Industry field guide
Secondary-market loan tape revisions for AI testing
Two loan tapes may represent different dates or corrections to the same date.
- Sector
- Finance and insurance
- 2022 NAICS
- 522299
- Practice pack
- Example and answer key (JSON)
- Updated
What this guide tests
Reconstruct the corrected day-10 inventory and compare its count with the day-20 inventory without double-counting versions.
See the specimen and fields · Work through the fictional records
Task and source context
Two loan tapes may represent different dates or corrections to the same date. Combining them can double-count exposures. FHFA’s public use database illustrates public mortgage data and its documented fields. A secondary-market operator can make a mock inventory with intentionally different versions. FHFA public use database
Make two fictional snapshots with invented asset IDs and a correction file. Add one revision that changes the status of an existing row and one genuine new asset. Ask the assistant to reconstruct the snapshot as of a stated date and explain why a corrected record is not an additional loan. Have a reviewer confirm expected counts and visible changes before testing.
Do not copy borrower data, property or loan IDs, investor reports, pricing, or licensed loan tape data. Use clearly synthetic terms and no underwriting decisions.
Industry research
The U.S. Census Bureau’s industry definitions establish which activities these codes cover. The workflow sources in this guide describe the task or record type. Fiund’s pilot below uses fictional material to test a specific decision within that scope.
522299 · International, Secondary Market, and All Other Nondepository Credit Intermediation
This U.S. industry comprises (1) establishments primarily engaged in providing working capital funds to U.S. exporters, lending funds to foreign buyers of U.S. goods, and/or lending funds to domestic buyers of imported goods; (2) establishments primarily engaged in buying, pooling, and repackaging loans for sale to others on the secondary market; and (3) establishments primarily providing other nondepository credit (except credit card issuing, sales financing, consumer lending, and real estate credit).
Build a pilot sample
Set up this pilot
Build two dummy loan-tape snapshots and an authorized correction to one old row. Count assets by stable key and retain snapshots rather than appending versions.
Prepare the test packets
- Start with the worked case. Use the fictional record types below as a guide to the source files you need. Give each packet a stable ID and save its source versions with it.
- Change one fact. Make one packet with enough evidence to answer the question and one where the evidence points to a different answer. Keep the other fields fixed so a reviewer can tell what changed.
- Remove the decisive evidence. Give the reviewer a third packet with a plausible near-match in its place. Have them write what can still be established and what evidence would resolve the rest.
Fields the reviewer needs
- Snapshot date and parent
- Defines the inventory time.
- Asset alias and correction target
- Preserves stable identity.
- Status, replacement rule and distinct count
- Makes the reconstruction checkable.
Use How to prepare a financial workflow data sample for AI to turn the specimen into a brief with a defined task, source and review owner.
Fictional exercise
A corrected snapshot row and a new asset affect counts differently
A synthetic inventory uses invented asset aliases and simple status labels. No borrower, property, actual loan term or licensed loan tape is present.
| Record ID | Source | What it records |
|---|---|---|
| LT1 | Snapshot P1 | As of day 10: asset A active, asset B pending. Snapshot count 2. |
| LT2 | P1 correction file | Explicitly replaces B's day-10 status with active. No new asset row; references LT1/B. |
| LT3 | Snapshot P2 | As of day 20: A active, B active, new asset C active. P2 is a later snapshot, not a correction to P1. |
Question for the model. Reconstruct the corrected day-10 inventory and compare its count with the day-20 inventory without double-counting versions.
Show the reference answer
LT2 updates B's status within LT1's day-10 snapshot, leaving two assets, A and B, both active. It does not create a third asset. LT3's later day-20 snapshot adds distinct C, giving three assets for that date. Keep snapshots separate and apply the correction to its named parent rather than appending every row from all versions into one inventory. The fixture supplies no pricing or underwriting conclusion.
How to check the result
Ask for the answer, the record IDs that support it and any fact the evidence leaves unresolved. Have a reviewer compare the response against the reference before deciding whether the sample is ready to expand.
Errors to flag in this case
- Appending a corrected row as an additional exposure.
- Combining later snapshots with earlier ones and counting the same asset repeatedly.
Change the inputs and run it again
Add a correction for B dated day 20 with no snapshot-parent reference. Require a scope clarification before applying it to P1 or P2.
Score the decision and the evidence separately
Save the reviewer’s expected answer, supporting record IDs and unresolved fields in a separate answer key. Score the model on the decision, its citations and its handling of the deliberate gap. A correct answer citing a different case counts as an evidence error.
Before expanding collection, have a second reviewer run the same checks on the held-out packet. Keep related versions, views and exports of the same case together so they stay on one side of the training and evaluation split.
Check permission and scope
Use original invented rows. Exclude actual IDs, investor reports, pricing and borrower information; no credit decision belongs in the benchmark.
Exclude real personal and financial data, credentials, underwriting decisions and individual eligibility conclusions.
Describe an authorized archive through Fiund’s existing-data assessment, or describe a new capture capability through the collection brief. The contributor agreement covers the terms for participation.
Sources and further reading
Resources for this workflow
Includes worksheet
How to prepare a financial workflow data sample for AI
Build a fictional or authorized finance example with the reporting unit and time cutoff intact.
Includes worksheet
Plan a sample of business data for AI
Describe the evidence without transmitting private account or customer information.
Company and program reviews
Use these reviews to compare sourcing, capture or evaluation approaches. Confirm the applicable program, accepted sources and terms with the provider.
Compare the approach
Scale AI review: data partnerships, Outlier and buyer services
Read the public financial-services partnership criteria and scope controls.
Compare the approach
Invisible Technologies review for expert-led AI training
Compare expert calibration and reviewer agreement for narrow financial document exercises.
Related industry examples
Industry guide · NAICS 522292
Real estate lender payoff records for AI evaluation
A principal balance and a payoff quote are different records.
Industry guide · NAICS 522291
Consumer installment payment records for AI review
A payment receipt confirms an amount received.
Industry guide · NAICS 522210
Credit card dispute status records for AI review
A credit appearing on a statement can be temporary.