Every whole-loan bid ends in a number, and from the seller's side two very different processes can produce numbers that look alike. The difference shows up in three places: how much of the pool was actually examined, how long a re-cut takes, and which parts of the decision a person still owns.
The manual desk bid: a sample, extrapolated
A traditional desk prices from a stratification. The tape is bucketed by FICO band, LTV band, coupon, state and occupancy, and each bucket gets a spread drawn from recent comparable trades and the trader's read on where the market is. A sample of files is pulled for a closer look. The bucket-level marks are then rolled up, weighted by balance, into a pool-level price.
That process is not sloppy. It is how the secondary market has worked for decades, and an experienced trader carries pattern recognition no model has. But it has structural limits that are worth naming plainly:
- It prices buckets, not loans. Two loans in the same FICO/LTV cell can have materially different risk and both receive the same mark.
- It is expensive to repeat. If the seller carves out fifteen loans, or adds a hundred, the strats move and much of the work is done again.
- It rewards familiar collateral. Paper that does not fit the standard buckets tends to get a conservative mark, because conservatism is the cheapest way to handle uncertainty on a deadline.
Loan-level model pricing: every asset gets its own read
A model-based approach inverts the order of operations. Instead of scoring buckets and rolling up, it scores each loan on its own attributes and then aggregates. Every row in the tape gets a value; the pool price is the sum of what the individual assets are worth, not the average of the neighbourhoods they live in.
Three consequences follow, and they are the reason the approach exists:
Coverage instead of sampling
There is no sample. A thousand-loan tape is a thousand scored assets. That matters most on heterogeneous pools, where the loans that do not look like the others are precisely the ones a bucketed approach handles worst.
Re-cuts are cheap
Because the unit of work is the loan, changing the pool does not restart the analysis. Pull a hundred loans out and the remaining scores still stand; the aggregate is recomputed and that is the whole job. Sellers use this more than they expect to. The first tape is rarely the tape that trades.
The reasoning is inspectable
A loan-level score can be attributed back to the attributes that drove it. When a seller asks why a particular segment priced the way it did, the answer is a list of drivers rather than a judgement about a bucket.
What does not change: a person still signs the bid
This is the part most worth being clear about, because "AI-driven" is often read as "no humans involved." It is not how our desk works. Model output is a starting point that a trader reviews before any bid is released. The model does the volume work: reading every loan, applying consistent logic, holding the whole tape in view at once. The trader does the work that requires context the data does not contain: where the market has moved since the comparable trades, how a specific seller's underwriting has performed historically, what a servicing transfer will realistically cost, whether a data anomaly is a real risk or a formatting artefact.
The honest framing is that models change the economics of attention. When scoring a loan is cheap, you can afford to look at all of them, and the scarce human judgement gets spent on the handful of decisions that actually need it rather than on arithmetic.
What it means for you as a seller
- Data quality pays you directly. Under bucketed pricing, a missing field mostly moves a loan into a conservative bucket. Under loan-level pricing, a populated field is a chance for that specific loan to price on its own merits.
- Heterogeneous pools are less penalised. You do not have to pre-sort your production into clean, homogeneous lots to get a serious look.
- Turnaround compresses. Most of the elapsed time in a traditional process is queueing for analyst attention, not analysis. Removing the queue is where the days come from.
- Ask what happens to your data. A tape is competitively sensitive. Ours are used to price the trade in front of us and are not used to train models shared across counterparties, and any buyer should be willing to answer that question in writing.
The short version
A manual desk bid asks "what is a pool like this worth?" Loan-level model pricing asks "what is each of these loans worth, and what do they add up to?" Both are legitimate. The second one scales to every loan in the file, survives a re-cut without restarting, and leaves the human in the seat where judgement actually matters, which is why it returns a firm, executable number in days rather than weeks.
