Candidate Pool Assessment: How to Screen a Pool Before You Bid

A candidate pool is defined by what has not happened yet. The loans are identified, the tape has been shared, and nothing is settled: rows can be excluded, the price can move, the structure can change. Everything a buyer establishes inside that window is leverage. Everything discovered after it closes is a lesson.

What makes a pool a candidate pool

The term describes a state, not a size. A candidate pool is any collection of loans or assets under evaluation for acquisition, financing, securitization or addition to a managed book, where the composition is genuinely still open. Full visibility into the pool's strengths and weaknesses is what lets a buyer negotiate price, request exclusions, or structure the purchase to match the actual risk profile. Once the trade is papered, all three of those become much more expensive.

That is the whole reason assessment is time-shaped rather than thoroughness-shaped. The question is not "what could we eventually learn about this pool," it is "what can we establish while it still changes the trade."

The stages of assessment, and what each produces

Exact sequencing varies by counterparty and asset class, but the shape is consistent, and so is the fact that the cheapest, broadest work happens first.

The practical consequence is that inputs which need only the tape are worth disproportionately more than inputs of similar quality that need files. Not because they are better, but because they arrive in time to matter.

What stratification shows — and what it cannot

Stratification is the backbone of pool review and it does its job well. It summarizes the population into distributions and lets you compare this pool against the last one on a consistent basis. Its limit is structural rather than methodological: it can only cut on fields that exist.

On a residential tape, the loan and borrower fields are numerous — balance, rate, term, amortization, maturity, leverage ratios, credit score, liquidity, sponsor experience, entity structure, internal grade. The property fields are typically an address, a property type, a value, and sometimes square footage and year built. Three groups describe the obligation and the obligor in depth. The fourth carries an address and a number. Every property-level aggregate in the review is built from that fourth group.

Three failure modes follow, and they show up in almost every pool:

• Geography is not a market. A pool spanning many metros reads as diversified, but properties compete for buyers at a much smaller scale than an MSA. Correlated submarket exposure survives a state-level cut intact. This is the subject of geographic concentration risk.
• Averages absorb their own tails. A pool can carry an unremarkable weighted average while its weakest collateral clusters in one submarket or one vintage. The aggregate hides the cluster until the loans resolve.
• Count and balance disagree. A segment that is a modest share by loan count can be a large share by balance. A composition view reported on one basis understates exposure roughly half the time.

The full field-by-field version of this argument is in what aggregate loan-tape metrics miss.

Exclusion logic and the three levers

A candidate pool has exactly three things that can still change, and a finding is only useful if it maps onto one of them.

Exclusion criteria themselves are always institution-specific and always the buyer's call. Common categories include loans outside stated credit or leverage parameters, property types outside the mandate, geographies the buyer does not hold, unresolved title or data issues, and diligence findings that cannot be cured. Analytics inform which rows deserve a closer look. They do not determine eligibility, and any tool that presents itself as deciding eligibility is overreaching.

Where a property-level pass fits

A property-level read runs from the tape, because it needs only an address and basic property attributes. That places it in the pre-bid window alongside stratification, rather than in the post-bid file review — which is the whole argument for it operationally.

What it contributes to a candidate-pool review:

A within-market rank on every row

Each property is positioned against comparable properties in its own local market rather than a national or regional benchmark. Because the output is a relative position, rows from different geographies become comparable to one another on a consistent basis — which is precisely what a metro-level average cannot deliver. The answer is always a rank; it is not a predicted price, a forecast, or a projected return.

Coverage, reported neutrally

Some properties cannot be assessed with sufficient local evidence: thin transaction history, unusual property types, incomplete recorded characteristics, recent major renovation, new construction with no local record. Those rows should be marked unsupported and routed to a person rather than assigned a number with invisible error bars. Coverage is not a defect rate and should not be read as one — it is the analysis reporting the limit of its own evidence, and where unsupported rows cluster is itself a finding about the pool.

Tails rather than a pool average

Property-level signal is strongest at the extremes. The useful output is four groups — a strong supported group, a weak supported group, a middle that stays in the standard process, and an unsupported group — not a single pool score. A pool average computed from property ranks would discard the only part of the distribution that carries information.

Divergence flags

Where the local property read and the value or leverage on the tape point in different directions, that disagreement is worth surfacing. It does not establish that either input is wrong; it identifies rows where two independent views of the same asset do not agree, which is a sensible way to spend a diligence hour.

Questions worth asking about any pool

• Does the geographic cut reflect submarkets or administrative boundaries, and would the diversification claim survive a finer cut?
• Does every concentration finding read the same by count and by balance?
• How old are the valuations behind the leverage ratios, and is mark age evenly distributed or clustered?
• Which valuation instrument produced each value, and does the cheapest instrument correlate with the hardest-to-value properties?
• What share of the pool is genuinely unsupported by local evidence, and where does it cluster?
• Where do independent views of the same property disagree?
• For any analytic in the process: has it been tested incrementally against the variables you already use?

What analytics should not decide

A property-level read is a narrow input, and the distinctions matter more than the capability:

• It is not diligence. It does not review files, verify data, or produce graded findings, and it does not replace any of the four established review streams.
• It is not valuation. It does not appraise a property or establish an after-repair value.
• It is not a credit opinion. Credit risk depends on the borrower, the structure, the servicing and the project as well as the property.
• It is not a rating. Rating a transaction involves loss distributions, structural analysis and credit enhancement. This is an input several steps removed.
• It is not an eligibility decision. What belongs in a pool stays with the institution.

And before any of it enters a workflow, it should be tested on your own historical population under a blind protocol — agree the measures first, deliver ranks using only information available as of the decision date, lock the output, then unblind. The full protocol is in the institutional guide to residential property risk assessment.

The narrow version of the argument

Candidate-pool assessment is a race between what you can establish and how long the composition stays open. Stratification and pricing models do most of the work and do it well, within the limit that they can only describe fields the tape contains. The property fields are thin, so the property-level variation that ultimately determines recovery is not among them. Adding a read that runs from the same tape, in the same window, does not replace any part of the process — it uses the window better.