Every filed figure should lead back to a loan file

In late 2021 my team was building a loan data mart that joined collateral records from our lien system with the core's loan categories. The comparison found 412 operating and livestock lines, totalling US$138 million, reported as loans to finance agricultural production. Each carried a farmland mortgage large enough to meet the call report's definition of a loan secured by real estate. Our core recorded the type of collateral but not its value at origination, so classification had followed the loan officer's purpose code.

The error did not change capital or earnings, but it misstated the composition of a federal return that examiners and peer analysts use, and it understated farmland-secured exposure in our own board concentration report.

I recommended amending the two most recent quarterly reports, for which the evidence was complete, rather than correcting the next filing only. Annelise Ostby, our chief financial officer, accepted once a 150-file sample confirmed the pattern. The chief credit officer objected strongly. The full review meant 1,960 files, pulling credit administration from operating-line renewals in their busiest six weeks. We gave up the launch date of the board's concentration dashboard and paid about US$84,000 in overtime and temporary help. Earlier quarters were not amended, because their lien records could not be matched reliably, and we explained that limit to our FDIC field office before filing.

Credit administration reviewed every file against our matched evidence. We made real-estate collateral value at origination a required field and added a quarterly reconciliation that flags disagreements before the return is prepared.

Both amended reports moved US$138 million from the production line to farmland-secured loans. Since then the reconciliation has flagged an average of six loans a quarter, each corrected before filing. The board's concentration report now shows exposure by collateral and by purpose separately.

What I would do differently: the chief credit officer heard the finding second-hand, after the chief financial officer, and much of his resistance was about that. Next time, the owner of the records hears first.

Each year my fair-lending team tests pricing on agricultural operating lines and small-business loans for our chief compliance officer, Theron Vasbinder. Our pricing grid lets lenders grant rate exceptions to meet competition. In the 2023 test of 2,870 loans with exceptions, one nine-branch region in south-west Kansas showed a gap. Borrowers whose principal owners were estimated, from surname and address, to be Hispanic received exceptions averaging 0.18 percentage points smaller than others, after controlling for deposits, collateral and credit grade. The region accounted for 610 of those loans.

At stake was possible unequal treatment under the Equal Credit Opportunity Act, measured by a method least reliable in exactly such small rural samples.

My lead analyst wanted three more months to refine the estimates before anything was reported. I decided to report the result at once, with its error range on the first page, and to say plainly that we could not yet explain it. The regional lending president disagreed with what followed. Theron recommended, and the management compliance committee approved, mandatory reason codes for every exception and a second approval above 0.50 percentage points. Lenders compete with dealer finance and other farm lenders, and they said the extra step would cost them borrowers. In the first six months they recorded 19 applicants who went elsewhere citing timing.

Compliance reviewed all 610 files while my team built reason codes into the origination system and a monthly exception report.

Most differences had documented, legitimate explanations, such as relationship deposits held at another branch. Forty-four loans had no recorded reason at all; the bank repriced them and refunded US$68,300 of interest. The following year's test showed a gap of 0.04 points, within the method's uncertainty.

Exceptions are now reviewed monthly rather than discovered annually. I would make that change before any finding, not after one. I would also show lenders the method earlier; part of their objection was to being judged by a model they had never seen.

Our seven-year agreement with Arbelune Core Systems, which processes our loans and deposits, came up for renewal in 2024. Annual core and processing fees were about US$7.4 million. Arbelune offered an 11% reduction for a new seven-year term with data access unchanged: a nightly extract of a fixed set of 640 fields, custom extracts billed at about US$96,000 a year, and definition changes announced only in release notes. I had worked at Arbelune until 2011 and hold no interest in it; I disclosed that history to our chief operating officer, and procurement led the negotiation.

The stakes were less visible than the fee. Our filings, fairness tests and models depend on fields Arbelune defines, and a silent change could alter a filed figure unnoticed.

I proposed accepting a 6% reduction in exchange for a full daily extract of the loan, deposit, collateral and customer tables, a maintained data dictionary, ninety days' notice of any definition change and a documented exit extract at no added charge. Procurement and Annelise Ostby preferred the larger discount, and they were right that the difference was real: about US$370,000 a year, or roughly US$2.6 million over the term. The chief operating officer decided in favour of the data terms.

My team drafted the contract's data schedule, mapped every field used in a return or model against the dictionary, and set up a review of each change notice with the affected data's owner. Retiring the custom extracts recovered US$96,000 a year of the forgone discount.

Since the new terms took effect in February 2025 we have received 17 advance notices. One would have changed how renewed operating lines stored their original origination date, which feeds both our loan-age reporting and our small-business application records. We adjusted before the release reached production.

I would start the data requirements a year before renewal rather than four months, and price each term separately. Finance could then have weighed each term rather than one bundle.