First payment default, and the agent that watches it

Of all the numbers a lender reports, first payment default is the one that arrives soonest and hurts most. It is also the one most often looked at monthly, in a pack, three weeks after it could have been acted on.

This is the first in a series taking each KPI in a lending business, explaining what it means commercially, how to define it so it survives an argument, and what an autonomous monitor should do with it.

What it is, in money terms#

First payment default is the share of new loans where the borrower misses the very first scheduled instalment. FPD 5 counts a miss of five days or more; FPD 30 counts thirty.

Both matter, for different reasons.

FPD 5 is your speed. It is observable about five weeks after issue. Nothing else about a cohort is available that early. If underwriting loosened in March, FPD 5 tells you in April rather than in September.

FPD 30 is your money. A loan that misses the first instalment by thirty days rarely recovers to a normal repayment profile. In most consumer books, an FPD 30 account is several times more likely to charge off than the average account, and it will consume collections capacity all the way there.

The commercial consequence is not the loss on those loans by itself. It is that FPD is a leading indicator of the whole vintage. A cohort with elevated FPD 30 will show elevated defaults at month six, month nine and month twelve, and by the time you can see that, you have written five more cohorts on the same policy.

Everything else about a vintage is confirmation. FPD is the only part of it you can still do something about.

Define it before you measure it#

Four definitional choices determine whether your FPD number is comparable with anything, including your own history.

Number based or amount based. Counting loans and counting euros give different answers, and the gap between them is itself information. If amount-based FPD is materially higher than number-based, your larger loans are defaulting first, which is usually a limit-setting problem rather than a scoring one. Report both.

Which loans are in scope. New loans and repeat loans behave completely differently and should not be blended. Top-ups, drawdowns and refinances have no meaningful "first payment" in the same sense and should be excluded outright — including them is the single most common way an FPD number gets quietly diluted.

The observation window. FPD 5 is not visible until roughly thirty-five days after issue, and FPD 30 not until about sixty. If your dashboard shows the current month it is showing an incomplete cohort, and it will look artificially good. Suppress incomplete cohorts rather than displaying them; a blank is honest, a partial number is not.

The cohort anchor. Group by issue month, not by calendar month of the default event. Anchoring on the event date mixes cohorts together and destroys the only thing that makes the metric useful.

The dimensional model underneath#

For an agent to reason about FPD rather than just report it, the model has to carry the drivers alongside the outcome.

• fct_loan at one row per loan at issue, carrying issue date, amount, term, product, market, score at decision and the score band
• dim_decision_rule, versioned with valid-from dates, so "what was the cutoff in March" is answerable
• dim_channel, resolving UTM, affiliate and broker to one conformed channel
• fct_instalment at one row per scheduled instalment, so the first instalment is identified by schedule rather than inferred from payments
• A metric definition for fpd_5 and fpd_30 with the exclusions inside it

That last point is the whole game. If FPD lives as a metric with exclusions in the definition, an agent asks for it. If it lives as a query pattern, an agent reconstructs it, and it will include the top-ups.

What actually moves it#

In our experience there are five candidates, and they are usually tangled:

1. A score cutoff or rule change. The most common, and the easiest to confirm because it has a date.
2. Acquisition mix. A channel that scales fast brings a different applicant, and it will not announce itself.
3. Product or limit changes. A longer term or a bigger first loan changes the affordability picture.
4. Seasonality. Real, and smaller than people claim when they want an explanation.
5. An upstream data problem. An income field that started arriving null and defaulted to zero — or worse, to a plausible average — will change decisions silently.

The last one is the one people forget, and it is the one an agent with lineage can catch that a human reading a chart cannot.

The agent#

Here is what a monitor on this KPI should do. Note that nothing in it is generative except the final sentence.

SignalFPD 5 for an issue-month cohort sits outside its learned band — here, 3.1 points above the trailing eight-week baseline for March.

ReadsReads fct_loan at issue-date grain, joined to dim_decision_rule and dim_channel. Excludes top-ups and refinances by metric definition, not by filter.

Holds fireCohort is 41 days seasoned. FPD 5 is valid; FPD 30 is not yet, so the agent will not comment on it.

CheckDid acceptance rate move in the same window? Yes — up 0.9 points from 8 March.

CheckDid any decision rule change? Yes — new-client score cutoff moved 0.22 to 0.26 on 8 March.

CheckDid acquisition mix move independently? Yes — affiliate share up 9 points over the same period.

CheckAny upstream data anomaly? Income field null rate unchanged. Rules out the silent case.

VerdictAttributes roughly 2.1 points to the cutoff change and 1.0 to the mix shift. Both are still recoverable this month.

ActionBrief to the risk lead naming the cohort, the split and the 1,840 rows behind it. Opens a ticket proposing a partial revert for a human to accept or reject.

Three things make this different from an alert.

It waits for seasoning. The most common false positive in FPD monitoring is reading an incomplete cohort. The agent knows the observation window because the model declares it.

It decomposes before it speaks. "FPD is up" sends four people to build four queries. "FPD is up, two thirds of it is the cutoff you changed on the eighth" sends one person to one decision.

It refuses to over-reach. FPD 30 is not yet observable for this cohort, so the agent says nothing about it. An agent that speculates on unobservable data is worse than no agent, because it teaches people to discount everything it says.

What good looks like#

A lender with this KPI properly instrumented can answer, without preparing for it:

• What FPD 5 and FPD 30 are for every cohort, number and amount based, excluding top-ups
• Which cohorts are not yet observable, and when they will be
• What changed in policy on any given date, and what the policy was before
• How much of a movement is policy, how much is mix, and how much is unexplained
• Which of those explanations was offered at the time, and whether it turned out to be right

That last one is the difference between a team that learns and a team that has the same conversation every quarter.

Next in this series: the recovery funnel, and why self-cure is the number nobody wants to measure.