Why COMPED prices are more accurate — and getting better weekly
Accuracy in card pricing is not a claim, it is a method. Here is exactly what COMPED measures, which sets other tools get badly wrong, and why an early-access pricing engine that learns from humans gets more correct every week rather than drifting.
The five things that make a comp accurate
- 01
Completed sales, not asks
Every comp starts from sales that actually happened. Live listings are used only to sanity-check the number, never to set it, so you are not pricing against optimistic asking prices.
- 02
One printing, one price
1st Edition, Shadowless, 4th print, Unlimited, reverse holo, promo stamp and every regional release are priced as separate cards. Averaging printings together is the single biggest cause of wrong prices elsewhere.
- 03
Condition anchored to reality
Vintage raw cards are anchored to the played grades they actually sell in; modern cards anchor to Near Mint because that is what the market trades. Pricing a 1999 holo as NM inflates it substantially.
- 04
AI outlier scrubbing
Digby, our proprietary engine, reads titles and conditions like a vendor and drops lots, bundles, sealed product, proxies, damaged copies and mislabelled grades before the weighted comp is calculated.
- 05
Stated uncertainty
Sale counts, ranges and an explicit 'not comped yet' rather than a fabricated figure. A price you can audit beats a price you have to trust.
The sets other apps get badly wrong
Coverage gaps cluster in predictable places: eras with unusual card types, regions priced from English data, and printings that share artwork. These are the areas COMPED was built around.
| Set / era | What goes wrong elsewhere | How COMPED prices it |
|---|---|---|
| e-Card era: Expedition, Aquapolis, Skyridge | Crystal cards such as the Aquapolis Crystal Lugia get averaged with the ordinary holo, or carry a stale US figure years out of date | Crystal, holo and reverse printings priced separately from their own completed UK sales |
| Neo era: Genesis, Discovery, Revelation, Destiny | 1st Edition and Unlimited merged; shadowless-adjacent print differences ignored | Edition-split comps with the sale count shown for each |
| Japanese vintage old-back (1996–2001) | Often not searchable at all, or priced from the English equivalent | Priced from Japanese sold data, searchable by English species name, 'old back', 'kisei' and set shorthand |
| Vending Series and Carddass | Uncut sheets and lots contaminate single-card prices, inflating them wildly | Sheets, lots and bundles are rejected; vending singles priced against vending singles |
| Southern Island, promos and jumbo prints | Sealed sets, singles and oversized prints averaged into one number | Each product type priced from its own sold evidence |
| Graded slabs across PSA, CGC, ACE, TAG, SGC | A raw price multiplied by an assumed multiplier | Grade- and grader-specific comps from verified sales, auctions included for trophy cards |
Why early access means more accurate, not less
Most pricing tools are static: they scrape, average, and publish. COMPED is a learning system, so the direction of travel is toward the market rather than away from it.
- Price Trainer: collectors and vendors verify or correct comps against the sold evidence in front of them, and verified verdicts outrank scraped data.
- AI price training: every verdict, rejection and correction teaches the engine which listings were never comparable, so the same mistake is not repeated across thousands of cards.
- Continuous warming: the most-searched printings are re-priced continuously, and coverage widens set by set into the long tail.
- Market tracking: completed sales, live asks and reference prices are cross-referenced so a divergence is investigated rather than published blindly.
- Community verification: where raw data is genuinely too thin, human evidence fills the gap instead of a guess.
Practically, that means a card showing a wide range today can be a tight, well-evidenced comp in a few weeks — and a card that was wrong stays fixed, because the correction lives in the engine rather than in one person's notes.
How COMPED builds a number
- UK sold data first. Completed eBay UK sales in GBP are the primary evidence. If a card is thinly traded, the search widens to global sold data rather than falling back to a guess.
- Exact printing only. Set, number, language, finish and edition must match — 1st Edition, Shadowless, 4th print and Unlimited are priced as separate cards, as are reverse holos and regional printings.
- Outliers scrubbed. Digby drops damaged copies, lots, proxies, mislabelled graded sales and single freak results before the weighted comp is calculated.
- Condition-anchored. Vintage raw cards are anchored to realistic played grades rather than pretending every copy is Near Mint; modern cards anchor to NM.
- Community-verified. Vendors and collectors confirm prices in the Price Trainer, and that consensus nudges the headline comp on cards where sales are sparse.
- Uncertainty is shown, not hidden. When there is not enough comparable evidence, COMPED says so and shows a range with the sale count behind it.
Keep reading
Common questions
- Why is COMPED more accurate than other Pokémon pricing apps?
- Because of what it measures and what it refuses to measure. Comps start from completed sales of the exact printing in GBP, live listings are only a sanity check, and our proprietary AI removes lots, bundles, sealed product, damaged copies and mislabelled grades before the weighted figure is calculated. Every number carries its sale count so you can judge it.
- Which cards are other apps most wrong about?
- Mid-era e-Card cards such as the Crystal Lugia from Aquapolis, vintage Japanese old-back printings, Vending and Carddass singles, Southern Island, and any card where editions and finishes get averaged into one price. These are exactly the sets COMPED prices as separate markets.
- How does price training make the app more accurate over time?
- Human comping judgement is fed back into the engine through the Price Trainer, where collectors and vendors verify or correct comps against the sold evidence shown to them. Verified verdicts outrank raw scraped data, so every session narrows the gap between the model and the market.
- What does the AI actually do?
- Digby, our proprietary engine, reads listing titles and descriptions the way an experienced vendor does: identifying printing, edition, language, finish, grade and condition, then rejecting sales that were never comparable in the first place. It also weights recency and corroboration so one freak sale cannot become the price.
- Is COMPED still improving?
- Yes — it is early access and improving continuously. Coverage widens as more printings get warmed, accuracy tightens as trainer verdicts accumulate, and thin-evidence cards convert from ranges into confident comps as sold data arrives.
- What happens when there is not enough data?
- COMPED says so. You get a range and a sale count, or an explicit 'not comped yet', instead of a fabricated precise figure. Refusing to invent a number is part of being accurate.
- Do you track competitor prices?
- We track the wider market — completed sales, live asks and reference prices — as corroboration, not as a source of truth. If our comp disagrees with a market average, it is because we priced a specific printing in a specific condition in GBP and the average did not.