Credit: auto ABS note classes

The note classes of an auto ABS: the trust as a container, collections as accounts, and ordered waterfalls paying seven classes by seniority, reconciled against the issuer's published percent-outstanding grid at every distribution date.

Every number below is checked against an independent reference implementation on every commit — period by period, and on each metric, inside a declared tolerance. See benchmark methodology.

The case

An auto-receivables trust issued seven classes of notes against a pool of car loans. Each month the borrowers pay interest and principal; the trust collects it, pays the servicer and the administrator, pays interest on each class at its coupon, and repays note principal strictly in order of seniority: nothing reaches the second class until the first is gone, nothing reaches the third until the second is gone, and so on down to the most subordinate. The question a noteholder asks is how fast their own class comes back, and that depends entirely on where they sit in the queue.

The reference

Exhibit 99.4 to the Form 8-K of Ally Auto Receivables Trust 2017-3, filed 17 September 2018. It states, for each of six classes and every monthly distribution date, the percent of that class still outstanding, and the assumptions the tables rest on: a 1.00% servicing fee, a $1,500 monthly administration fee, each class's coupon and day count, and no defaults, losses or repurchases.

auto_abs_wal reconciles the collateral in the same exhibit — 43 sub-pools amortizing to an aggregate the issuer states to the cent — and stops there, recording that the per-class columns need a priority of payments. This case is that axis, on the same 43 sub-pools unchanged.

Class principal amounts come from the trust's own Form 10-D. See SOURCE.md.

What it exercises

The trust as a container, the notes as claims on its cash, and the priority of payments as two ordered allocations from the two amounts the indenture defines.

Interest collected, net of the servicer's fee and the trust's own expense, is one account; principal collected is another. Each class is a credit.note the trust issued — a face, a coupon, and the account its holder's principal is paid into — and each distribution date allocates the first account to the classes' interest and the second to their principal, by seniority. Every holder owns an account that receives its principal, and that account IS the class's position: what a class is still owed is its face less what its holder has been paid, which the note lowers as its claim, so a step pays the claim and says which note and line it pays:

pay a3_principal to party.a3_holders for contract credit.note.a3 line principal =
      min(remaining, container.trust.credit_note_claim_a3)

Declaration order is seniority. A retired class contributes zero because its claim is zero, without being switched off. Nothing restates the waterfall and no class carries a balance of its own.

The published grid is therefore asserted directly, not by differencing: each class's percent outstanding is its face less its account, and the account balances sit in expected.csv beside the payments they explain.

The result

Worst disagreement 0.0054 percentage points, across all six classes and 208 published cells.

The exhibit rounds to 0.01, so 0.005 is the floor a reader can check against. This sits on it — the same place auto_abs_wal lands on the collateral. Of the 208 cells, 205 agree within that floor: the model reproduces the issuer's own printed number. Every class retires on exactly the grid's date.

The exhibit prints each class's weighted average life to maturity beneath its column. Folded from the principal each class's step pays, on the model's own axis, every one lands inside the print floor of 0.005:

Three cells exceed the floor, by 0.0003–0.0005 percentage points — C at 04/15/22, D at 07/15/22 and 08/15/22. Net of rounding, the disagreement those cells prove is at most $74 on the $537.6m pool.

Interest is paid in full on every class at every distribution: the coupons never exhaust the interest collected, which is what a deal with no losses should show. The interest the trust collects beyond the coupons, and the $13.75m by which the pool exceeds the notes, accumulate in the trust's own accounts and are reported as their balances.

The delta

That residue is traceable to the reference's own inputs, not the waterfall. The exhibit's pool table is exact where it can be — balances to the cent, integer remaining terms — but prints each pool's APR to three decimals, while the issuer ran on unrounded receivables data. Half of that last printed digit (±0.0005%) is enough to move tail cumulative principal by up to $248, concentrated in the seven large 51–53-month pools still amortizing in 2022; the $74 the cells prove fits inside it several times over. The signature agrees: the excess is small, one-signed, and appears only in the deal's final four months, in the last two classes to pay — accumulated input drift surfacing at the bottom of a sequential waterfall.

The engine's side of the ledger is clean. An independent month-by-month recursion from the printed table reproduces the pack's closed-form output exactly, and no payment-rounding convention (level payment to the cent, or payment, interest and balance together) moves the aggregate by more than $10. Nothing was tuned to close the residue: adjusting APRs within their printed half-digit would fit the benchmark to its own reference, and the model already sits at the information floor of the published data.

The exhibit's tables assume the receivables prepay at a constant ABS rate "with no defaults, losses or repurchases". With no losses, overcollateralization never has to build and no trigger can trip, so neither is modeled here. The clean-up call is not exercised: the tables asserted are the to-maturity columns.

This validates the priority of payments and nothing about the loss-driven machinery. That belongs to a deal that can lose money.

model.total is a regression anchor from this model, not an external figure. Every external assertion is a per-period class column in expected.csv, derived from the published grid.

Run configuration

{"deterministic":{"annual_discount_rate":0.03}}

Verified results

Checked period by period: 12 series across 48 periods — 496 values in all, each within the tolerance shown.

• notes.principal.a2_principal — within ±11202.66
• notes.principal.a3_principal — within ±27137.0
• notes.principal.a4_principal — within ±8601.0
• notes.principal.b_principal — within ±2222.0
• notes.principal.c_principal — within ±1851.0
• notes.principal.d_principal — within ±1375.0
• account.a2_principal — within ±11202.66
• account.a3_principal — within ±27137.0
• account.a4_principal — within ±8601.0
• account.b_principal — within ±2222.0
• account.c_principal — within ±1851.0
• account.d_principal — within ±1375.0

Summary metrics for the base run: