Marketing mix modelling for banks, insurers, fintech and B2B SaaS: long sales cycles and adviser-led sales

Banks, insurers, fintech apps and B2B software companies share a doubt about marketing mix modelling: our sale does not happen the day someone clicks. An application takes weeks to become a paid-out loan. A policy is often sold by an adviser. A SaaS deal closes a quarter after the first visit. This guide explains how a marketing mix model handles that, which outcome to model, and where the method stops working.

The short answer

Why these businesses doubt MMM

The doubt usually comes in one of three forms.

• The conversion is weeks away from the click. A mortgage application is submitted, assessed, granted, accepted and paid out. A B2B trial becomes a paying customer after a sales call. Platform attribution windows of 7 or 28 days miss most of it.
• The sale goes through a person. Insurance and corporate banking are often sold by advisers. A marketing team can see the application, but the sale depends on what happens in the branch or on the phone.
• The tracking breaks. Many banks run their application flow on a third-party system that looks like their own domain but cannot pass approved loans back to Google or Meta as conversions. They match in the CRM afterwards, with a gap between the start of the application and the booked volume.

All three are real problems for attribution, which follows tracked users through a funnel. None of them is a problem for a marketing mix model, which never follows a user at all.

The model does not need platform conversions

A marketing mix model is a statistical model of the relationship between two time series: how much you spent per channel each week, and how much business you did each week. Both are aggregates. The spend side comes from the ad platforms, your TV and radio bookings, sponsorships and so on. The outcome side comes from your own systems: the number of applications, the volume of approved loans, the number of funded accounts.

Whether or not your application flow can report back to Google is irrelevant to the model. It would use the same two series either way. Whether you should send conversions to Google so that its bidding improves is a separate question, and it belongs to whoever runs your campaigns.

Pick the outcome that matters

The most important setup decision is which outcome the model learns from. Our rule is to model the business outcome, not a top-of-funnel count, as long as it has enough volume. For a lender that means approved loans or paid-out volume, not form starts. For a subscription or SaaS business it means signed customers or revenue, not trials. The model can be built against most things, including loan size.

There is a case for an earlier step too. Applications and bookings are closer to marketing, so no other part of the organisation influences them: sales are affected by adviser incentives and the offers of the month, and in some markets the link from the final sale back to the original lead is lost in the CRM, while the earlier step never loses tracking. When that is the situation, we take both series in, build the first model on both, and agree with the customer which to carry forward. Some customers run several models permanently.

How the delay is handled

The gap between a marketing touch and the outcome is part of the model setup. We set the expected number of days between spend in a channel and the outcome it should be credited with, so the model lines up cause and effect. That setting carries through to the recommendations and the scenario planner. A multi-step funnel, say submit, grant, accept and pay out, is modelled on the step you chose as the outcome, with the whole delay in front of it.

The model also catches up on late conversions. One of our customers introduced a retention programme that turned old leads into sales months later. When the model is retrained, it identifies that the lead converted and gives the credit to whichever component it believes produced the lead. If a search campaign produced the lead, search gets the sale.

Interest rates, advisers and other things that are not marketing

In financial services a lot of demand has nothing to do with advertising. Rates move, competitors change their offers, a regulator changes a rule. We handle this with fewer control variables than classic sales modelling, on purpose: every variable you add takes explanatory power away from the media channels, and every variable you add has to be forecast too. So we take in the obvious factors with a direct effect, and let the rest sit in the baseline as a trend.

For a bank, the obvious factors are the dates of policy-rate decisions and the bank's own best mortgage rate against the policy rate. A rank on a rate-comparison site is nice to have, because it explains history, but it is hard to know a year ahead, and a control variable has to be knowable ahead to be useful in a forecast. The advice we give is to start simple and add better controls as the months go by.

Adviser-led sales are treated the same way. The model does not try to explain what happens in the branch; it measures how much the volume moved when the marketing moved, with everything else held in the baseline. The baseline is usually the largest part of the result in these sectors, and knowing its size is itself a useful answer: it tells the board how much of the business marketing is actually moving.

Where the method stops working

Two conditions have to hold. First, the outcome has to be in your data. A step that completes offline and is never recorded, or is recorded without a date, cannot be modelled. Second, there has to be enough of it. A sales signal the model can learn from means, as a rough rule, around a hundred conversions a week or more. Below that, the weekly series is too noisy for any model to separate marketing from everything else.

That second condition is what makes pure B2B a weak fit. Account-based selling with a handful of large deals a quarter gives the model almost nothing to learn from, however large the marketing budget. B2B software with high-volume self-serve sign-ups is a different case and works like any subscription business. If you are unsure which side of that line you are on, the volume of your weekly outcome is the thing to look at.

What it looks like in practice

Aprila Bank, a Norwegian bank, runs its marketing budget on Odins. Between October 2024 and January 2025 the bank got 37% more loan applications from the same budget by reallocating spend on the model's recommendations. That is one customer's result, not a forecast for anyone else, but it shows what the method does in a business where the loan is approved and paid out well after the first visit. Read the Aprila Bank case.

Odins is a managed marketing investment platform. We connect the data, digital and offline, build and maintain the model, and deliver monthly recommendations that our team reviews before they reach you. It fits companies with a marketing budget from EUR 1 million to several hundred million a year and a sales signal the model can learn from.

Frequently asked questions

Our loan application runs on a third-party system, so we cannot send approved loans to Google as conversions. Can you still measure marketing?

Yes. The model does not use conversions sent back to the platforms. It uses what you spent per channel over time and an aggregated outcome you already have, such as approved loans per week, and models the relationship between the two.

Does marketing mix modelling work for B2B SaaS with long sales cycles?

It works when the outcome exists in your data with enough weekly volume, which is typically the case for self-serve and mid-market SaaS with many sign-ups. Account-based selling with a few large deals a quarter is a weak fit, because the outcome is too slow and sparse to learn from.

Should we model applications or paid-out loans?

With enough volume, paid-out loans or approved volume, because that is the business outcome. Where the final step is influenced by advisers and offers, or loses its link to the lead in the CRM, we take both series in, build the first model on both, and agree which to carry forward.

How do you handle the delay between a click and a funded account?

We set the expected number of days between spend in a channel and the outcome it should be credited with. The setting is part of the model configuration, and it carries through to the recommendations and the scenario planner.

How do you account for interest rates and competitor offers?

With a small number of control variables that have a direct effect and can be known ahead, such as policy-rate decision dates and your own rate against the market. Everything else stays in the baseline as a trend, so the media channels keep their explanatory power.

How do we measure which marketing drives new active users in a fintech app?

Model funded or active accounts per week as the outcome, not installs, and set the delay between acquisition spend and activation. The model then shows which channels moved activations, and how much of the growth came from marketing versus the baseline.

For why the platforms' own reports cannot be added up across channels, read why Google Analytics and Meta report different numbers. For the general requirements, read what data a marketing mix model needs. For how the model separates marketing from the baseline, read how the model works. To see the setup on your own funnel, book a demo.