An optimised mortgage campaign inherits the bank's old map

For most of the history of American lending, where a bank advertised was a decision someone made: which newspapers, which radio stations, which neighbourhoods received branches and billboards. Redlining cases have long used those decisions as evidence. Today the decision is increasingly made by bidding software tuned towards the cheapest application and by audiences modelled on the bank's existing customers. My argument is that this does not make the geography of credit advertising anyone else's responsibility. It makes it the chief marketing officer's, and it belongs in fair-lending reporting rather than in a buying-efficiency review.

How a model learns a map nobody drew

In March 2024 our fair-lending officer at Ledgemoor showed me her annual comparison of our mortgage applications with those of peer lenders, built from public home-mortgage disclosure data. In the Milwaukee metro, 4.9% of our 2023 applications came from majority-Black and majority-Hispanic census tracts. Lenders of similar size averaged 12.8%. Our underwriting and pricing reviews had found no unexplained differences in outcomes. The difference lay earlier, in who applied, and who applied depended largely on who saw our advertising.

When we pulled the delivery data, 3.1% of our digital mortgage impressions in the metro had reached those tracts. Nobody at the bank or the agency had excluded a neighbourhood. Our programmatic buying set bids by ZIP code according to past cost per application, and it had learnt that applications were cheapest in the suburbs where most of our borrowers already lived. A modelled audience bought from a data vendor, built to find people resembling our existing mortgage customers, reinforced the pattern. Each tool did what it was asked. Together they had rebuilt in software a map that a loan committee of the 1960s might have drawn by hand.

I want to be precise about what this was. It was not evidence that anyone intended to discriminate, and I do not claim it was unlawful. It was evidence that a cost-tuned campaign, left alone, reproduces the bank's past, including whatever was unequal about it. Fair-lending law covers the whole credit process, and redlining enforcement has examined where lenders market as well as how they lend. An explanation that the algorithm chose the audience would not satisfy a supervisor, and it should not satisfy a board.

Report delivery by census tract to the people who own fair lending

The first change I would urge on any bank is reporting. Marketing teams report impressions, clicks, cost per application and return on spend. Very few report delivery by census-tract demographics, because nobody asks them to. At Ledgemoor every campaign for mortgage, home-equity or small-business credit now produces a monthly report comparing impressions, enquiries and applications in each tract group with that group's share of owner-occupied homes or businesses. The report goes to our fair-lending officer, who co-signs each credit campaign's delivery plan before launch, and twice a year to the board risk committee.

This reporting is not expensive. The tract data is public, the delivery data belongs to the bank and the analysis takes an analyst a few days a month. What it requires is a decision that marketing's performance includes whom it reached, and that the fair-lending function, which has the expertise to judge the result, sees the numbers before a supervisor does. It also requires the agency contract to give the bank its delivery data at tract level, which ours did not until we renegotiated it.

Pay for reach the model will not choose

The second change costs money, and a board should know that before approving it. We removed ZIP-level bid optimisation from all credit advertising, stopped using the vendor's modelled audience and set delivery floors for each tract group in proportion to owner-occupied housing. We added Spanish-language advertising and paid for referral partnerships with three housing counselling agencies approved by the Department of Housing and Urban Development. Our head of mortgage lending objected that cost per application would rise on a flat budget. He was right. In the second half of 2024 it rose from US$290 to US$430, and our metro applications fell 7%.

By the end of 2025 the share of our metro applications from those tracts had reached 10.7%, and cost per application had settled at US$360 as our creative and our partners' referrals improved. We also added four loan officers in two branches, because the first months showed that advertising without anyone to answer the telephone merely moves the failure further down the process. Approval rates for applications from those tracts remain lower than elsewhere, mainly for credit reasons, and our credit and community-development teams are working on that with the counselling agencies. I do not regard the work as finished.

There are fair objections. Delivery floors override an optimisation that genuinely reduces cost, and a bank owes its shareholders care with their money. My answer is that cost per application is not the bank's real objective; lending profitably and lawfully across its whole market is, and a campaign that saves a third of its cost by ignoring a fifth of the market is not efficient. A second objection is that demographic reporting invites the bank to sort people by race. The reporting is by census tract, from public data, at the aggregate level that fair-lending analysis already uses, and no individual is ever classified. A third is that the gap may simply reflect where demand lies. Perhaps in part, but a bank cannot know what demand exists in neighbourhoods its advertising never reached.

The questions for a board are concrete. Can the chief marketing officer show where the bank's credit advertising was delivered last quarter, by tract group? Does the fair-lending officer see that before launch, or only after a disclosure-data comparison a year later? And has the board been told what it costs to reach the whole market, so that a decision to pay for it, or not to, is made deliberately? A bank that cannot answer the first question has handed part of its lending decisions to software it has never examined.