Commoditizing Cashflow Models, Starting With EDGAR Filings

1. Why a cashflow model costs a fortune

For a public deal, nothing in the model is secret. The capital structure, the payment priority, the triggers and the loans are printed in the prospectus. The monthly loan data and what the trustee paid each class are filed with the SEC. What has been scarce is the labour: someone reads a 750-page prospectus, codes the waterfall, ties it to the decrement tables, and keeps it current as the deal pays down. That work is why a model arrives as a licence to an encrypted file, priced per deal and per seat.

Agents change the cost of that labour. When reading the documents, building the model and checking it are done by software in minutes, the model stops being the product. It becomes a commodity, and two questions are left: can you trust it, and can you take it with you?

2. EDGAR is the supply

Since Regulation AB II took effect in late 2016, every public securitization files with the SEC on the same schedule: a prospectus (Form 424B) when it prices, then every month the trustee's report of what each class was paid (Form 10-D) and, for CMBS and auto deals, a file with every loan in the pool (Form ABS-EE). We built a registry of every filer, classed by the asset type each filing declares in its own header. This is what it holds:

And it keeps arriving. In the third quarter of 2026 alone, EDGAR's filing index shows 920 trusts filing 2,852 trustee reports, and 706 filing 2,426 loan-level files: 380 CMBS trusts and 326 auto. Every one of those is another month of evidence a model can be checked against.

The private market shows up too, but only by name. 303 RMBS trusts, all but two of them sold under Rule 144A and the non-QM shelves among them, appear on EDGAR solely through Form ABS-15G, a repurchase disclosure. Their offering memoranda and trustee reports never reach the SEC, which is why Graam takes those documents by upload.

Graam's agents turn that standing flow of filings into models, one deal at a time when you click Structure, and then hold each model to the trustee's reports, statement by statement.

3. How a filing becomes a model

The pipeline in the graph has four steps. Agents do the reading and the reasoning; the arithmetic is done by the engine, and every check is code comparing one number with another.

• Read. The prospectus is parsed as filed: for C44, all 754 pages and 263 tables in 21 seconds. Agents take the capital structure, the payment priority, the triggers and the loan terms from it, each with the page it came from.
• Build the cashflow model. The deal is assembled from what was read: the classes and their coupons, including the WAC strips and interest-only notionals CMBS is full of, the waterfall, triggers, reserve accounts, overcollateralization, and the loans themselves.
• Validate and repair. The model runs through the engine and is held to three things: cashflow invariants (every note retires, every interest-only strip is paid), the prospectus's own sensitivity tables, and the trustee's reports. A break sends it back. The agent diagnoses it, finds the provision that settles it, and makes one edit, which has to cite the page. The edit is kept only if no measure gets worse, and the model is validated again.
• Independent review. A separate agent, which did not build the deal, reviews the result. It passes, or it goes to a human with every remaining break listed.

4. The proof: one deal, checked against everything it filed

C44 has 19 classes over 44 loans. The classes add up to $766.7 million, the pool's cut-off balance to the dollar. The model was then checked against everything the deal has filed:

The first check failed

Validation is where a model earns trust, so it is worth seeing one fail. The first replay found 21 class-months that didn't tie, $1.44 million in all. Twenty of them were in June 2018, and every class was off by exactly eight days of interest. The model had accrued from the closing date, 38 days, where the trustee paid a 30-day first period. The rule was in the prospectus, but printed across a page break: the day-count basis at the foot of page 290, the definition of the first period at the top of page 291. The reader looked for both on one page. Now it reads the page beside it too.

The 21st was H-RR in January 2021, paid $13,484 more than its interest for the month while its shortfall column showed zero. That was a payback of earlier shortfalls, which the check now recognises as its own event.

The second check passed

Every month that doesn't tie lines up with an event the trustee printed: money held in the interest reserve, a shortfall on a lower class, a payback. That is the difference between a model you are asked to trust and a model that shows you its check.

5. The model is a file, and the engine is open

Open the deal's Deal Json tab and the whole model is there in one file: the 19 classes, the waterfall, the triggers, the prospectus's WAL table, and all 44 loans with their interest-only, lockout and defeasance terms. Search it, read it, download it. It is your file, not a licence to look at someone else's encrypted one.

Every cashflow in this post comes from graam-flows, the engine behind Graam, and graam-flows is open source. The collateral projection, the waterfall and the bond math are code anyone can read.

Put the pieces together. The structure comes from public filings, the check comes from public trustee reports, the engine is public code, and the work of building the model takes minutes. That is what a commodity looks like. What you pay for should be the analysis you do with the model, not access to it. Why we opened the engine is in Open Sourcing the Execution Layer of Structured Credit.

6. Ask it what happens next

A model you own is only useful if it answers questions, so every deal comes with a conversational agent that runs on it. C44 has live ones. At the September 2026 remittance, four of its 41 remaining loans, $97.9 million or 14.5% of the pool, were delinquent or in special servicing, and most of the pool matures in 2028.

Ask the agent to run 2% CDR at a 40% severity with no prepayments to maturity and show which classes take a loss, or to compare that with a faster prepayment case. It runs each scenario on the same model that tied to 100 trustee reports, starting from the balances the trustee reported for whichever month you pick, and comes back with the bond cashflows, the loan-level projections and the loss by class. For the collateral side in more depth, on another conduit, see A CMBS Deal, Read From EDGAR.

7. Coverage, and what is not covered yet

From EDGAR, Graam builds public CMBS conduits and auto loan and lease ABS, the shelves that file loan-level data, along with equipment ABS from its prospectus and trustee reports. Non-QM RMBS is sold under Rule 144A and never reaches EDGAR; for those deals Graam works from the offering memorandum and loan tape you upload. Credit card and student loan trusts are on EDGAR but file no loan-level data, so they are not covered yet.

Not every deal passes its checks on the first build. Each deal's Build Status tab shows where it stands, check by check, so you know which models to rely on.

8. Try it yourself

1. Sign up. It takes a minute, and the EDGAR data is already loaded.
2. Open EDGAR search and find a public CMBS or auto deal.
3. Click Structure and watch the Build Status tab fill in.
4. Ask the agent what happens next, and download the deal JSON.

You just built a cashflow model. Sign up and build yours.