Redemption Behaviour Is a Funding Strategy

Most of the analytics ideas I read in bank proposals are about selling. Better segments, better next product, better conversion on a digital application. Those ideas travel well because everybody in the room understands what winning looks like and roughly what it is worth.

The idea I want to write about does not travel well at all, which is why it has survived in my notes far longer than the document it came from.

It sits in an old proposal put to an Austrian retail and corporate bank, on a panel of opportunities addressed to the finance side of the house. Two ideas on that panel, both marked for discussion, neither carrying a figure. The first argued that reshaping lending toward asset-backed form would improve loss-given-default estimation and lower the capital the bank has to hold, an effect the deck called dramatic without ever saying how dramatic. The second is the one I want. Sharper insight into portfolio growth and redemption rates, it said, gives you "better funding strategies to lower cost and improve margins."

Read at speed that is a sentence about reporting. Read slowly, it relocates a modelling problem from one building to another, and the relocation is the whole point.

What the borrower decides, the treasury pays for

Every loan has a contractual maturity, and every loan has an actual life, which is whatever the borrower does. People overpay when a bonus arrives. They settle early when they sell the asset. They refinance when the rate differential gets wide enough to be worth the paperwork. Corporates repay a facility ahead of schedule because a disposal completed or because a cheaper line opened somewhere else.

The distance between those two numbers is not scatter around a schedule. For a large book it is the schedule.

Funding is priced on term. A bank that funds a book as though it will run to contractual maturity, on a book that in practice repays years earlier, has bought duration it did not need and paid a spread for the privilege. A bank that funds short against a book that turns out to run long carries roll risk into whatever conditions happen to be in the market when the refinance falls due. Neither error announces itself. Both of them settle quietly into the interest margin, where they are easy to attribute to rates, to competition, to anything except an assumption nobody revisited.

That is why the sentence on the finance panel matters more than its position on the page suggests. Repayment timing wears the clothing of a credit question. It involves borrowers, loans and behaviour, so it lives with people who think about default. But its consumer is the treasury, and the treasury is not usually in the room when the analytics roadmap gets drawn.

Contractual maturity is a document. Actual life is a behaviour, and the treasury funds the behaviour whether or not anybody has modelled it.

The claim with no number attached

The proposal asserts the link and stops there. Across four executive domains it names eight opportunities and attaches a figure to none of them. On the finance panel specifically, one idea gets the word dramatically and the other gets nothing at all. The document is no more precise about its own commercials. The effort and investment page is a set of placeholders that proposes a shared investment and then, in its own body text, asks what the ratio should be. No day rate, no effort estimate and no duration appears anywhere in it.

I am not being unkind about the drafting. A proposal written before any data has been seen cannot honestly quantify anything, and I would rather read an unsized claim than a fabricated one. But the consequence is worth naming, because it explains why this idea keeps failing to get built while weaker ones get funded.

Prioritisation meetings are a contest between claims that carry numbers. A conversion model arrives with a test design, a per-application uplift and an owner in the commercial function who wants it. A behavioural maturity model arrives with a treasury conversation and a sentence. It loses, every time.

Breaking that requires a first measurement rather than a first model, and the first measurement is embarrassingly simple: take the expected life the funding plan currently assumes for a book, take what the book actually did over the last several cycles, and put the two side by side. Nobody needs a data scientist for that. What they need is the data, and this is where most estates fail.

Two ideas

Named on the finance-domain opportunity panel

Zero figures

Attached to either of them

'Dramatically'

The only magnitude word in the passage, unquantified

Eight opportunities

Across four executive domains, none sized

Balances are kept. Events are not.

Ask for redemption data and what usually comes back is a monthly balance snapshot per account. It is complete, it reconciles to the ledger, and it is close to useless for this.

A behavioural maturity model needs the event: date, amount, and above all a reason code that separates a scheduled instalment from a voluntary overpayment, an overpayment from a full early settlement, and a settlement that refinanced into another of the bank's own products from one that refinanced away to a competitor. Those last two produce the same balance movement and carry opposite meanings. One is a customer you kept on different terms. The other is a funding assumption and a customer walking out together.

Then there is lineage. Restructures, product migrations, portfolio transfers and system consolidations all rewrite account identity, and a model that cannot follow one borrower across an identifier change will read a migration as a redemption. Feed enough of those into a prepayment curve and it will confidently tell the treasury that a book runs shorter than it does.

None of this is modelling work. It is definition, capture and lineage, the same unglamorous layer that decides whether any operational estate can carry a model at all. It is also the part that gets skipped, because a proposal describing it does not sound like an opportunity.

An average is not a behaviour

The second failure is subtler than missing data, and it survives even in banks that have the events.

Redemption is normally held as a constant per product, a single rate applied to a book and refreshed occasionally. But repayment behaviour is conditional. It responds to the gap between the rate a borrower holds and the rate available, to how long the loan has been running, to the channel it was sold through, to whether the asset behind it changes hands. Estimate one average over a stretch during which those conditions were stable, and the average will hold beautifully until the conditions move. Which is precisely the moment the funding plan needed it to be right.

A conditional model is not exotic. The reason it stays unbuilt is not technique, it is that the value only becomes visible when someone measures the assumption error, and measuring the assumption error is the work nobody has been asked to do.

Building it so a regulator, and a treasurer, can follow it

A model like this eventually shapes the funding plan and, through it, the price of credit. That puts it in a different governance class from a campaign model, and it should be built that way from the first week rather than retrofitted after someone asks.

In practice that means a held-out period rather than a held-out sample, because the failure mode is regime change and a random split hides it. It means an evaluation set the treasury agrees to before the model is built, expressed as expected life against realised life on named books. It means lineage carried with every feature, so the reason code that drives the curve can be traced back to the system that wrote it. And it means one named owner for the assumption, a person rather than a committee, who can be asked why it moved.

The AI Act taking shape in Europe will eventually sort systems by what they touch, and lending sits near the top of that sorting. A behavioural maturity model is not a creditworthiness assessment, but it lives close enough to one that the documentation burden is worth assuming voluntarily and early. Institutions that already keep this standard for capital models find the extra work is mostly paperwork. Institutions that built their analytics culture around marketing tests find it is a rebuild.

There is a useful role for the current generation of assistants here, and it is a narrow one. A retrieval-augmented copilot over the funding policy, the product terms and the model documentation can answer, in seconds, what expected life a given book is held at and which document set that number came from. What such a tool cannot do is manufacture the event history underneath it. The modelling has never been cheaper and the retrieval layer is close to commodity. The reason code on a repayment is still a decision somebody has to make in a source system, and it is still the thing standing between a bank and this whole line of work.

That ordering, definition first and model second, is how RealAI's Platform team sequences an engagement in a lending book, and it is why we ask which assumption you want to test before which use case you want to run.

Drawn from a proposal put to an Austrian retail and corporate bank for data analytics support. It was a pitch, not a delivery: it produced ideas and a recommended approach, and no figure is attached to any of them in the source. The claim that redemption insight improves funding strategy is the document's. The reading of it as a data readiness problem, and everything said here about what it would take to build, is ours.