Where a bank's figures acquire their meaning

I treat the call report as the end of a chain that starts with a lender's entry. My team maintains lineage for every loan and deposit line we prepare: which core field feeds it, what transformation applies and who owns the definition. Before each filing we run reconciliations comparing collateral records with loan categories, maturity data with repricing buckets and nonaccrual flags with payment histories. Differences go to the business owner, not to my analysts to settle quietly. The chief financial officer signs the return and directors attest to it, so the people relying on the figures must be able to see where judgement was applied. I report exception counts each quarter to the audit committee. I also keep a register of known weaknesses not yet fixed, because a clean filing can still rest on a field nobody verifies.

For most agricultural and business credit, a lender may not ask the applicant's race or ethnicity, so testing relies on statistical estimates drawn from surnames and addresses. I build those tests with the chief compliance officer, who owns the conclusions. We examine approval, pricing and exception patterns, controlling for the factors our credit policy recognises, and we publish the error range beside every result. An estimate can suggest a pattern worth investigating; it cannot establish how any individual was treated. When a gap appears, compliance reviews the files and decides on remediation. My contribution is a method that is reproducible, documented well enough for examiners to rerun, and honest about what it cannot see, including small rural samples where one loan can move a percentage by several points.

The Consumer Financial Protection Bureau's small-business lending data rule, issued under section 1071 of the Dodd-Frank Act, will require lenders of our size to collect and report information about credit applications from small businesses and farms, including owners' self-reported demographics. Its compliance dates have moved more than once and its content has been revisited, so I plan for the practice rather than a fixed deadline. We have mapped where an application begins in each channel, because an agricultural operating request often starts as a conversation in a pickup rather than on a form. Collection is designed so that the people deciding credit do not see demographic answers unless that is unavoidable. I would not build an expensive process around one assumed date. I would build data that shows accurately who asked us for credit and what happened next.

Most of our data begins inside systems we rent. Our core processor, loan origination platform and online banking provider each define fields their own way, and a release can alter a definition without changing its name. Having led conversions for a core vendor, I read contracts for data rights as closely as for price: the completeness and frequency of extracts, a maintained data dictionary, notice before definitions change, and the format of data at exit. I keep an inventory of every vendor-supplied field used in a regulatory return or a model. Procurement and the chief financial officer own the commercial decision. My role is to show what a cheaper term would cost us in reconciliation work, custom extracts and the risk of filing a figure whose meaning shifted beneath us without warning.

I use machine-learning methods where a person can verify each output. Our most useful tool flags loans whose collateral records, purpose codes and amounts look inconsistent, then routes them to credit administration for review; it decides nothing. Our credit and pricing models sit under the bank's model risk process, and my team owns the data each one consumes, including how its development sample was drawn. I have been cautious with generative tools. A vendor offered to draft credit memo summaries from borrowers' tax returns and farm financial statements. The drafts read fluently, but in our trial they sometimes misread a schedule and presented an inferred number as a reported one. Until someone can show how every such error would be caught before a lender relies on it, I would rather lenders read the return themselves.