A score is a measurement produced by a model
A credit score compresses information into a number, but the number has meaning only in connection with the model, underlying data and population. A consumer’s score in a personal-finance app can differ from a mortgage score without either calculation containing a simple mistake. Changing a model can move the number even when the person’s financial behavior is unchanged.
Mortgage lending also combines several decisions: whether a loan meets an investor’s eligibility rules, the borrower’s overall underwriting assessment, how the loan is priced and how investors evaluate the pool. A third-party score contributes to those decisions; it is not the whole decision. This article focuses on the Fannie Mae and Freddie Mac modernization initiative, with Fannie Mae’s current operational rules used as the detailed example. It is not a statement of universal mortgage law.
The dated status: approval did not mean simultaneous implementation
FHFA approved FICO 10T and VantageScore 4.0 in 2022, but approval was not a same-day replacement of Classic FICO across mortgage production. As checked October 4, 2026, FHFA’s policy page says lenders may choose Classic FICO or VantageScore 4.0 for loans sold to the Enterprises; FICO 10T remains approved for planned future use. Earlier transition schedules should not be read as today’s implementation instructions. [1]
Fannie Mae’s September 9 lender letter, updated September 30, expands VantageScore 4.0 to all Fannie Mae-approved lenders without prior written approval for eligible DU loans. DU Version 12.1 supports the reports. Classic FICO remains available, and manually underwritten loans must continue using it. FICO 10T is not yet eligible for Fannie Mae delivery. The letter supersedes the narrower limited-rollout language still visible in the April Selling Guide topic. [2][3]
The September 30 revision also aligns loan-level pricing adjustments across Classic FICO and VantageScore 4.0. That change applies to whole loans purchased in Purchase Ready status on or after October 1, 2026 and MBS with issue dates on or after that date. Operational acceptance, pricing effectiveness and a consumer’s closing date are therefore not interchangeable milestones. [2]
A score model and a credit-report merge are different choices
Fannie Mae’s current initiative page retains the three-bureau merged-report approach. When VantageScore 4.0 is chosen, the lender requests that model from each bureau and uses the same model for all borrowers on the loan. The newer score does not itself create permission to substitute a two-bureau report. VantageScore loans also require identification in delivery data, including Special Feature Code 067. [4]
The distinction is mechanical. A scoring model converts credit information into a measure. A merged report assembles information from repositories whose records may differ. A rule for combining borrower scores produces another derived number. Changing any one of those layers can alter the result, and changing several together makes attribution difficult.
This is why a stored field called credit score is incomplete for longitudinal analysis. The same database column can silently mix different models, reporting sources and aggregation methods. An apparent improvement in average portfolio score may then reflect a measurement change rather than a safer pool of borrowers.
What trended information adds, and what it cannot create
The Enterprises’ September 2026 playbook describes the newer models as using trended credit data and additional payment history, such as rent when available. It also notes that their automated underwriting systems already consider some such information separately. A model change can therefore overlap with information already used elsewhere in the underwriting process. The playbook’s expected benefits are agency and Enterprise characterizations, not measured results for every lender. [5]
Trended data is conceptually different from a snapshot. Two people can each show a $5,000 revolving balance today; one may have reduced it steadily from $12,000 while the other has increased it from $500. A trajectory may contain information beyond the ending balance. It still does not establish the cause of the change or the borrower’s complete budget.
Likewise, a model capable of considering rent cannot evaluate rental payments that never reach its usable data. Expanding who can receive a score is different from proving that each newly scored person qualifies for a mortgage. Income, obligations, property, product rules and the lender’s full assessment continue to matter.
Why equal numerical cutoffs need not mean equal risk
Consider a deliberately hypothetical comparison. On the same 10,000 applications, Model A places 6,000 people above a threshold and Model B places 6,500 above the same numerical threshold. Suppose 5,500 appear in both groups. Model A alone admits 500; Model B alone admits 1,000. The larger approved group does not prove an improvement in prediction or a deterioration in standards.
If the common group later has 55 observed adverse outcomes, the A-only group has 10 and the B-only group has 15, the combined observed rates would be 65 ÷ 6,000 = 1.08% for A and 70 ÷ 6,500 = 1.08% for B, rounded. B produces more loans and more total adverse outcomes while retaining a similar rate. Different outcomes in the two unique groups would change that conclusion.
These are invented figures, not FICO or VantageScore results. They show why approval counts, event counts and event rates tell different stories. In a real lending dataset, outcomes are usually not observed on the same terms for rejected applicants, so this complete comparison may be unavailable. A shared pricing grid is an operational policy choice, not proof that two models assign identical risks to every equal score.
Historical data improves analysis but does not recreate every original decision
The Enterprises released FICO 10T historical scores and additional VantageScore 4.0 history on July 1, 2026. Fannie Mae’s public documentation covers acquired loans from approximately April 2013 through September 2025. Scores are reconstructed from bureau archives, whose dates can differ from the original report date. Historical archives also do not retroactively incorporate every subsequent reporting change. [6]
The analytical implication is that observed score differences can combine model differences with timing and data differences. A late payment appearing between the archived snapshot and the actual report pull can move the result. A historical population of purchased loans also represents prior selection: it cannot, by itself, establish outcomes for everyone who would become eligible under a different future policy.
Outcome definitions matter as much as the score. Thirty-day , serious delinquency, default and realized loss are different targets. Performance can look different across origination years, loan-to-value bands and economic conditions. A ranking improvement does not automatically prove that predicted probabilities are calibrated, or that an unchanged numerical threshold has the same financial effect.
The transition has operating and competitive economics
Score competition can affect purchasing choices and vendor leverage, but a quoted score price is not the full cost of a mortgage credit report or completed loan. Bureau data, reorders, integrations, exceptions and downstream investor requirements can all contribute. No universal borrower savings figure is established by the primary policy documents reviewed here.
Technology changes also propagate beyond the origination screen. Pricing engines, mortgage insurers, delivery files and investor analytics can interpret the same number differently if the model identifier is lost. A valid underwriting result can still encounter delivery friction when one downstream system has not implemented the new data definition.
The evidence that would clarify the transition includes actual delivered-loan mix, comparable full-report costs, exception rates and matured outcomes for the newly included population. The primary sources support a meaningful expansion in available scoring options. They do not yet establish a universal improvement in access, pricing or losses, and they do not make future FICO 10T implementation a completed event.
Sources
- FHFA, Credit Scores policy and implementation history, checked October 4, 2026Official sourceBack to text: ↑
- Fannie Mae Lender Letter LL-2026-06, September 9, 2026, updated September 30, 2026SourceBack to text: ↑1↑2
- Fannie Mae Selling Guide B3-5.1-01, General Requirements for Credit Scores, April 22, 2026; read with subsequent lender letterSourceBack to text: ↑
- Fannie Mae, Credit Score Models and Reports Initiative, current implementation pageSourceBack to text: ↑
- Fannie Mae and Freddie Mac, Credit Report and Credit Score Model Playbook, September 2026SourceBack to text: ↑
- Fannie Mae, Historical Credit Score Files, data scope and limitationsSourceBack to text: ↑