Collateral risk analysis has to reach the house, and a loan tape stops at the address.

What is collateral risk analysis at the property level?

Collateral risk analysis asks what happens to the lender if the borrower stops paying. Credit risk asks how likely that is. They are separate assessments, and residential lending has far better tooling for the second than the first. Borrower credit files are dense, standardized and refreshed continuously. The property behind the loan is usually represented by an address, a type and a value.

For a residential file, collateral review covers four things: whether the stated value is supported, whether the property is what the file says it is, whether title and condition carry any impairment, and how far the value can move before the position is exposed. The first three are questions about the present, and the established toolkit answers them well. The fourth is a question about where the house is going, and today it is answered with the rate for the whole area.

That fourth question is where Good Investment sits. Appraisal and file verification stay with your team. The score places the property inside its own local market, reads that position as a rate of appreciation above or below the area, and attaches neighborhood context and a confidence flag.

Why does a loan tape carry no collateral risk?

A tape describes loans in depth and properties in about three fields, and every aggregate computed from it inherits that asymmetry. Weighted-average LTV, FICO bands, state mix and vintage distribution describe the obligation and the obligor. They say nothing about how any single house is positioned among the homes it competes with, because that field is not on the tape.

Two properties can carry the same value and the same loan-to-value ratio while sitting on very different paths. One is a well-positioned home in a neighborhood with durable demand. The other is the soft example of its type in an area that quietly lags the metro. A valuation treats them as equivalent, and the area rate every model applies treats them as equivalent too. A within-market read separates them.

The panel below is the read a collateral reviewer gets on a single file. It places the property among comparable homes in its own market and shows where in that distribution it lands, with the basis it was scored on labelled on the panel itself.

How does a property-level appreciation read enter a strat table?

As one more bucket dimension. A bid desk already stratifies a tape into FICO bands, LTV bands, occupancy, documentation type and state, assigns an expected loss to each bucket, and rolls the buckets up through a cashflow model to a price. The property read arrives as a band on every row, so it sits beside the bands you already cut and flows through your own loss and pricing models. An adjustment applied after a blended price is a committee argument; an extra strat is a Tuesday.

What we hand a desk is the expected loss differential between the bands, measured on resolved history and carried with its interval. The desk turns it into price with its own hold, discount rate and transmission view. The loss differential is the product; the price conversion is yours.

The score is measured on property appreciation, the thing the tape leaves out, and it is benchmarked against the columns you already hold. Removing the same share of a tape by LTV reproduced none of the effect, and FICO moved expected loss by almost nothing on the same test, which is what a property signal should look like beside two borrower signals.

What does a bid desk do with a band?

One of two things, never both. The first is to bid the weakest band down and the strongest band up, so the blended price stops paying the same for collateral positioned to slip as for collateral positioned to hold. Re-pricing a band moves money between buckets and creates no loss on its own; the gain is that the seller no longer keeps the difference.

The second is to drop the weakest band onto the kick list you already run, and take the expected loss it carries out of the book. That is a different figure on a different basis, and a desk chooses one or the other. Adding them is the first error diligence finds, so we do not print them side by side and neither should a memo.

In a competitive bid, dispersion is expensive in a way that averages hide. Bidding against others, you disproportionately win what you overvalued, and one blended price overvalues exactly the band the score finds.

Where does the exposure concentrate?

In the weakest band, and unevenly. When a pool is priced at one blended number, most of the mispricing sits in the bottom band of the score, which holds a smaller share of balance than its share of the miss. That is why one column added to a strat earns its place: the correction is concentrated where a desk can act on it, and the middle bands stay in the standard process.

Applying the same read across a book changes what the composition question can be computed on. Every property is ranked against comparable homes in its own local market, so a row in Ohio and a row in Arizona sit on one axis. Concentration is then measured on something other than location, such as whether weak-ranked rows cluster by price tier, vintage or originator.

Report it on both bases. A segment that is a small share of a book by loan count can be a large share by balance, and a report on a single basis understates exposure roughly half the time. The gap between the two is a finding in its own right.

What happens when the evidence is thin?

Any analytic run across a whole population meets properties it cannot assess well. Rural and low-turnover markets go long stretches without enough comparable activity to support a local read. Unusual configurations have few genuine local competitors. Recorded characteristics go stale after a renovation, and new construction has no local record for the specific asset.

Those rows come back marked unsupported, with the reason named from a controlled list, and routed to a person. Coverage is reported as an output, at the file level and by segment, because a book whose unsupported rows cluster in one market, one originator or one vintage is telling you where your evidence is weakest before a single scored row has been examined.

The read supplies a position within a local market and a coverage status on every row. Appraisal, AVM and after-repair value keep their jobs and the read puts no price on anything. Credit keeps its job, since default turns on the borrower, the structure and the servicing as much as on the property. Ratings sit several steps downstream, and which files belong in a book stays with your team.

What does a pilot look like?

Four lines. You send one resolved tape, keyed on your own identifiers, with the outcomes held back. We score every property as of its original decision date, using only what existed before it, and return a score, a band and a coverage status per row. The ranks are locked and digested before you release anything, so a rank changed afterwards would stop matching the digest you hold. You join the outcomes and grade the result against the variables you already use.

Two weeks, and a negative result is a result you keep. The debt brief, the model card and the validation pack go out on the first call, and a model risk team gets the full internal pack under NDA as a paid engagement.