What a risk score is allowed to say
A single number for a whole file is either a convenience or a decision wearing the clothes of an observation. Ours is meant to be the first: a weighted summary of six things the statement showed, published with its recipe, because a score you cannot take apart is a score nobody should be acting on.
- Six factors, published with their weights
- An unobservable factor is never scored as clean
- Three floors that override the average
Six factors and what each is measuring
Weights in brackets. Every factor is shown beside the score in the report, and every factor traces back to transaction rows.
Financing burden (25%) and cash flow (25%)
Burden scales total position payments against revenue, with the top of the scale set well above where anyone would fund. Cash flow scores how far net flow has gone negative relative to revenue. Together they are half the score, which is the right proportion: they are the two things the advance will actually be repaid out of.
Stack depth and new positions (20%)
This takes the worse of two readings: how many positions began during the period, and how deep the active stack is. The existence of a stack is risk on its own and it deepens with each funder — two active advances score a third of the way up, three two thirds, four or more at the top — independent of how many of them are new this month.
Liquidity (15%)
The share of observed days that closed below the low-balance threshold. Crucially the denominator is the days the ledger actually observed rather than the length of the period: forty low days out of forty-five observed is an account that is illiquid about ninety percent of the time, not forty-four.
Returned items (10%) and document authenticity (5%)
Each return moves the factor a quarter of the way up, so a handful saturates it. Authenticity carries the lightest weight in the model on purpose: a flagged document is a reason to look at the document, and it should not be able to decide an underwriting outcome from inside an average.
Unobservable is not the same as clean
Everything above is ordinary. The part worth arguing about is what happens when a factor cannot be measured at all, because the obvious implementation is also the dangerous one: score it zero and carry on. Zero, in a risk model, means perfect.
Two cases hit this regularly. When the daily balance ledger cannot be vouched for, liquidity is not good and not bad; it is unknown. And when a statement carries financing positions but no observable revenue base to measure them against — income arriving entirely as transfers, say — both burden and cash flow are unknown rather than excellent.
Scored as zeros, those two cases would hand a stacked merchant a comfortable number built out of things nobody could see. So an unobservable factor is not scored: its weight is removed and the remaining weights are renormalised, and the score is an average over what the statement actually showed. The report says which factors were observable, so a score built on four of six is never mistaken for a score built on six.
Three situations where an average is the wrong summary
A weighted mean is a reasonable way to combine independent pressures, and there are states where it quietly lies, because a genuinely alarming fact gets diluted by the good behaviour around it. Three of them override the average outright.
An account that lives under the line. A business that spends most of the period below the low-balance threshold while burning cash is not low risk merely because it carries no debt. With liquidity at fifteen percent of the model, a debt-free and empty account would otherwise land in the green, so the score is floored.
Positions with nothing to measure them against. Active financing and no observable revenue means the burden could be anything at all. That uncertainty is itself elevated risk, and it is floored rather than renormalised into comfort.
Payments to a debt-settlement arrangement. A merchant paying a settlement firm or its dedicated-account processor is actively working out existing debt. That single fact dominates whatever the rest of the statement looks like, and it pins the score deep into the high band.
Three things the number is not
It is not a probability of default. We have not built a model that estimates one, and a score that looks like a percentage is not the same thing as a percentage that means something. Anyone publishing a default probability should be asked what it was fitted on and over what period.
It is not a decision. Nothing in the report approves or declines. The score orders a queue and points at what to read first; the deal has to survive a human reading the evidence.
It is not comparable with somebody else’s score. Another tool’s 60 and this one’s 60 are different statements about different mixtures. Comparing them means comparing recipes, which is exactly why this one is published rather than described as proprietary.
Common questions
What goes into the risk score?
Six factors with fixed weights: financing burden and cash flow at a quarter each, stack depth and new positions at a fifth, liquidity at fifteen percent, returned items at a tenth, and document authenticity at five percent. Each is shown alongside the score.
Is the score a probability of default?
No. It is a weighted summary of six observations from the statement, not an estimate fitted to outcomes. Treating it as a default rate would give it a precision it does not have.
Why is a merchant with no debt not scored as low risk?
Because an account that spends most of the period below the low-balance threshold while burning cash is not safe just because nothing is being debited. That state floors the score rather than being averaged away by the absence of positions.
What happens when a factor cannot be measured?
Its weight is removed and the remaining weights are renormalised, rather than the factor scoring zero. Zero in a risk model reads as perfect, and an unknown must never be published as a clean result.
Does a document authenticity flag dominate the score?
No, and deliberately so — it carries the lightest weight in the model. A flagged document is a reason to examine the document, which an underwriter should do directly rather than through a number that already averaged it away.
Related
See the score and the six factors behind it
Upload a statement and read each factor next to the rows it came from.