Case Study 16.2: Rafael Helps a Regional Bank Design Its First DFAST Submission

Context

Clearwater Bank is a fictional regional bank headquartered in Columbus, Ohio. It has grown steadily over the preceding five years through a combination of organic growth and two community bank acquisitions, crossing the $10 billion total asset threshold in one direction eighteen months before this case study begins, then growing further through another acquisition to reach $12.4 billion in total assets.

Crossing $10 billion in total assets carries significant regulatory consequences in the United States. Under the Dodd-Frank Act, banks above that threshold become subject to the annual Dodd-Frank Act Stress Test (DFAST) requirement — a materially different compliance environment from the examination-based regime that applies to sub-$10 billion banks.

Clearwater had known the threshold was approaching. Its management team had discussed it in strategic planning sessions. But knowing and preparing are different things. The operational reality of conducting a formal, model-driven stress test — projecting credit losses across nine quarters under three regulatory scenarios — was something Clearwater had never done.

The Head of Enterprise Risk Management, Kathleen Brennan, was given ownership of the DFAST implementation project. She had 11 months before the first submission deadline. She engaged Rafael Torres as the lead external consultant. Rafael had built stress testing programs at two large banks before his departure from Meridian Capital; he had since specialized in helping regional and mid-sized institutions navigate their first encounter with model-driven regulatory stress testing.

Their initial call lasted three hours.

Month 1: The Inventory

Rafael's first instinct, honed through experience, was that Clearwater's leadership did not yet know what they did not know. The first task was an honest inventory: what data did Clearwater have, and in what form?

The findings from a two-week data assessment were sobering.

Finding 1: No segmented historical loss data.

DFAST credit loss modeling requires satellite models — quantitative relationships between macro variables and default rates by loan category. Those models must be estimated from historical data. The minimum data requirement for a credible satellite model is typically 8–10 years of quarterly loan-level loss data by segment (commercial real estate, consumer credit, C&I loans, etc.).

Clearwater had been through three system migrations in the preceding decade. The most recent migration, two years prior, had produced a data warehouse with granular loan-level data going back two years. Before that migration, the data existed in legacy formats that had been partially converted but were not categorized to DFAST reporting standards. Before the acquisition-related systems, the data was on legacy acquired bank systems, partially migrated, partially paper-based.

Effective historical loss data by segment: approximately two years of reasonable quality data, falling far short of what was needed for model calibration.

Finding 2: No macro scenario modeling capability.

The Fed publishes DFAST scenarios specifying values for 28 macro-financial variables (GDP growth, unemployment, house prices, CRE prices, corporate spreads, interest rate curves, and others) over a nine-quarter projection horizon. Translating those macro variables into credit loss projections requires — at minimum — a spreadsheet model that maps macro variable paths to loan category default rates and LGD estimates.

Clearwater had nothing of this kind. Credit loss projections were done at the relationship manager level (annual loan reviews, classified asset tracking) or through simple allowance coverage ratio analysis — neither approach was structured for macro scenario-conditional projection.

Finding 3: No DFAST governance infrastructure.

DFAST requires that the stress test be governed by a documented process — a DFAST Policy, a Model Risk Management framework for the satellite models, a board review and approval requirement, and a documented process for scenario to results to disclosure.

Clearwater had credit policies, a credit review process, and an audit function. It did not have a DFAST Policy, a formal model validation function, or board-level experience with stress test output review. The Audit Committee had not been briefed on DFAST requirements.

Finding 4: No regulatory reporting templates pre-population experience.

DFAST submission requires populating the Fed's Supervisory Stress Test Information Collection (SSTf) templates in specific formats. The reporting infrastructure to extract, transform, and load data from Clearwater's systems into those formats did not exist.

Rafael presented these findings to Kathleen Brennan on Day 15. His opening line: "You have eleven months. We need to treat this like a product launch, not a compliance filing. Let's build the minimum viable stress testing program that is credible enough to survive the Fed's review, and then build from there."

The Build Plan: Six Months, Four Workstreams

Rafael and Kathleen designed a four-workstream build plan with clear monthly milestones.

