The KPI framework we build for a consumer lender in the first 30 days
The metric set we put in first for a consumer lender, in the order we build it — and the four definitions that cause the most arguments.
Every consumer lender we meet has a dashboard. Most of them report the same three things: applications, approvals and arrears. Those are the numbers that are easy to get out of the loan management system, which is a different thing from the numbers that run the business.
The gap shows up in a specific way. Someone asks why last month was worse, and four people go away and come back with four answers. Nobody is lying. They are each looking at a true number that answers a slightly different question.
What follows is the metric set we put in during the first month of a lending engagement, in the order we build it. It is not exhaustive and it is not meant to be. It is the smallest set that lets a leadership team argue productively.
Start with the funnel, not the portfolio#
The instinct is to start with the loan book, because that is where the money is. Start with the funnel instead. Portfolio problems are usually funnel problems that happened four months ago, and you cannot fix a vintage after you have written it.
The funnel we build has six stages, and every stage is a count and a conversion rate off the previous stage:
- Traffic to the application form, split by channel and by device
- Applications started
- Applications completed
- Decisioned, split into approved, declined and referred
- Offered and accepted
- Disbursed
Two things matter more than the stages themselves.
Each stage needs one owner. If marketing owns traffic and risk owns decisioning, then the drop between completed and decisioned belongs to someone specific. Metrics without an owner become weather — everyone comments, nobody acts.
Declines need a reason code taxonomy that a human wrote. Most loan management systems emit decline reasons designed for an audit trail rather than for analysis. You end up with forty codes where six would do, three of which mean "policy" in different words. Group them into a small set the business actually reasons about: affordability, credit bureau, identity and fraud, policy exclusion, data quality, and duplicate or abandoned. When your top decline reason for a market turns out to be data quality, that is a product bug, not a risk outcome — and you want to be able to see that in one chart.
Then unit economics, before the loan book#
The second thing we build is contribution per disbursed loan, by cohort and by channel. Not revenue. Contribution.
The minimum version needs four components:
| Component | Where it usually lives | The common mistake |
|---|---|---|
| Interest and fee income | Loan management system | Recognising it at origination rather than as earned |
| Expected credit loss | Risk models or IFRS 9 stage allocation | Using actuals only, which lags by months |
| Acquisition cost | Ad platforms and affiliate invoices | Allocating spend to the month it was paid, not the cohort it bought |
| Servicing and collections cost | Finance, usually an allocation | Leaving it out entirely because it is hard |
The third row is the one that gets skipped and the one that changes decisions. If you attribute August's affiliate invoice to August's disbursements rather than to the cohort that invoice actually acquired, every channel looks fine in a growing month and terrible in a flat one. That is an artefact of your reporting, not a fact about the channel.
Once contribution per loan exists by channel, a lot of arguments end. The channel that looks expensive on cost per acquisition is often the cheapest on contribution, because it brings customers who repay.
Then the book, in vintages#
Only now do we build portfolio reporting, and we build it in vintages rather than as a single balance.
A vintage view groups loans by the month they were disbursed and tracks each group's performance over its own lifetime. It is the only view that lets you compare December's underwriting with June's without growth distorting the answer. A book-level arrears percentage falls when you write more loans, which is the exact moment you should be most suspicious.
The standard set:
- Vintage curves — cumulative percentage of each cohort that has reached 30, 60 and 90 days past due, plotted against months on book
- Roll rates — the probability of moving from each delinquency bucket to the next, month on month
- Recovery curves — cumulative cash recovered as a percentage of the balance at charge-off, by months since charge-off
- Early payment default — the share of a cohort that misses its first or second instalment
Early payment default is the one we build first if we have to pick one. It is the fastest signal you have that something changed in underwriting or in acquisition, and it is available within sixty days rather than after a year of curve.
The definitions that cause fights#
Roughly the same four definitions cause arguments at every lender we have worked with. Write them down, get them signed, and put the signed version where the dashboard is.
What counts as an application. Does a form abandoned at step two count? Does the same person applying twice in a week count as one or two? Marketing wants the higher number, risk wants the lower one, and both are defensible. Pick one, name the other something else, and report both if you need to.
When a loan is "active". At disbursement, at first payment, or at contract signature? The gap between these is small in days and large in reporting, because it moves loans across a month end.
What "default" means. IFRS 9 stage 3, ninety days past due, sent to collections, and legally in default are four different populations. Your regulator, your funder and your board may each mean a different one.
Which day the month ends. Calendar month end and processing month end are not the same date in most loan management systems. The difference is usually one working day and it will produce a permanent, unexplainable gap between finance's numbers and yours.
The point of writing definitions down is not that the definitions are hard. It is that when a number moves, you want the conversation to be about the business, not about whose query was right.
What we do not build in the first 30 days#
Some things are worth naming as deliberately excluded, because a lot of first-month analytics work is spent on them.
We do not build self-serve until the semantic model is stable. Handing out a broken model faster does not help anybody.
We do not build channel attribution beyond last non-direct click. Multi-touch attribution is a genuinely interesting problem and a genuinely poor use of your first month. Get contribution per cohort right first — it answers most of the same questions with far less machinery.
We do not build predictive scorecards. Not because they are not valuable, but because a scorecard built on a data model you are about to change is a scorecard you will rebuild.
The order matters more than the list#
If there is one thing to take from this, it is the sequencing. Funnel, then unit economics, then portfolio. Most lenders build it backwards — portfolio first, because that is where the auditor is looking — and end up with excellent reporting on outcomes they can no longer influence.
The funnel is where you still have choices.
Want this framework built against your own book?
The Insight Pod does exactly this — a KPI framework signed off by your leadership team and a governed model underneath it, with the first executive dashboard live in 21 days.