Fair Servicing: What Examiners Test in Servicing Data, and How to Test It First

Fair lending examination does not stop at the credit decision. A servicer that grants a forbearance to one borrower and denies it to a similarly situated borrower on a prohibited basis has a fair lending problem, even though no application was declined and nothing about it will ever appear on a HMDA loan application register. The CFPB instructs its examiners to test for precisely this. The data they need to do it does not live in your LAR.

This article covers what those servicing decisions are, why origination-side testing cannot see them, and how the statistical methodology used on application data transfers to servicing data. It is also specific about where that methodology stops transferring cleanly, because that is the part most treatments of this topic skip.

What fair servicing means, and why it is examined separately

The Equal Credit Opportunity Act and Regulation B govern credit transactions, and servicing a loan is part of the credit transaction. The Interagency Fair Lending Examination Procedures, issued in August 2009 by the OCC, FDIC, Federal Reserve Board, Office of Thrift Supervision and NCUA, states in its introduction that a lender may not "treat a borrower differently in servicing a loan or invoking default remedies."

That single sentence is, notably, close to the whole of what the interagency procedures say about servicing. Read the rest of the document and you find an examination methodology built entirely around origination: Part III walks through underwriting comparisons, pricing analysis, steering, redlining and marketing, and every one of those procedures assumes a population of applications with approve and deny outcomes. Appendix IV supplies sample size tables for application samples. Appendix V explains how to identify marginal applicants. There is no servicing module, no servicing sample table, and no servicing focal point definition anywhere in it.

The servicing-specific instruction lives somewhere else. The CFPB's Mortgage Servicing Examination Procedures (June 2016) devotes Module 8 to loss mitigation, early intervention and continuity of contact, and inside that module are two sections named Disparate Treatment in Loss Mitigation and Disparate Impact in Loss Mitigation. Module 9 applies the same analysis to foreclosure. These sections tell examiners to look at how loss mitigation outcomes are distributed across prohibited-basis groups, to compare modification terms, and to compare foreclosure timing and rates.

Module 8 also says how to do it. The examination of whether a loss mitigation program involves disparate treatment relies, in the CFPB's words, on the procedures in its ECOA examination program manual, the ECOA baseline review modules, and the Interagency Fair Lending Examination Procedures. So the methodology examiners bring to your servicing data is the origination methodology. That is not an inference about how examiners might approach it. It is written instruction.

The practical consequence is a gap that sits between two documents. The instruction to apply origination methodology to servicing is explicit and citable. The mechanics of doing so — what a servicing sample looks like, how large it should be, what counts as a marginal case, which variables legitimately explain a decision — are not written down in either document. Those have to be worked out. Most of this article is about working them out.

The servicing decisions examiners test

Regulation X gives the loss mitigation process a defined shape, and that shape is where the testable decision points are. 12 CFR 1024.41 sets out receipt of a loss mitigation application and what makes it complete at (b), evaluation at (c), denial of loan modification options and the notice of denial reasons at (d), borrower response at (e), the prohibition on foreclosure referral at (f), the prohibition on foreclosure sale at (g), and the appeal process at (h).

Each of those transitions is a decision, and each decision is a place a disparity can appear:

• Whether an application was treated as complete. A borrower whose application is repeatedly deemed incomplete never reaches evaluation. If completeness determinations skew across groups, the disparity is upstream of every outcome you would otherwise measure.
• Which loss mitigation option was offered. Reinstatement, repayment plan, forbearance, modification, short sale and deed-in-lieu are not equivalent outcomes. Two borrowers can both be "approved" and receive materially different relief.
• Modification terms. Rate, term extension, principal forbearance and capitalized arrears all vary. Approval rate parity with term disparity is still a finding.
• Time to decision. How long a complete application sat before evaluation, and how long the borrower waited for the offer.
• Foreclosure referral timing. How many days of delinquency elapsed before referral, and whether that differed across groups holding delinquency and hardship constant.
• Fee assessment. Property inspection fees, late fees and corporate advances are assessed under discretion or automated rules that were rarely designed with fair lending testing in mind. This is the most frequently overlooked item on the list.
• Appeal outcomes. Who appealed, and who succeeded.

One thing worth being straight about: we could not identify a public CFPB or DOJ enforcement action alleging that discrimination occurred in a servicing decision on a prohibited basis. The discrimination cases are origination cases, about pricing and underwriting. The servicing consent orders that exist allege process failures under RESPA, TILA and the Consumer Financial Protection Act rather than fair lending violations. So the risk described here is a documented examination priority, not a documented enforcement pattern, and anyone telling you otherwise should be asked for the case name.

A related caution, because this specific error circulates: the CFPB's Supervisory Highlights Issue 25 (December 2021) contains a mortgage servicing section describing failures to evaluate complete loss mitigation applications within the required window, and it separately contains a fair lending section describing statistically significant pricing exception disparities. Those are two unrelated findings, and the pricing one is an origination finding. They are frequently cited together as though the report documented a servicing fair lending violation. It does not.

Why servicing disparities hide from origination-side testing

An institution with a mature fair lending program can run a full origination analysis every quarter and still have no visibility into any of the decisions listed above. Three structural reasons account for it.

There is no register. Origination testing begins with a defined universe: applications acted on in a calendar year, in standard fields, with a required demographic collection. That universe arrives ready-made because Regulation C requires you to build it. Servicing has no equivalent. Before you can test anything, you have to define the population yourself — which borrowers were eligible for which decision, during which window, under which investor guidelines — and reasonable analysts will define it differently. Much of the work in a servicing self-assessment is population definition, and none of it exists in origination testing.

