Asset-pool and tranched credit framework
This framework rates a defined pool of assets held in a vehicle with no business to fail. The risk is the pool's own loss distribution, and a claim attaches at some point on that distribution. Scope is off-chain credit pools: receivables, SME and consumer loans, leases, CRE, private-credit portfolios and tranched claims on them.
Nothing here is a company. The issuer is a vehicle with no employees, no strategy and no discretion: it holds a defined pool and pays according to a contract. It cannot fail in the sense that a trading firm fails. So there is no business-risk axis, no management assessment and no liquidity modifier.
What is rated is a slice of a loss distribution. A tranche is defined by an attachment point, below which it is untouched, and a detachment point, above which it is wiped out. Everything else is detail on how the distribution is obtained and how the result is converted into a letter.
Framework Architecture
Six layers. The first is binary; everything downstream of it assumes it passed.
| Layer | What it does |
|---|---|
| Layer 0 · Legal isolation | Four legal conditions, tested before any quantitative work. |
| Layer 1 · Portfolio anchor | Produces four numbers that fully specify the loss distribution: portfolio default rate, portfolio loss given default, effective obligor count, and correlation |
| Layer 2 · Loss distribution | A one-factor correlated model run on those four numbers. Loss volatility falls out of it rather than being sourced separately |
| Layer 3 · Pool rating | The analogue of an issuer credit rating: what the whole pool would be rated, unlevered and untranched. Two axes, five modifiers |
| Layer 4 · Enhancement and waterfall | Where structure enters. Moves a tranche several notches with the pool held constant |
| Layer 5 · Caps | On this side, caps bind more often than the pool analysis does |
| Layer 6 · Facility rating | Converts the distribution into a rating on the specific claim, and reports the quantities that a single letter conceals |
The pool rating at Layer 3 is retained as a summary statistic of the distribution, so that a pool can be quoted while a capital structure is still being negotiated and so that deals with different capital stacks remain comparable. It is a summary of the tranche calculation rather than an input to it; if it ever begins to drive tranche outcomes, that is a defect.
Layer 0. Legal isolation and bankruptcy remoteness
Four conditions, tested first. They are not scored and they do not trade off against each other. A failure at this layer means there is no pool to rate; the correct response is to say so and route the assessment to the corporate framework on the originator.
| Condition | What has to hold |
|---|---|
| True sale | The receivables have left the originator's balance sheet rather than being pledged against a loan. The opinion has to survive recharacterisation risk, where a court re-reads the transaction as a secured loan. That turns on whether the originator retained the credit risk, the servicing economics and a repurchase obligation. A deep first-loss position retained by the seller is the fact pattern that attracts the argument |
| Non-consolidation | A court will not pull the vehicle's assets back into the originator's insolvency estate. This turns on behaviour as much as on drafting, separate books, arm's-length dealings, no commingled cash, no holding out as a single enterprise, and it degrades over the life of a deal in a way the closing opinion does not capture |
| Bankruptcy remoteness | The vehicle cannot incur other debt, has no employees, has an independent director whose consent is required for a voluntary filing, and has limited-recourse and non-petition language binding every counterparty |
| Continuity of servicing | Collections keep arriving if the originator fails. This requires a named back-up servicer or a demonstrated ability to appoint one, plus loan data in a form another servicer can ingest. A back-up servicer clause with no data-transferability test carries no value |
Layer 1. The portfolio anchor
The distribution is built at portfolio level. The model consumes those four inputs and does not require obligor identities. Obligor-level analysis is therefore an exception, taken where there is both a specific reason and a specific way to do it.
| Reasoning | Evidence |
|---|---|
| A framework whose main route requires obligor-level data produces "cannot rate" as its default answer, which is honest and useless. Obligor-level data almost never arrives with an inbound deal; balances always do | A pool of eleven exposures generates wide loss dispersion from its concentration alone. Establishing that requires only the balance schedule, not obligor ratings or vintage history |
The anchor panel
The portfolio default rate is triangulated from up to four methods rather than derived once. No single method is authoritative.
| Method | What it reads | Preference |
|---|---|---|
| A · Realised default rate | The issuer's own tape, on a static-pool basis | Primary where available |
| B · Sector and geography reference | An industry-risk × geographic-risk grid, mapped to a letter and then to a default rate | Always run |
| D · Bottom-up obligor PD | Exposure-weighted obligor ratings or scores | Only when detailed in the loan tape |
Method B is always available, which is what makes the panel usable on a thin file. Methods A and B together are the minimum for a published rating.
