the future in a rigorous manner. Further complicating the conceptual, numerical, and data concerns already present in this and other expedients, vintage analysis may be prima facie inappropriate for important, material portfolio segments. Vintage analysis can be both powerful and predictive for portfolio segments constituting loans that are homogeneous, but the following homogeneity is required: Similar in term and structure (e.g. 3-year installments) Similar in purpose or collateral grade (e.g. autos) 9

10 Similar in exposure size A portfolio with a blend of terms, structures, or other characteristics will typically not produce meaningful vintage results. Vintage analysis is inappropriate for revolvers 2, lines, and other common commercial loan structures as well. Finally, we find from a practical standpoint that historical origination data is often suspect and suffers from lack of controls. For example, the specific process by which a loan is boarded and its commitment versus initial draw recorded can present problems for reliable measurement for many portfolio segments. LIFETIME PD AND LGD ANALYSIS In this analysis, a closed-cohort approach is used to measure life-of-loan probabilities of default and conditional losses given default. This modeling expedient presents similar concerns and conceptual hurdles to the closed-cohort migration to loss, but does present the advantage of decomposing the measurement into two components; an institution may produce sensible PD estimations based on its own experience but supplant the LGD assumption with an LGD function such as the Frye-Jacobs model or otherwise employ industry data. In other words, this expedient is more robust to numerical problems than the migration analysis described above, but still presents data and conceptual concerns, especially for longer-lived assets. ROLL RATES AND TRANSITION MATRICES These forms of analysis are conceptually more cleanly aligned to the guidance s requirement, in that they are fundamentally projective rather than look-back based. Essentially, an institution or entity measures the periodic tendency for a loan to move from one category (such as performing) to another (such as days past due), and extrapolates that tendency over the average life of the assets to produce a tendency for an asset to migrate to loss. Other simulation approaches, such as the use of Markov chains and Monte Carlo simulation, present similar advantages and concerns. We believe these are useful approaches insofar as they are projective, and not limited to specific or narrow methods for asset pooling and portfolio segmentation. A time series of observed transition matrices could be regressed against economic indicators to produce a sensible and justifiable forecasting model, as well. However, we consider these approaches at least as conceptually and computationally complex as the DCF approach we describe in our Practical Transition methodology, but without the added advantage of providing financial intelligence and symmetry to other valuation and accounting exercises. Further, the use of these techniques do not provide clean inputs where external and internal information can be leveraged according to an institution s judgment. 2 Vintage can be incredibly powerful for revolving structures where each draw is categorized as its own vintage, but this is beyond the scope (or appropriateness) of almost all depository financial institutions. 10