The demographics are often missing. HMDA requires collection of applicant race, ethnicity and sex at application. Servicing systems frequently do not carry those fields forward, and for purchased loans or transferred servicing the origination demographics may never have arrived at all. Proxy methodology therefore does more work in servicing than it does in mortgage origination testing. Bayesian Improved Surname Geocoding, which is a supplement on the origination side, often becomes the primary demographic source on the servicing side. That is a meaningful difference in the confidence you can attach to a result, and it should be stated in the analysis rather than buried.

The outcome is sometimes a duration. An underwriting decision is a point event with a categorical outcome. Days-to-referral and days-to-decision are continuous, and the interagency comparative method does not accommodate them. That method ranks denied applicants by qualification, identifies a benchmark, and looks for approved control-group applicants no better qualified than that benchmark. There is no way to express "waited 71 days instead of 34" in that framework. Timing disparities need a different statistical treatment, and if you only port the comparative file review across, you will not find them.

Running the same disparity analysis on servicing data

The parts of the origination methodology that do transfer, transfer well. We have run this class of analysis on lending data since 1987, and the sequence is the same one an origination self-assessment follows: establish the population, measure raw outcome differences across groups, control for legitimate factors, test what survives, then read files to understand what the numbers mean.

What changes is the control set

Origination regression controls for credit score, debt-to-income, loan-to-value, loan amount and collateral. None of those describe a loss mitigation decision. The legitimate factors in a servicing model are different:

• Delinquency stage at the time of the request, and payment history before the hardship
• Documented hardship reason and whether it was characterized as temporary or permanent
• Verified income at the time of the request, which is not origination income
• Investor or guarantor, because Fannie Mae, Freddie Mac, FHA, VA and portfolio loans run under different eligibility rules
• Occupancy status and current property valuation
• Prior modification history, since prior relief often constrains what is available
• Escrow shortage and advance balances

The investor variable deserves particular attention, because it undermines an assumption the origination methodology relies on. Origination comparative analysis measures consistency against a single set of underwriting guidelines. A servicing portfolio operates under several sets of investor guidelines at once, so two borrowers who look identical on every borrower-level characteristic may be genuinely, legitimately eligible for different outcomes. Testing that ignores investor will manufacture disparities that are not there. Testing that controls for investor without checking whether investor composition itself correlates with a prohibited basis will miss a real one.

What a marginal case means in servicing

Appendix V of the interagency procedures defines the marginal applicant, on the reasoning that discrimination is likelier among applicants who are neither clearly qualified nor clearly unqualified. That reasoning holds in servicing, but the definition has to be rebuilt. The servicing analogue is a borrower who was neither clearly eligible nor clearly ineligible for the option under the governing investor guidelines — sitting near a debt-to-income threshold, or near a delinquency boundary, or presenting a hardship that could reasonably have been documented either way. Those are the files where discretion operated, and discretion is where the risk is.

Where naive significance testing breaks

One borrower can submit several loss mitigation applications over the life of a default, and can be evaluated for several options within one application. Origination testing treats applications as independent observations. Servicing observations cluster within borrower, and treating them as independent inflates your sample and makes results look more significant than they are. Either analyze at the borrower level or account for the clustering. This is a genuine statistical trap and not a technicality.

The data you need, and where it lives

The fields for this analysis are spread across systems that were not built to be joined. Loan-level and delinquency data sit in the servicing system of record. Loss mitigation application dates, completeness determinations, evaluation outcomes and denial reasons often sit in a separate default management or workout platform. Fee assessments are in the transaction ledger. Foreclosure referral dates may be in the attorney or trustee system. Demographics, if they exist, are back in the origination file.

Assembling that into one analyzable table is most of the effort, and it is the reason servicing testing gets deferred. It is also the reason the exercise pays off once rather than repeatedly: the join logic is the hard part, and it only has to be solved once per portfolio. Comply DataMine exists to run analysis across data that is not a LAR, which is exactly the shape of this problem.

At minimum you need, per borrower and per decision: the decision type, the request date, the completeness date, the decision date, the outcome, the denial reason where applicable, the terms offered, the investor, delinquency stage at request, documented hardship, verified income, occupancy, current valuation, prior modification indicator, and every fee assessed with its date and code. Plus a demographic field or a documented proxy.

What a servicing self-assessment looks like

The sequence below mirrors the origination self-assessment, adapted for the differences described above.

• Define the population and write the definition down. Which decisions, which window, which loans. Because there is no register, the definition is a methodological choice you will have to defend, so record the reasoning rather than only the result.
• Establish the demographic basis. Use collected data where it exists. Where it does not, apply BISG and state plainly in the write-up that the analysis rests on proxied demographics.
• Measure raw outcome distributions by group, for each decision point: completeness, option granted, terms, timing, referral, fees, appeals.
• Control for the servicing factor set, investor included, and check whether investor composition itself correlates with a prohibited basis.
• Handle timing separately. Duration outcomes need their own treatment; do not fold them into the categorical comparison.
• Analyze at the borrower level or account for repeat observations.
• Identify marginal cases and read those files, comparing outcomes for borrowers who sat near the same eligibility boundary.
• Test fee assessment as its own analysis. It is separable, it is often automated, and it is rarely examined.
• Document the whole method, including what you could not test and why. An assessment that names its own limits is more credible to an examiner than one that implies completeness it does not have.

The statistical work here is the work we have been doing on lending data for 39 years, applied to a different set of decisions. The regression, the proxy methodology, the matched comparison and the significance testing are the same tools described on our Comply Fair Lending page, including the regression analysis that explains a disparity to an examiner. What differs in servicing is the population you build, the variables you control for, and the honesty required about the demographic basis you are standing on.

Servicing is where fair lending testing is thinnest across the industry, and it is an examination area the CFPB has written procedures for. Those two facts together are the argument for testing it before someone else does.

Schedule a demonstration to talk through how this analysis runs on your data.