The definition of default used on the pool side is as per the originator definition. That definition may be different from Credora's internal definition and adaptations may be used to reach it.
The other three inputs
Portfolio loss given default. Level and range, from realised recoveries on the tape where they exist, otherwise from asset-class reference data. The range carries as much weight as the level, because a hard maximum on severity becomes a hard ceiling on pool loss.
Effective obligor count. The inverse Herfindahl of exposure weights. Measured, available from any tape that lists balances, and load-bearing.
Portfolio Stratification. Obtained from the loan data tape.
Reading the stratification
The four inputs are read off a stratification, and the cuts that matter are the ones that move them: obligor count and concentration; seasoning, meaning how much of the loss curve has already run; sector and geography, which is the only observable handle on correlation; weighted LTV or debt-service coverage; and the asset-specific driver.
Layer 2. The loss distribution
One method: a one-factor correlated model, run on the four numbers from Layer 1. Loss volatility falls out of the model rather than being sourced separately, which is what makes a pool rating computable on a file with no vintage history. For genuinely granular pools, published rating-scenario stress multiples serve as a cross-check on the result rather than as a second route to it.
The standard copula formulation has zero tail dependence: conditional on an extreme common factor, defaults still decouple in the limit. Real credit does the opposite: the worse the environment, the more tightly outcomes bunch. The model therefore understates the tail at every correlation level, and understates it most in the region a senior tranche occupies.
A heavier-tailed copula, a mixture of factor loadings, a heavier-tailed common factor, each add a parameter, and none of those parameters is more observable than correlation itself. Rather than trading one unobservable parameter for two, the framework treats tail dependence as a known direction of error: modelled senior tranche loss is read as a floor rather than an estimate. That supports reluctance about the top of the scale on concentrated pools, and it is a second independent argument for the correlation-uncertainty cap.
Layer 3. The pool rating
The analogue of an issuer credit rating: what the whole pool would be rated, unlevered and untranched. Two axes, five modifiers, and the same grammar as the corporate anchor so that the two remain comparable.
Pool modifiers
Originator and servicer quality. The pool was underwritten by someone and is collected by someone, and both can change what the vintage curves mean. This is the closest analogue to the corporate framework's management axis, and it earns its weight for the same reason: it is where the tail lives.
Data sufficiency. Good data does not make a pool better; it makes the estimate of the pool trustworthy, and that is already priced by the absence of a penalty.
Obligor concentration. Driven by the effective obligor count against the pool's nominal count, plus the single largest exposure. Concentration is not harmful in expectation; it converts a distribution into a small number of bets, which raises loss volatility and is therefore already partly counted on the anchor's second axis.
Performance track record. Realised against expected loss on this originator's prior series, and on this pool where it is seasoned. The only modifier with a genuine empirical handle.
Layer 4. Enhancement and the waterfall
Half of credit enhancement can be expressed as an attachment point and half cannot. Getting that distinction wrong is how a structure with thin subordination and thick excess spread gets rated as though it were the reverse.
Layer 5. Caps
The rating is the worse of the modelled outcome and every applicable cap. On this side caps bind more often than the pool analysis does. The credit analysis of the pool determines the mezzanine; the caps determine the senior: a senior tranche at high coverage does not touch an obligor in any plausible scenario, and its rating is a function of whether the account bank fails, whether the swap counterparty fails, whether collections are commingled with an insolvent servicer's operating cash, and whether the sovereign imposes controls. Caps include, but are not limited to: Counterparties, Servicer, Sovereign and country, Legal Enforcements.
Layer 6. The facility rating
One pool, many claims. A claim is defined by two numbers, an attachment point and a detachment point, and this layer converts the loss distribution into a letter on that specific claim.
| Quantity | Definition | Role |
|---|---|---|
| Attachment probability | The probability that pool losses exceed the attachment point | The rated quantity |
| Expected loss | The average loss across the tranche's band, as a share of the tranche | Reported |
| Conditional severity | Expected loss divided by attachment probability | Reported: the thinness signal |
A rating that discloses only the probability conceals how much is lost when the event occurs. All three are reported. Attachment probability maps onto the same rating scale used throughout this documentation, see DeFi Rating Scale.
Committee overlay
The bounded, auditable home for discretion, applied once, at facility level: written rationale, peer validation, logged. The rate and direction of overrides is itself tracked, because the pressure on an override is structurally upward.