CECL and lifetime loss estimation

Study purpose

This chapter is written as a serious banking study guide, not a technology brochure. The aim is to explain cecl and lifetime loss estimation in a way that a business analyst, architect, developer, tester, risk specialist, operations lead, compliance reviewer or student can use inside a real bank. The discussion stays close to banking decisions, data, controls, customer impact, model governance and audit evidence.

The chapter also explains how this topic connects to AI and machine learning without pretending that every banking problem should be solved by AI. Some questions need deterministic rules. Some need scorecards. Some need statistical models. Some need human judgement. The practical skill is knowing which method belongs where, and how the evidence travels from source data to final bank action.

How to study this chapter

CECL stands for current expected credit losses. It is a US GAAP credit loss methodology under FASB ASC Topic 326. For banking AI students, the important point is that CECL requires an expected loss view using relevant information about historical experience, current conditions and reasonable forecasts.

In practical banking terms, this means the bank must identify the decision point, the data available at that moment, the control owner, the allowed action, the expected customer or regulatory impact and the evidence that will be stored afterwards. A model output becomes useful only when it changes a real workflow in a controlled way. If the output cannot be tied to an action, a user, a policy rule, a monitoring metric and an audit trail, it is not yet production-grade banking AI.

For delivery teams, the safest approach is to describe the use case as a chain: source event, data validation, feature or rule calculation, model or score output, policy orchestration, human review where needed, customer or operational action, reporting and feedback. This keeps the project grounded. It also prevents the common mistake of discussing AI as if it floats above the bank instead of sitting inside payment hubs, credit platforms, risk engines, case tools, ledgers, data warehouses and monitoring dashboards.

CECL in plain language

CECL asks an institution to estimate expected credit losses over the life of financial assets measured at amortized cost and certain other exposures. Instead of waiting for loss events to become probable, the bank recognises an allowance based on expected losses.

Why lifetime estimation matters

Lifetime loss estimation forces the bank to think beyond near-term delinquency. A loan booked today may produce losses years later. The bank has to estimate expected losses across the contractual life, adjusted for expected prepayments, recoveries and reasonable forecast periods.

Difference from simple delinquency reporting

Delinquency reporting shows what is already late. CECL asks what loss is expected. That requires modelling, segmentation, assumptions and judgement. A clean past-due report is not enough. The bank needs a forward-looking allowance process.

Relevant information

FASB describes expected credit loss measurement using relevant information about past events, current conditions and reasonable and supportable forecasts. This gives banks flexibility, but also responsibility. The method must fit the portfolio and be supportable.

Portfolio segmentation

CECL models usually group assets with similar risk characteristics. Segmentation may use product, geography, borrower type, credit score, vintage, collateral, industry, risk grade or other factors. Poor segmentation can hide risk or exaggerate loss.

Historical loss experience

Historical data provides the foundation, but it must be relevant. A bank must ask whether history reflects today's underwriting, products, economy, customer mix and servicing practices. If not, adjustments may be needed. This is where governance matters.

Current conditions

Current conditions may include unemployment, interest rates, inflation, collateral values, sector stress, borrower behaviour, delinquency trends, utilisation, payment holidays or operational changes. These factors make the estimate more current than a purely historical average.

Forecast information

Forecasts can influence expected losses during a reasonable and supportable period. The bank needs a disciplined process for selecting scenarios, assumptions and reversion methods. Forecasting is not guesswork; it must be controlled and documented.

Methods banks may use

CECL can be estimated using different methods such as loss rate, vintage, discounted cash flow, probability of default/loss given default, roll-rate or other supportable approaches. The choice depends on data, portfolio and governance requirements.

AI and CECL

Machine learning can support risk segmentation, default prediction, prepayment analysis, loss severity estimation or anomaly detection. But CECL is a financial reporting process. Any model used must be explainable, validated, controlled and acceptable for audit.

Allowance governance

The allowance estimate usually goes through finance, credit risk, model risk, governance committees and audit review. Inputs, assumptions, overlays and movements must be explained. AI cannot replace that accountability.

Management judgement

CECL allows judgement, but judgement must be evidenced. If a bank adjusts a model output due to an emerging sector risk or data limitation, it should document the rationale, amount, owner, approval and monitoring plan.

