How the valuation model works

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.

The current chart is a normalized scenario based on fair value, floor, and sale-price dispersion. Its 0–100 lines are relative price-appeal indexes—not probabilities and not promised time-to-sale. Appraisr now stores timestamped listing and floor snapshots. Measured sale probabilities will remain unavailable until listing outcomes and durations form a large enough sample for chronological validation.

Every collection's reported accuracy comes from a chronological 80/20 hold-out. The model trains on the older 80% and scores the later 20%; each target sale is absent from training, comparables, prior-sale inputs, and the causal floor at that timestamp. Weekly charts and recent-sale tables are immutable rows from that same evaluation ledger—not predictions recreated later with the final model.

“Accuracy” is displayed as 100 minus absolute percentage error, floored at zero. We also report median and mean absolute percentage error, ±5/±10/±20% coverage, signed bias, and separate results for floor, middle-market, and grail sales. Positive bias means the model over-predicted. Thin segments remain visible even when aggregate results look strong.

Each registry entry declares its valuation archetype and high-value regime fields. Hard regimes keep unlike assets out of the comparable pool—for example Punk, Azuki, and Meebits Type; Opepen Set and edition structure; and Otherdeed land, Koda, and resource signals. Interaction-led policies cover major trait combinations such as BAYC Fur, Eyes, and Mouth.

The analytics table shows unique sold tokens observed in the stored sample. It does not infer total collection supply from transaction frequency.

Confidence is calculated per token from effective comparable count, best similarity, comparable age, regime support, recognized-trait coverage, disagreement between model components, prior-sale age, and rare-tier support. Weak evidence widens the range and can produce an explicit insufficient-evidence state. Grails use same-collection rare-tier sales first, then token history, active asks, and executable item or trait bids as bounds. Cross-collection grail multipliers are research context only and are never substituted as a token price.