The Financialization of NFTs
Non-fungible assets have resisted every attempt to make them behave as collateral. We argue the failure is not one of liquidity but of representation: the industry prices a distribution with a single number. This paper sets out a trait-conditioned pricing model that produces per-item intervals rather than point estimates, and shows how downstream primitives — lending, tranching, liquidation — change when they consume uncertainty explicitly.
The floor price is not a price
A floor price is the lowest outstanding ask on a thin, adversarially-observable book. It is trivially manipulable by anyone willing to buy the bottom of a collection, and it carries no information about the item actually posted as collateral. Every lending market that adopted it inherited both properties.
The deeper issue is representational. A collection is a distribution of values with structure — traits, provenance, position in the collection’s own history. Collapsing that to a scalar discards the structure, and the discarded structure is precisely what determines whether a given item can be sold inside a liquidation window.
Trait-conditioned estimation
We model item value conditional on observable traits and transaction history, producing a posterior rather than an estimate. The width of that posterior varies enormously within a single collection: heavily-traded common items are priced tightly, while one-of-a-kind items may have intervals spanning an order of magnitude.
This is not a defect of the model. It is a measurement of something real that floor-price systems were reporting as certainty.
Consuming uncertainty
Once a primitive receives an interval, its design changes. A lending market can size against the lower bound rather than applying a flat haircut to a point estimate, which means items are neither uniformly over-collateralised nor uniformly dangerous. Liquidation logic can condition on interval width, extending the window for items the model admits it cannot price quickly.
The practical effect is that far more of a collection becomes financeable, at terms that reflect what is genuinely known about each item.
Tranching as the natural structure
Uncertainty that cannot be eliminated can be allocated. We show that pooling items and tranching claims against the pool dominates per-item lending across most of the parameter space we tested, for the same reason it dominates in structured credit: it lets participants with different tolerance for valuation error hold different parts of the same risk.
The junior tranche is, in effect, a market-priced estimate of how wrong the pricing model is — which is a considerably more honest oracle than a floor.
What we got wrong
Our first implementation propagated the interval but let the liquidation engine act on the midpoint, which reintroduced exactly the failure mode we had set out to remove. It took a live cascade to surface it. Uncertainty has to be carried end-to-end or it is not being carried at all.