Technology implementation

CECL requires data pipelines from core systems, loan systems, collateral systems, recoveries, finance ledgers, risk ratings and macroeconomic sources. Reconciliation is crucial because allowance numbers must tie back to reporting populations.

Business analyst role

The BA should map asset scope, segmentation, source data, model method, assumption workflow, adjustment process, approval steps, reporting output and audit evidence. CECL delivery is a finance-risk-technology control process, not just a model build.

The main lesson

CECL makes lifetime expected credit loss estimation a governed reporting discipline. AI can support insight, but the bank must support the allowance with data, assumptions, validation and audit evidence.

Chapter-level control checklist

Business analyst study prompts

• Draw the process before and after the model is introduced. Mark exactly where the bank decision changes.
• List the deterministic controls that must remain outside model discretion.
• Identify which source systems provide the data and whether the data is available before the decision.
• Write the reason a user would trust, challenge or override the output.
• Define the customer impact if the model is wrong in both directions.
• Describe what operations, risk, compliance, finance and technology each need to see.
• Explain what would happen if the model service is unavailable during a peak period.
• Define the monitoring report that proves the use case is still safe after go-live.

Main lesson

CECL moves allowance thinking toward lifetime expected losses and makes model evidence, segmentation and governance critical. The best banks will not adopt AI by replacing every existing rule, scorecard or control. They will adopt AI by understanding where learning systems improve judgement, where deterministic rules remain stronger, where human review protects customers, and where evidence must be retained for audit and regulatory challenge.

For Malla Banking Academy, the takeaway is simple: this topic belongs to banking first and technology second. AI and ML become valuable only when they are connected to capital, provisions, payments, fraud, sanctions, liquidity, reconciliation, reporting, customer treatment and operational resilience. That is the difference between a generic AI explanation and a bank-grade learning chapter.

The allowance question under U.S. GAAP

Current expected credit losses (CECL) is an accounting measurement under U.S. GAAP Topic 326 for financial assets within its scope, principally those measured at amortized cost and certain off-balance-sheet credit exposures. It asks what credit losses are expected over the contractual life of an asset using relevant past events, current conditions, and reasonable and supportable forecasts. The estimate is made at each reporting date. It is not a lending approval score, a regulator's capital formula, or an assertion that a particular borrower will default. The FASB credit-loss project and the FDIC's CECL resource center are starting points for primary material. Accounting staff must use the applicable Codification, amendments and entity policy.

The scope matters before mathematics. A portfolio may include amortized-cost loans, held-to-maturity debt securities, trade receivables, loan commitments and financial guarantees. Available-for-sale debt securities have a different impairment model under Topic 326. The bank must identify what it holds, its classification, the legal entity, and whether the exposure is in the relevant scope. A data scientist cannot infer that from a row labelled "credit." Similarly, an undrawn commitment may require a liability for expected credit losses subject to the applicable recognition conditions. The team should ask whether the bank can unconditionally cancel it and what exposure is expected to be funded. Do not simply add a commitment limit to loan balances.

CECL differs from the IFRS 9 general approach. The latter generally starts with 12-month expected credit losses for Stage 1 and moves to lifetime expected losses after a significant increase in credit risk, subject to the standard's scope and exceptions. CECL generally measures lifetime expected losses on in-scope amortized-cost assets from initial recognition. A U.S. bank should not copy an IFRS staging field into its CECL estimate, and an international group should not assume one reserve engine's output satisfies both standards. Shared data and scenarios can be useful, but accounting policy, population, term, measurement and disclosures require separate reconciliations.

Start with a dated reporting population

Suppose a fictional U.S. bank closes its books on 30 June. It has installment loans, revolving lines and a few individually troubled relationships. The analyst reconciles the eligible instrument inventory to the general ledger and to servicing balances. The snapshot contains legal entity, product, origination and maturity dates, amortized cost basis, contractual terms, payment schedule, accrued items as appropriate, charge-offs, modifications and commitment status. It also contains source and processing timestamps. A loan sold before the reporting date should not remain in the allowance population because a stale servicing file arrived late. A new loan booked at the cutoff should not be omitted because a batch job ran an hour earlier.

Population reconciliation is more than a row count. Compare balances by ledger account, product and entity, and explain differences in charge-off timing, accrued interest policy, foreign exchange or consolidation. Account for acquired assets and special treatments under the applicable rules rather than treating all originations as identical. Preserve each exclusion with a reason, approver and source. The preparer and reviewer should be able to trace a reported allowance to the precise records and estimate version that produced it. A replay should use the data available and accounting policy effective at the close, not today's corrected balance without a documented adjustment.

