On-chain credit score: underwriting without a credit bureau

There is no file to pull. The repayment signal has to be built from behaviour the chain already records.

The missing credit file

Traditional underwriting starts by requesting a record someone else has already assembled. On-chain, that record does not exist. What exists instead is a complete, public, tamper-evident history of what an address has actually done — which is both less and more than a credit file.

Less, because it says nothing about income, employment or the obligations a borrower carries off-chain. More, because it is a record of behaviour rather than a summary of it. A bureau file tells you a borrower was 60 days late in 2023. An on-chain history shows you the hour the position went underwater, whether they topped up collateral or walked away, and how long they took to decide.

What an on-chain credit score substitutes for it

The signal comes from behaviour over time rather than a declared position at a moment:

• history of borrowing and repayment across lending protocols, including partial and late repayment;
• behaviour under stress — what the address did during sharp drawdowns, when liquidation was close;
• stability of funding sources, and whether collateral is owned or rented;
• the age and consistency of activity, which is the closest available proxy for intent to stay.

An address with two years of small, boring, fully repaid loans is a better risk than one with a large balance and no history, and the model should say so.

Behaviour under stress is the strongest feature

If we could keep only one input it would be this. Anyone repays comfortably in a rising market; the informative observation is what an address did at 3am during a 30% drawdown with a liquidation price three percent away. Topping up collateral, partially deleveraging, or doing nothing and accepting liquidation are three very different signals about the same borrower — and all three are recorded permanently, with timestamps, for anyone to read.

Rented collateral

Whether collateral is owned or borrowed is invisible in a snapshot and obvious in a history. An address that acquired its collateral through a flash loan, a looped position, or a transfer from a related address minutes before drawing credit is a materially different risk from one that has held the same assets for a year. Provenance analysis here uses the same graph machinery as risk scoring, applied to a different question.

What it cannot do

It cannot see income, it cannot see identity, and it cannot stop the same person opening a fresh address with no history. Underwriting on-chain is therefore about pricing the absence of evidence honestly rather than pretending to certainty: a thin address is not scored as bad, it is scored as unknown, with a wide band.

That distinction has practical consequences a lender must design around. Because a borrower can always create a fresh address, on-chain reputation is an asset that accrues and can be abandoned — which means it deters bad behaviour only in proportion to how much history has been accumulated. The implication is not that the signal is useless; it is that a large first loan to a thin address is not underwritten by the score, and no amount of modelling will change that.

How lenders should use an on-chain credit score

Signal shape	Reasonable use
Long history, consistent repayment	Improved terms or a higher limit against the lender's own policy
Long history, stress-period defaults	Price the risk; do not treat one liquidation as disqualifying
Thin history, clean	Small limit, tighter re-assessment interval; the score is not the underwriter here
Rented or freshly acquired collateral	Discount the collateral, not the borrower

Where it is used

The output is an input to a lender's own policy — a repayment signal with its confidence attached, not an approval decision. The lender sets thresholds, and keeps a human path for anything consequential.

That is not only good practice, it is frequently a legal requirement. Credit decisions made about individuals by automated means attract explanation and human-review duties in most of the jurisdictions our customers operate in, and a lender who cannot articulate why an application was declined has a problem no vendor can solve for them. We publish contributing factors alongside every signal for exactly that reason.