Methodology

How the valuation model works

One transparent model, fit live to each collection's real OpenSea sales. No black box, no paywall.

In plain terms: an NFT is worth what similar ones have recently sold for. We look at thousands of real sales, work out what each trait adds to the price, and use that to estimate any item's value. Here is exactly how the model does it, step by step.

Hedonic regression on log-price

A ridge regression learns a price for every trait_type=value feature (plus a trait-count signal) directly from recent sales. Working in log-price keeps multiplicative trait effects linear and well-behaved.

Monthly time fixed-effects

A per-month market-context index lets many months of sales contribute without the price trend leaking into trait estimates. This is the single biggest accuracy lever. It de-trends every comparable to “today.”

Comparable-sales ensemble

A trait-similarity (Jaccard) kNN, recency-weighted, finds the closest real sales and blends with the hedonic estimate. This captures nonlinear trait interactions a linear model alone would miss.

Robust trimming + bias correction

Wash trades are removed with a log-space MAD filter (per type bucket, so rare grails are never trimmed as “outliers”), then a median bias correction centers the estimate.

Liquidity curves

The price/speed tradeoff on every item

The model measures each collection's clearing spread, the dispersion of sale prices around fair value. From that we derive two S-curves: the probability a listing sells at each ask, and the probability an owner accepts a bid at each price. They cross near fair value. List below fair value and your sale-odds rise; hold out above it and you wait. It's directional guidance grounded in real dispersion, not a guarantee.

How accuracy is measured

Honest, hold-out, per collection

Every collection's accuracy is computed on a seeded 80/20 hold-out: the model trains on 80% of recent sales and is scored on the 20% it never saw, valuing each test sale as of its sale date(no look-ahead). We report median accuracy, the share of estimates within ±5/±10/±20%, and bias. You can see it live on each collection's accuracy tab.

Accuracy varies by collection. Liquid, floor-dominated collections are easiest to value precisely; thin markets and rare-trait grails carry irreducible noise. The same token can sell at different prices depending on buyer, seller, and timing. We don't hide that.

Type-aware collections

When a trait stratifies value into tiers

Some collections have a trait that splits the market into value tiers (you should never compare a rare type to a common one). For those, comps and trimming are bucketed by that trait:

🌸 Azuki · Type 🧍 Meebits · Type 🧢 Nakamigos · Type 👾 CryptoPunks · type 🕹️ CryptoPunks V1 · Body 🧬 CLONE X · DNA 😺 Cool Cats · tier ⬛ Opepen Edition · Edition Size

Every other collection is valued across its whole set via trait similarity.

Data via the OpenSea API. Estimates are probabilistic and not financial advice.