A control table can express the work:

The table does not prescribe one bank's accounting policy. It makes missing evidence visible. A model can produce a precise dollar amount from an incomplete cohort; the precision is meaningless until scope and reconciliation pass.

Contractual life, expected funding and cash flows

Lifetime does not mean "forever" or the full legal maximum without qualification. The contractual term and applicable guidance define the period, including treatment of prepayments, renewals and extensions. A credit card or other revolving product poses particular questions about how the estimated exposure and term are represented. The bank's accounting team documents its interpretation, while risk and data teams supply credible usage and payoff evidence. A three-year installment loan with expected early repayments should not be modelled as three years of full outstanding principal in every period.

Take a simplified 100-unit loan with scheduled principal balances of 100, 70 and 35 at the starts of successive years. If expected payment behaviour shortens the exposure, the cash-flow estimate must represent that behaviour consistently. If a forecast predicts a default after a borrower has prepaid, counting both the prepayment and the full later exposure would overstate loss. If interest or fees are included in the relevant measurement, their treatment must match the bank's accounting policy and basis. A simple PD × LGD × EAD calculation can teach the drivers of loss, but the final CECL method has to meet the measurement requirements and avoid omitted cash flows or double counting.

An off-balance-sheet commitment has a different source record. The analyst needs the undrawn amount, cancellation rights, expected draw behaviour and terms. Draw probability may rise when customers face stress, which makes a historical average from benign periods suspect. The funded exposure and commitment liability should be reconciled so the same expected loss is not counted twice. The bank must document the actual policy and legal facts; a teaching example cannot decide whether a specific commitment qualifies for a liability.

Pooling by similar risk characteristics

CECL generally calls for collective evaluation of assets that share similar risk characteristics, with individual evaluation when an asset no longer shares those characteristics. Segmentation can use product, collateral, vintage, risk grade, geography, term or other relevant attributes. More segments are not necessarily better. A pool with very few observations may yield unstable rates, while an overbroad pool hides a material underwriting change. The bank should explain why chosen characteristics relate to expected credit loss and how it tests whether those relationships still hold.

Imagine two small-business cohorts issued before and after a material policy change. They share a product code but differ in verification and collateral practices. Combining them without adjustment could make an old loss rate understate current risk. Alternatively, splitting them into many tiny cells could produce volatile estimates. The team can assess the observed outcomes, underwriting evidence and available external data, then choose a supportable segmentation and adjustment. Keep a record of rejected alternatives and sensitivity, especially when a manual override changes the estimate.

An individually evaluated relationship should not vanish from the collective reconciliation. The close pack should move it from the pool with an explicit transfer and track its separately estimated allowance. If a loan later returns to a pool, document when it again shares the relevant characteristics. Test both directions on cutoff dates. An analyst should also identify collateral-dependent assets where applicable guidance affects measurement rather than blindly applying a portfolio loss rate. The source of any valuation, its date and costs to sell where relevant need a clear lineage.

Choosing a method without pretending there is one formula

Topic 326 does not prescribe one modelling technique for every portfolio. A bank might use a loss-rate, vintage, probability-of-default and loss-given-default, discounted cash-flow or other reasonable method suited to available data and the nature of its exposures. The FASB's Topic 326 staff Q&A discusses developing an estimate; the interagency allowance policy statement discusses governance for FDIC-insured institutions. A sophisticated machine-learning model is not required. The method needs supportable assumptions, controls and results that can be explained.

For a loss-rate example, assume a homogeneous illustrative pool has 10 million units of amortized cost and an initial historical lifetime loss rate of 1.2%. That gives a starting amount of 120,000 units. Management then assesses whether underwriting, portfolio mix, current delinquency and reasonable and supportable forecasts differ from the historical experience. An adjustment of 0.3 percentage points would produce a 1.5% rate and 150,000 units, but only if the evidence supports that adjustment and it is not already embedded in the historical sample. The 30,000 difference is a documented assumption effect, not an AI-generated fact or universal CECL rule.

