How we grade risk.
And how we know it's right.
Every market gets a letter grade from five measured axes. This is exactly how it's built — and the live, on-chain proof that a low grade actually predicts real losses, not just a hunch.
Five axes.
One letter.
Each market is scored 0–100 on five axes, weighted into a composite, then mapped to a letter. An axis we couldn't measure is excluded, never guessed — the grade only reflects real, on-chain signal.
The five axes & their weights
The letter
The weighted score maps to a grade — and three rules can only ever CAP it lower, never raise it.
Does it actually
predict losses?
The honest test of any risk grade: do the markets we graded low turn out to be the ones that took real losses? We measure it directly.
Across — markets that took realised bad-debt losses and — that didn't, the grade ranks the loser lower — of the time. A coin flip would be 0.50.
Which axes carry the signal?
Each axis's own discrimination against real liquidations (AUC) — the label the weights are tuned to. The axes that predict get the weight; the ones near 0.50 (a coin flip) don't.
Every liquidation,
on-chain.
The grades are checked against a live dataset of real liquidation events — decoded from each protocol's own logs and reconstructed per market. Nothing is simulated.
Any liquidation is an independent check the grade never sees as an input — the grade separates markets that were liquidated at — AUC.
Coverage by protocol (markets liquidated / tracked)
Tuned against reality,
not guessed.
The axis weights aren't a product hunch. They were fit against real liquidation outcomes with nested cross-validation — regularised toward the prior judgement, so the data nudges the weights rather than overfitting five numbers to a sparse label.
What changed
Prior judgement → the weights the data actually supports.
Fit on liquidations, then measured against realised bad debt — a different adverse outcome. Grade-vs-bad-debt AUC rose 0.74 → 0.91. Generalising to a label it never saw is the sign it's a real improvement, not curve-fitting.
The limits
we won't hide.
Good risk work shows its own edges. Here's what these numbers do not prove — because a grade you can't interrogate isn't worth trusting.
Everything on this page is computed live, in your browser, from the same public data and the same code our risk engine runs. If the numbers look thin, that's the honest state of the evidence — and it gets stronger every day the dataset grows.