Workstream 1: Data Remediation and Historical Loss Development

Months 1–3

The first priority was historical loss data. Without it, satellite model calibration was impossible.

Rafael brought in a data analyst experienced in banking system migrations. Together, they undertook a systematic extraction and reconstruction of Clearwater's historical loss data from legacy systems, focusing on the seven loan categories most material to Clearwater's balance sheet: 1. Owner-occupied commercial real estate (CRE) 2. Non-owner occupied CRE 3. Construction and land development 4. Commercial and industrial (C&I) 5. Residential first lien mortgages 6. Home equity loans and lines 7. Consumer installment

For the pre-migration period, they recovered five additional years of data — imperfect, inconsistently categorized in places, but sufficient to extend the usable dataset from two years to seven years of quarterly observations. The data reconstruction required multiple reconciliation passes: matching legacy system definitions of "charge-off" and "non-performing" to FR Y-14M reporting definitions, resolving categorization inconsistencies from the acquired banks (which had used different segment definitions), and establishing data lineage documentation for each reconstructed period.

The output was a historical loss rate database by quarter by segment — covering Q1 2016 through Q2 2023. Seven years, not ten, but sufficient to calibrate models through one recessionary period (the COVID-19 shock of 2020, which provided the loss experience the models needed to estimate recession-period behavior).

Kathleen reflected later: "We thought the data was our biggest problem. It turned out the data was just the beginning of understanding what we didn't know about ourselves."

Workstream 2: Satellite Model Development

Months 2–5

With historical loss data in hand, Rafael's team developed satellite models for each loan category. Given Clearwater's limited model development resources (a team of two quantitative analysts who were also responsible for ALCO modeling), the approach prioritized simplicity and interpretability over sophistication.

For each loan category, the approach was:

Select candidate macro variables: Based on regulatory guidance and economic logic, identify the macro variables most likely to drive loss rates in each segment. For CRE, the candidates were GDP growth, commercial real estate price indices, and the unemployment rate. For consumer installment, they were unemployment and personal income growth. For C&I, they were GDP and corporate credit spreads.
Run OLS regressions: For each segment, run quarterly OLS regressions of the historical net charge-off rate on lagged macro variables (macro variables are lagged 1–2 quarters to capture the delay between macro deterioration and loan losses materializing). Assess statistical significance, intuitive sign of coefficients (higher unemployment should predict higher loss rates, not lower), and in-sample fit.
Apply Fed macro scenarios: Once model coefficients are estimated, apply the Fed's published baseline, adverse, and severely adverse scenario macro variable paths to generate predicted loss rates for each of the nine projection quarters.
Validate against the COVID shock: The models were developed before COVID data was included in estimation, then used to "predict" 2020 charge-off rates as an out-of-sample validation exercise. Models that badly misfired on the COVID period required re-specification.

The most challenging segment was construction and land development. Clearwater had significant construction loan exposure from the Ohio residential development market. Historical construction charge-off data was thin (construction loans had performed well in recent years) and highly concentrated in a small number of large exposures. Statistical estimation with thin data is unreliable; a single large charge-off event could dominate the historical series. Rafael's solution was a combination: a simple statistical model for systematic risk (calibrated to industry data from the Fed's own stress test guidance documents) plus a qualitative overlay for idiosyncratic concentration risk in Clearwater's specific exposures.

The models were documented in a series of model development memos — six to eight pages per segment — covering data sources, variable selection rationale, regression specification, validation results, limitations, and intended use. This documentation was essential: the Fed's model risk examiners would review it.

A technical compromise that mattered: Rafael knew that Clearwater's models were simpler than what a large bank would produce. The satellite models were linear OLS regressions, not machine learning models with interaction terms and non-linear transformations. He made an explicit decision not to apologize for this in the documentation. Instead, the documentation was transparent about the simplicity, explained why simpler models were appropriate given the data limitations, and described the overlays and judgment applied to compensate for model limitations. The Fed's guidance explicitly acknowledges that smaller institutions are expected to use simpler models — complexity is not the goal.

