From Transactions to an Affordability View: What Is Safe to Infer

The distinction that matters is between reading and judging. Reading a statement is a solved engineering problem with a checkable answer — the balance has to reconcile. Deciding that a particular credit is a salary is a judgement, and no arithmetic verifies it. Systems that present both with the same confidence are the ones that get credit teams into trouble.

The ladder, safest first

Note how the uncertainty compounds downward. Each rung inherits the errors of the rungs above it, so the figures a credit decision most wants — total obligations, disposable income — are the least reliable things the system produces. That is worth surfacing in the interface rather than hiding behind a single number.

Recurrence is detectable; meaning is not

Finding that a credit of similar size arrives at a similar point each month is a tractable pattern problem, and it is robust because it does not depend on reading the description at all. Tolerance matters — pay dates move for weekends and holidays, and amounts vary with overtime or allowances — so a detector requiring exact equality will miss real salaries while one that is too loose will merge unrelated credits.

What recurrence cannot tell you is what the money is. A monthly credit of a consistent amount might be salary, a transfer from a family member, rental income, or a loan drawdown cycling through the account. Narration helps and is unreliable: it is truncated, abbreviated differently by every bank, frequently bilingual, and sometimes just a reference number. An honest system reports "recurring monthly credit, likely salary" with the evidence attached, and lets an analyst confirm. One that reports "monthly income: 850" has made a judgement and hidden it.

The asymmetry that should shape the design

Missing income and missing an obligation are not equally bad. Understate income and you decline someone you should have approved — a lost sale, visible in conversion. Miss an obligation and you approve someone who cannot afford the repayment — a credit loss, visible much later and much more expensively.

So obligation detection should be tuned to over-flag and income detection to under-claim, with both routed to a human. A system tuned symmetrically is optimising a statistical measure rather than a commercial one, and the two point in different directions.

Risk markers, and one caution

Some signals are close to arithmetic and worth surfacing plainly: returned cheques, insufficient-funds charges, balances trending toward zero before each pay date, and a rising ratio of fixed outgoings to inflows. These are countable rather than interpreted, which puts them near the top of the ladder.

The caution is that a marker is not a conclusion. One returned cheque in a year is noise; three in a quarter alongside a falling minimum balance is a pattern. And a marker's absence proves little, because the statement supplied may not be the applicant's only account — which is the limit of the whole exercise and worth stating to anyone who asks the system for a verdict.