A mortgage book can be diversified by ZIP code and concentrated in weak houses.
A funded book is cut by state, metro and vintage, and every cut reads the collateral through the growth rate of its area. If the homes behind the loans are each the weaker example of their type in their own market, the book is concentrated in a way no geographic table shows. Good Investment scores every house in the book within its own market and returns a band, a coverage status and a review route per row, so a risk team sees where the weak collateral sits, by count and by balance, before it surfaces at the next mark.
Good Investment supports analytical review and risk research. It is not an appraisal, credit decisioning system, or replacement for an institution's underwriting policy.
Where it fits
Find the loans whose collateral is positioned to run behind its own market.
Measure concentration on something other than location: price tier, vintage, originator.
Report exposure by count and by balance, because the two disagree.
Direct surveillance to the files where the property read and the mark disagree.
Support a portfolio review with a property read a committee can check.
Risk signals
Weak-band collateral clustered in one market or one origination vintage.
Neighborhood-level weakness inside a metro that reads as stable.
A segment small by loan count and large by balance.
Unsupported rows clustered in one originator, which says where the evidence is thinnest.
Review workflow
1Share the book schema and the review cadence you run today. No customer data is needed for this step.
2Agree the population, the fields available, and how coverage exceptions are handled.
3Receive every row back keyed on your identifiers: a score, a band, a coverage status and a review route.
4Read composition by band, by count and by balance, and open the files where the weak band sits.
What does a diversified-looking mortgage book still hide?
A residential book can span many states and metros and still hold highly correlated positions, because homes compete for buyers at a scale far smaller than a metro, often smaller than a ZIP. Two homes a few streets apart at similar prices can be positioned very differently against what buyers in that submarket want, and when the only geographic field is a state or a metro none of that is visible.
The tape describes the loans in depth and the houses in three fields, so every aggregate inherits that asymmetry. Weighted-average LTV and a state mix describe the obligation and the map. They say nothing about how any single house is positioned among the homes it competes with, and that is the field that decides recovery.
pool_review_output · within-market rank distribution
Illustrative data
Share of the metro's supported rows, by local quintile
Metro A
8.4% of pool by balance · 41 supported rows
0%20%40%
19%
21%
20%
21%
19%
Q1Q2Q3Q4Q5
Rows spread across the local distribution. This is roughly the mix you would expect from buying broadly inside the metro.
Metro B
8.1% of pool by balance · 38 supported rows
0%20%40%
34%
27%
18%
13%
8%
Q1Q2Q3Q4Q5
Nearly two-thirds of the rows sit in the bottom two local quintiles. The headline share matches Metro A, but the position inside the market is materially different.
Q1 is the weakest local quintile, Q5 the strongest. Both panels use the same vertical scale. Rank is a position within a market, not a forecast for a region.
Two metros carrying an almost identical share of the same pool, cut by where each row ranks inside its own local market. A state or metro exposure table reports these as the same 8% bet. A share table cannot express the distributions, because the variation happens underneath the unit it measures on. Figures are invented to show the shape of the view. Institutional real estate analytics
How is a funded book read on one basis?
Every house is scored within its own local market, so a row in Ohio and a row in Arizona sit on one axis. The book comes back cut into bands on that score, and concentration is measured on something other than location: which price tiers, vintages and originators hold the collateral positioned to run behind.
Report it on both bases. A segment that is a small share of the book by loan count can be a large share by balance, and a view on a single basis understates exposure roughly half the time. The gap between the two is a finding in its own right.
Coverage is reported as an output. Rows the model cannot support with enough local evidence come back marked unsupported, with the reason named, and routed to a person. A book whose unsupported rows cluster in one originator or one vintage is telling you where its evidence is thinnest before a single scored row is examined.
pool_review_output · composition and coverage
Illustrative data
Submitted
500
Supported ranks
418 · 83.6%
Out of coverage
82 · 16.4%
By loan count
17.2%
50.2%
16.2%
16.4%
By balance
12.4%
46.8%
24.9%
15.9%
Strong tail 17.2% / 12.4%
Middle 50.2% / 46.8%
Weak tail 16.2% / 24.9%
Out of coverage 16.4% / 15.9%
Composition reported on both bases, because they disagree. In this illustration the weak tail is 16.2% of the rows and 24.9% of the balance, a gap a count-only report hides completely. Coverage gets its own segment instead of being folded into the middle, so a pool whose unsupported rows cluster somewhere stays visible. Figures are invented to show the format. See portfolio monitoring
Where does surveillance look first?
At the weak band. On a funded book there is no natural trigger to re-examine a quiet loan, and performance data arrives after the point at which intervention is cheap. The band and the coverage view give the team a list: the files whose collateral is positioned to slip, the files where the property read and the mark disagree, and the files the model could not support.
From the book the team opens the property report on the exposed collateral, with the neighborhood context and the confidence flag, and carries one vocabulary for property risk into the portfolio review. Your risk system, your marks and your credit judgment stay where they are; the read tells them where to look.
Which product is this?
The same band that prices a pool before it is bought reads the book after it is funded. The product is collateral risk for loan pools, described at collateral risk analysis and residential loan pool analysis, and a funded book is scored the same way a candidate pool is, with the ranks locked and digested before you release anything.
Fifteen minutes, then a pilot on the seat that fits.
Thirty seconds is enough: the desk this lands on, rough size, and the workflow you have in mind. We reply within 24 hours with the brief for that seat and a time to talk.
Frequently Asked Questions
What do mortgage portfolio risk analytics show at the property level?
How the collateral across a book is positioned relative to its own local markets: where the weak band sits, which price tiers, vintages and originators hold it, how that looks by count and by balance, and where the evidence is too thin to support a read. Good Investment adds that lens on top of the loan-level data your team already holds.
How is this different from a metro-level risk view?
A metro can read as stable while specific neighborhoods inside it weaken, and a book that appears diversified across ZIP codes can be concentrated in individually weak collateral. Reading each property within its own market surfaces that concentration before it shows up at the next mark.
Does it replace our risk system or our marks?
No. It adds a property-level read that most risk stacks do not produce, delivered as a scored file against your own book and keyed on your identifiers. Your marks, your risk system and your credit judgment stay where they are, and the read tells them where to look first.
Can we drill from the book to a single loan?
Yes. Every row carries a score, a band, a coverage status and a review route, and the property report for any exposed loan carries the neighborhood context and the confidence flag behind the read.
How is the book handled?
It goes to a cloud project created for your engagement, separate from the project that runs our product, and it is scored there by a job that exists only while it runs. No copy touches an employee machine. When the work ends we delete the objects, then the project, and send an attestation naming what was removed and when.