Workstream 3: Governance Infrastructure

Months 1–4 (parallel to Workstreams 1 and 2)

Kathleen owned this workstream. The governance build had four components:

DFAST Policy: A formal policy document establishing the annual DFAST process, the roles and responsibilities of each function (model owners, validation, finance, compliance, board), the escalation framework for material model limitations, and the board approval requirement. The policy was reviewed by outside counsel and approved by the board before the models were completed.

Model Validation Function: Clearwater did not have an independent model validation team. Creating one from scratch in eleven months was not feasible. Rafael negotiated an arrangement with a specialized model validation boutique: Clearwater would retain them to conduct "challenge reviews" of the satellite model documentation — serving, effectively, as an outsourced validation function for the DFAST models. This is explicitly permitted under the OCC's Model Risk Management guidance (SR 11-7/OCC 2011-12) and is common among smaller institutions. The validation boutique produced formal challenge memos identifying limitations and recommending enhancements.

Board Education: The Audit Committee received two educational briefings on DFAST — what it required, what the first submission would look like, and what the board's specific responsibilities were (approving the DFAST policy, reviewing the stress test results before submission, and understanding the limitations of the models used). These sessions were designed by Rafael to be conversational rather than technical: "What would have to happen in the Ohio commercial real estate market for Clearwater's CET1 to fall below 7%?" rather than "Here are the regression coefficients for the CRE satellite model."

Documentation and Audit Trail: All workstream outputs were filed in a structured DFAST documentation repository — a SharePoint folder structure with version control — so that any step from raw data to final submission results could be reconstructed. This audit trail is not specifically required in DFAST guidance for banks at Clearwater's size, but Rafael insisted on it: "The first time the examiners look at your DFAST program, they will ask to follow a number through the process. If you can't show them how the number was produced, they will question everything."

Workstream 4: Results Production and Submission Infrastructure

Months 4–7

The final workstream was the plumbing: the operational process of producing the actual submission.

The DFAST submission for a $12 billion institution requires population of Fed reporting templates with projected quarterly balance sheet, income statement, and capital ratios under each scenario. These templates require specific data mapping from internal system categories to DFAST reporting categories.

Clearwater's finance team, led by the Controller, built a DFAST consolidation model in Excel — a deliberate decision not to invest in specialized vendor software for the first submission. The consolidation model pulled quarterly balance and loss projections from the satellite models, combined them with NII projections (built by the ALCO team), operating expense projections (built by finance), and a tax model (built by the tax department), and produced the template outputs.

This approach had risks: Excel models are prone to error, difficult to audit, and do not have built-in version control. Rafael addressed this through a mandatory peer review of every formula in the consolidation model (two junior analysts working from a checklist, cell by cell), a reconciliation to general ledger balances for the starting position, and a "challenger model" exercise in which Rafael's team independently estimated the severely adverse scenario results using a different methodology and compared to the primary model's output. The two estimates agreed within 5% on total credit losses — a standard Rafael considered acceptable for a first-year submission.

Month 8: The Dress Rehearsal

Eight months into the project, the team ran a complete end-to-end dress rehearsal: applied the Fed's current year scenarios through all models, produced the consolidation model output, populated the submission templates, and held a mock board review session.

Three significant issues surfaced:

Issue 1 — Template population error: The mapping of Clearwater's internal loan categories to the DFAST reporting categories had an inconsistency in the treatment of owner-occupied CRE. Loans that Clearwater classified as "Business Banking" loans secured by owner-occupied real estate were being split across two DFAST categories (CRE and C&I) in a way that was internally inconsistent. The error would have affected approximately $340 million in reported exposures. It was caught, documented, and corrected.

Issue 2 — LGD assumption inconsistency: The satellite models produced PD-based loss projections. LGD assumptions were applied separately by the finance team. The LGD for construction loans under the severely adverse scenario had been applied at a flat 30% — an assumption that was not stress-conditioned (construction loans in a severe real estate downturn historically produce LGDs of 50–60%). The flat LGD understated severely adverse losses by approximately $18 million. Corrected.

Issue 3 — Board review too brief: The mock board review was a 20-minute agenda item at the end of an Audit Committee session. Two board members engaged; two did not appear to have read the results summary in advance. Kathleen extended the board review to a dedicated 75-minute session two weeks before submission, with results circulated to all members five business days in advance with a written Q&A document covering the most likely questions.