A PD/LGD-style estimate instead needs period-specific exposures, default probabilities, loss severity and timing. If a default can occur in several periods, the calculation must avoid counting the same exposure repeatedly as if it survived every earlier default. Recoveries and prepayments have to be handled consistently. Validate the result against actual charge-offs, recoveries and portfolio behaviour, while recognizing that later outcomes do not magically reveal the single "correct" estimate at an earlier reporting date. A discounted cash-flow method has its own rate and cash-flow rules. The bank should select and document a method that can be maintained, challenged and reconciled.

Forecast horizon and reversion

Historical loss information is a starting point, not a substitute for current evidence. The bank adjusts it for differences in risk characteristics and current conditions, then incorporates reasonable and supportable forecasts. It is not required to forecast macroeconomic variables credibly for every year of a long loan. For periods beyond its supportable forecast horizon, Topic 326 requires reversion to historical loss information that reflects the remaining contractual term, using a rational and systematic method. The FASB's 2025 amendment text reproduces the relevant Topic 326 guidance on forecasts and reversion. The bank should record the chosen horizon, reversion method and supporting evidence.

Suppose an illustrative portfolio has a two-year supportable forecast within a five-year remaining life. The forecast could capture an approved near-term employment and property-price view. Years three through five would follow a documented reversion to appropriate historical experience rather than an unsupported assertion that the adverse scenario persists unchanged. A straight-line reversion is one possible method, not a required one. The analyst should compare a faster and slower reversion as sensitivity, explain why the chosen path fits the data, and ensure the result is coherent across products and reporting periods.

Forecast governance needs a source and a date. Record the economic series, as-of date, scenario owner, approval, model version and how the forecast maps into losses. If different teams use the word "baseline," verify that they mean the same scenario. A scenario generated after the close cannot quietly replace the approved close view. Where management applies a qualitative adjustment for a risk not captured in the model, identify the underlying exposure, evidence, direction, amount, duration and the condition for removal. A permanent unexplained overlay can conceal model failure or double count a risk already captured in a new vintage.

A close bridge that finance can defend

The allowance changes for multiple reasons: new lending, repayments, charge-offs, recoveries, model updates, macro forecasts, risk migration and management adjustments. A quarter-end bridge should separate these drivers where practical and reconcile opening allowance, provision and other accounting entries to the closing balance. A rise in allowance does not necessarily mean borrowers suddenly became worse: growth alone can raise lifetime expected losses. Likewise, a decline after charge-offs may reflect balance removal rather than improved credit quality.

Imagine the fictional bank reports 1.8 million units of allowance at the last close and 2.2 million now. A bridge might identify 0.25 million from net new volume, 0.18 million from changed risk and forecast assumptions, a reduction of 0.10 million from repayments and 0.07 million from other identified movements. The illustrative components sum to the 0.40 million change. The actual bridge must reconcile exactly to the ledger and disclose how charge-offs, recoveries and provision activity are treated. Reviewers should see underlying cohort movements, not only a final number and a narrative that the macro outlook worsened.

A prudent review includes sensitivity to assumptions. What happens if a loss-rate adjustment is 0.2 instead of 0.3 percentage points? How much does a forecast-horizon change move the estimate? Is the effect concentrated in a small pool, one product or one source feed? A high sensitivity does not prove the estimate is wrong, but it directs validation and management attention. Compare realized outcomes with prior estimates using cohorts whose outcomes have matured; avoid declaring victory from a short, favorable observation window.

Model validation and controlled judgment

The revised U.S. interagency model-risk guidance in Federal Reserve SR 26-2 describes risk-based governance and validation, with applicability to be checked for the institution and model use. For a CECL estimate, challenge conceptual soundness, data lineage, implementation, outcomes and limitations. The accounting policy and management's judgment also need review; a model validator does not sign the financial statements. Smaller institutions may use simpler methods with proportionate controls, while complexity and materiality drive the depth of challenge.

Validation should examine the loss definition, cohort construction, seasoning, recoveries, prepayments, forecast mapping and reversion. An apparently excellent back-test may use later information in historical features or omit accounts that charged off. Compare predictions with outcomes by vintage and relevant risk segment. Investigate changes in underwriting or servicing practices before attributing differences to the macro scenario. A validation report should state what was independently replicated, what evidence was unavailable, material limitations and required actions. "The model has a high R-squared" does not answer whether the allowance is appropriate.

