Marketing budgets for banks and lenders, measured on the loans that pay out.

The sections below cover which outcome to model, how the delay is handled, what to do when the application flow can't report back, which controls are worth having in a bank model, and how to set the budget from what a loan is worth.

A loan passes several steps before it earns anything: application, credit check, offer, acceptance, payout. Each step loses people, and not evenly across channels. A channel that brings many cheap applications can bring applicants who fail the credit check. Measured on applications, it looks like the best channel you have.

We usually model the step closest to revenue that still has a steady weekly volume, and the amount where loan size varies a lot. Where it's unclear, both candidates go in and the first model is built on both, so the choice is made on what each shows.

A lender's sale happens weeks after the marketing. The model is set up with that delay, the number of days between the activity and the outcome it is credited with, so cause and effect line up. When a later step closes, the model credits the driver that started it: if a search campaign brought the application that paid out a month later, the search campaign gets the credit for the payout.

The same setup handles leads that close long after the first contact, for example through a follow-up programme. Once the model is retrained on the new data, those late sales are credited back to whatever brought the lead in, marketing or base.

Many banks run their loan application on a third-party system, often on a domain that looks like the bank's own. The result: approved loans can't be sent back to Google or Meta as conversions, the platforms optimize on a weak signal, and nobody can see which channels bring loans that pay out.

A marketing mix model does not need those conversions. It uses what was spent per channel over time, and an aggregated volume the bank already has: applications, approved loans, paid-out loans or loan amount. From the relationship between the two it estimates what each channel adds. Where platform data does exist, it helps set the starting assumptions for each channel. Whether conversions should also flow back to the platforms for bidding is a separate question, usually one for the media agency.

Interest rates move lending volume, so a bank model needs to know about them. But every extra variable takes explanatory power away from the media, and anything used to explain the past also has to be known when the model forecasts. That rules out a lot of tempting data.

• Start with the policy rate. The dates of central-bank rate decisions, and your own best rate set against the policy rate, are known ahead and have a direct effect.
• Mark process changes. A new application flow or a change in credit policy can shift volume overnight. Flag the date, or the model will credit the change to whichever channel ran that week.
• Treat rankings with care. Your position on rate comparison sites can explain history well, but you can't know it for next quarter, so it weakens forecasts. It is a nice-to-have once the basic model is in place.

Start simple and add controls as the months go by and the model shows where it needs them.

Young banks often have a launch period that looks nothing like normal operation: a rate cut or aggressive pricing brings in more applications than the bank can handle, and marketing is switched off for a while. Should the model learn from those months?

We try both: one model on the whole history, one starting where the numbers settled, and keep the version that fits better. Throwing data away is a cost, and so is teaching the model from a period that will not repeat.

A bank has far fewer conversions than a webshop, and each one is worth much more. With less data, the model leans more on what is already known about each channel. In a Bayesian model that knowledge goes in as priors: starting assumptions about how each channel works, drawn from your own past budgets, platform data and interviews with your team. The data then moves them.

That is what keeps a bank model stable on a few hundred loans a week rather than thousands of orders. Read why the Bayesian approach fits sparse data.

The budget question in lending has the same shape as in subscription businesses. A paid-out loan brings a margin over its expected life. That margin, over the horizon you are willing to wait for payback, is the most the next loan may cost to win. The model estimates what the next loan costs in each channel at today's spend, and the budget level is where that marginal cost meets your ceiling.

It changes the budget conversation with finance. Instead of defending last year's number, marketing can show what a larger budget would bring at the margin, for example that a 20% larger budget still pays back 1.75 times on the additional spend, or that it does not.

Banks spend more on brand than most performance-driven businesses, because trust takes time to build. TV's near-term pull, the extra searches and applications in the weeks after a burst, shows up in the model as TV's own effect. For the slower effect, the model can be extended with a brand component that captures sales arriving months later. See how we model brand effects.

Where the data is thin for a channel, a structured test settles it faster: change the spend in one channel at a time, pulse it on and off over a few weeks, and let the model measure the result. Media platforms often co-fund a test when they know a marketing mix model is measuring it.

Odins is a managed marketing mix modeling platform. We connect the ad platforms and your loan or account data, build and maintain the model with your team's knowledge as priors, and review every recommendation before it reaches you. Data flows in continuously and the model is retrained once a month, while scenarios and forecasts are yours to run whenever you need them, in the platform or from your own AI assistant. Companies like Aprila Bank and Penni run their marketing budgets on Odins. Read the full guide to MMM for banks and long sales cycles.