The Submission

Clearwater's first DFAST submission was filed on the regulatory deadline. The results showed: - Starting CET1 ratio: 11.2% - Severely adverse stressed CET1 (minimum over the nine-quarter horizon): 7.8% - Capital depletion of 3.4 percentage points under the severely adverse scenario - Primary driver of capital depletion: CRE losses (46% of total) and C&I losses (31%)

Rafael presented the results to Kathleen with a note of context: "7.8% under severely adverse is a credible result for a bank with your portfolio composition. It's not outstanding — a large well-diversified bank might show a 1–2% depletion. But it's honest, it's model-supported, and it reflects genuinely conservative scenario-conditional loss estimates. That's what the first submission should be."

The Fed's examiners did not contact Clearwater with substantive questions on the first submission — which, for a first-time filer, is a good outcome. The absence of a challenge does not mean the submission was perfect; it means it was within the range the Fed considers acceptable for an institution of Clearwater's size and complexity submitting for the first time.

Year Two: Building on the Foundation

Twelve months later, Rafael was engaged again for year two. The improvements were material:

The historical loss data had been extended by one year of organic data, improving model calibration
A formal model performance monitoring process had been established, tracking model predictions against actual quarterly loss experience
The construction loan module had been partially recalibrated using peer data from FR Y-14 public disclosures
The board review session had been restructured around scenario narratives rather than model outputs — the board now engaged with the question "what economic conditions does this scenario describe?" before seeing the loss projections

What did not change — deliberately: the basic satellite model structure remained the same. Rafael's view on this was firm: "Year two is not the year to redesign the model. Year two is the year to make the year-one model better — more data, better validation, better documentation. Redesigning is a year-three conversation, and only if you have the data to support a more sophisticated approach."

What Rafael Learned — and Taught

In a conversation with Priya Nair at a regulatory technology conference two years after the Clearwater engagement, Rafael was asked what the most important lesson from helping a bank build its first stress test program was.

"The hardest thing," he said, "is not the modeling. Any competent quantitative analyst can build an OLS satellite model in two weeks if they have the data. The hardest thing is changing how the institution thinks about its own risk. When Clearwater's board sat down for the first time and saw a table showing 'severely adverse: CET1 falls to 7.8%', the initial reaction was relief — it's still above 4.5%, so we're fine. That's the wrong reaction. The right question is: what would have to happen for 7.8% to become 5%? What's between us and non-viability? The models are a starting point for that conversation. If the board treats them as the end of the conversation, the stress test has failed its purpose regardless of whether the submission is filed correctly."

Priya wrote this down. She had heard a version of the same sentiment from Maya Osei six months earlier, in a different context. It seemed to be the principle that separated institutions that understood stress testing from institutions that merely complied with it.

Discussion Questions

Rafael made an explicit decision to use simple OLS regression satellite models rather than more sophisticated approaches (machine learning, ARIMA time series). What are the arguments in favor of this choice for a first-year DFAST filer? Under what circumstances, in future years, would it be appropriate to increase model sophistication?
Clearwater had only seven years of historical loss data for most segments, compared to the 10+ years typically recommended. How did the team address this limitation? What alternative approaches exist for institutions with limited historical loss data?
The outsourced model validation arrangement (retaining an external boutique to perform challenge reviews) was used in place of an independent internal validation team. What are the advantages and limitations of this approach? What conditions would make it inappropriate?
The dress rehearsal in Month 8 identified three material issues — template population error, LGD assumption inconsistency, and insufficient board engagement. How does the dress rehearsal fit into a well-designed stress testing governance framework? Should it be a formal regulatory requirement?
Rafael's closing observation distinguishes between viewing stress test results as a compliance output ("we're above 4.5%, so we're fine") and viewing them as a risk management tool ("what's between us and non-viability?"). How should a bank's risk governance framework be structured to institutionalize the second mindset — making it the default rather than the exception?
Clearwater's first stressed CET1 under severely adverse was 7.8% — a depletion of 3.4 percentage points from 11.2%. Is this result reassuring or concerning? What additional information would you need to form a view on whether Clearwater's capital position is genuinely adequate?