Management adjustments need a controlled journal and assumption trail. A reviewer should see why the model misses a risk, why the selected amount is supportable, whether the adjustment overlaps another input, who approved it and when it will be revisited. An adjustment that offsets an adverse model result every quarter deserves a separate challenge. An AI-generated narrative can help organize the close pack, but every number and factual claim in that narrative should be tied to verified data and reviewed by an accountable person.

The narrow 2025 receivables amendment

FASB Accounting Standards Update 2025-05 adds a practical expedient for current accounts receivable and current contract assets arising from Topic 606 transactions. For qualifying assets, an entity may elect to assume current conditions at the balance sheet date do not change during the remaining life. An entity other than a public business entity that elects the expedient may also elect the specified treatment of certain subsequent collections. The amendment is effective for annual periods beginning after 15 December 2025 and interim periods within them; early adoption is permitted. Read the official ASU 2025-05 for precise scope, conditions, transition and disclosures.

This is not a blanket exemption from forecasting for a bank's loan book. A team maintaining a commercial-loan CECL model should not turn off reasonable and supportable forecasts because a receivables expedient exists. The BA's acceptance test should use a qualifying current Topic 606 receivable and an ordinary loan as separate cases, check the entity's policy election and effective reporting period, and confirm that the exception is limited to its proper population. If the bank has no assets in that scope, the amendment may have no practical effect on its loan allowance. Current standards can change; version the policy by reporting date.

A two-pool worked review

Consider a second, separate classroom example at a 31 December close. The bank has 8 million units of performing secured installment loans and 2 million units of unsecured loans. Both totals reconcile to the ledger and exclude instruments sold before the cutoff. The bank has evidence that the pools have different loss behaviour, so it estimates them separately. Assume a historical lifetime loss rate of 0.6% for the secured pool and 2.0% for the unsecured pool, selected from cohorts with relevant terms and adjusted for differences in seasoning. Multiplication gives starting estimates of 48,000 and 40,000 units. The sum of 88,000 is a historical starting point, not the reported allowance.

The bank then evaluates current information. A new underwriting policy has changed the unsecured portfolio's risk mix; a property-price decline could affect secured recoveries. After documented analysis, assume the unsecured rate rises by 0.4 percentage points to 2.4%, while the secured rate rises by 0.1 percentage points to 0.7%. The resulting illustrative amounts are 48,000 and 56,000, for a total of 104,000 units. The 16,000 increase from the starting point has a traceable calculation: 8,000 for the secured pool and 8,000 for the unsecured pool. The bank must still consider reasonable and supportable forecasts, reversion, individually evaluated assets and any other applicable adjustments before it reports a final amount.

A reviewer should challenge each input, not only recompute multiplication. Is the 8 million balance a reporting-date amortized cost basis under policy? Did the historical loss windows include downturns and recoveries on comparable loans? Are the rates truly lifetime rates for the remaining term rather than one-year default rates? Has the property-price effect already been included in the historical experience or a separate forecast? Were charged-off accounts retained in the historical cohort long enough to observe their outcomes? Was a manual risk-grade override recorded before cutoff? These questions can change both the denominator and the adjustment.

If a 500,000-unit troubled loan no longer shares the secured pool's risk characteristics, the bank should remove it from that pool and evaluate it as required under its policy. Keeping it in both the 8 million pool and an individual estimate would double count it. Removing it without a separate estimate would omit it. The bridge should show the 500,000 transfer and the allowance effect. If the loan is collateral dependent under applicable guidance, the reviewer needs the relevant collateral evidence and valuation date. An automated classifier may suggest the case, but finance must confirm the accounting treatment.

Suppose a late file corrects 100,000 units of balances after the ledger reconciliation. The close owner assesses materiality and timing, then documents whether to rerun the estimate or book an approved adjustment. The original run, corrected file and final reported amount remain available. A dashboard that simply replaces the original 10 million population would prevent a later auditor from reconstructing what was approved. This example demonstrates why a reproducible population, differentiated pools, forecast evidence and a signed correction trail matter as much as the loss-rate arithmetic.

Reporting, disclosures and reviewer questions

The close pack should contain the population reconciliation, methodology by pool, forecast and reversion assumptions, qualitative adjustments, significant model changes, validation status, sensitivity and allowance movement. The financial reporting team then prepares the required disclosures under the current applicable rules. Disclosure is more than a generic paragraph about economic uncertainty. Readers should understand the relevant risks, how the estimate was developed and how the allowance moved. Legal and finance teams decide the exact filing requirements for the entity.

Before sign-off, ask: Can the bank reproduce the population and each material estimate from the reported date? Are funded loans and commitments accounted for without omission or double count? Does the historical window reflect today's underwriting and product mix? Is the forecast supportable over the chosen horizon, with documented reversion beyond it? Can a reviewer explain each overlay and its exit condition? Have individually evaluated assets been handled consistently? Has a model or data correction after close been assessed through the approved reporting process? Can accounting distinguish CECL from IFRS 9 and Basel capital numbers for the same portfolio?

A reliable CECL process is a chain from contract and ledger to historical evidence, current information, reasonable forecast, method, management judgment and signed report. Machine learning may improve forecasts or data quality, but it cannot settle accounting scope or approve a financial statement. The bank earns trust by making the estimate transparent, repeatable and open to challenge, including when the model is simple.

Source notes for further study

FASB credit losses project and ASC Topic 326 CECL materials; FDIC CECL resources.

Additional banking practice note 1

A real bank should never treat this chapter as only a model-building exercise. The model sits inside policy, architecture, workflow, risk appetite, customer communication, operational support and evidence retention. That full chain is what makes the solution bank-grade rather than experimental. In the context of cecl and lifetime loss estimation, this means the team should document the exact portfolio, channel, product, process and control boundary before making design decisions. A retail credit example, a corporate treasury example, a sanctions alert example and an instant payment example may all use data-driven scoring, but the risk owner, evidence requirement and customer impact are different.

A useful working question is: if this output is challenged later, who can explain why the bank trusted it? The answer should include the business owner, data owner, model owner, validation evidence, monitoring result, user action and stored audit trail. If that answer is weak, the bank may have analytics, but it does not yet have a controlled banking capability.

Additional banking practice note 2

The practical difficulty is usually not the algorithm. It is agreeing the definition, finding the trusted source, proving the data timing, making the output usable for staff, preventing misuse, monitoring outcomes and explaining the result months later to someone who was not part of the delivery team. In the context of cecl and lifetime loss estimation, this means the team should document the exact portfolio, channel, product, process and control boundary before making design decisions. A retail credit example, a corporate treasury example, a sanctions alert example and an instant payment example may all use data-driven scoring, but the risk owner, evidence requirement and customer impact are different.

Additional banking practice note 3

This is also why business analysts matter so much in banking AI work. They translate between risk language, product language, operations language, data language and technology language. Without that translation, a technically good model can still fail because the bank cannot use it safely. In the context of cecl and lifetime loss estimation, this means the team should document the exact portfolio, channel, product, process and control boundary before making design decisions. A retail credit example, a corporate treasury example, a sanctions alert example and an instant payment example may all use data-driven scoring, but the risk owner, evidence requirement and customer impact are different.

Additional banking practice note 4

The strongest implementation pattern is staged adoption. First understand the current process. Then run the model silently. Then compare with existing decisions. Then expose it as decision support. Then automate only low-risk paths when monitoring evidence proves that the use case is controlled. In the context of cecl and lifetime loss estimation, this means the team should document the exact portfolio, channel, product, process and control boundary before making design decisions. A retail credit example, a corporate treasury example, a sanctions alert example and an instant payment example may all use data-driven scoring, but the risk owner, evidence requirement and customer impact are different.

Additional banking practice note 5

Customer impact must remain visible. A false positive can delay a genuine payment, decline a good customer, create a complaint or overload operations. A false negative can allow fraud, credit loss, financial crime exposure or regulatory breach. Both sides of error need business cost and control ownership. In the context of cecl and lifetime loss estimation, this means the team should document the exact portfolio, channel, product, process and control boundary before making design decisions. A retail credit example, a corporate treasury example, a sanctions alert example and an instant payment example may all use data-driven scoring, but the risk owner, evidence requirement and customer impact are different.

Additional banking practice note 6

The chapter should therefore be read as part of the larger AI and ML banking journey. Data foundation, feature design, model governance, production monitoring, fallback, human oversight and audit evidence are not separate topics. They are the operating model that allows AI to be adopted safely. In the context of cecl and lifetime loss estimation, this means the team should document the exact portfolio, channel, product, process and control boundary before making design decisions. A retail credit example, a corporate treasury example, a sanctions alert example and an instant payment example may all use data-driven scoring, but the risk owner, evidence requirement and customer impact are different.