How We Value Cards

Valuation Standard v1.0 · every number on this page is read live from our database, not typed in by hand.

Most price guides show you one number and never tell you how much to trust it. We show you two: a value, and a range around it. The range is the honest part — it is wide when we are guessing and narrow when we are not.

The short version: we price a card off its own recent sales when it has them. Only 8% of cards do. For everything else we work outward — the same card in a different grade, then cards like it in the same set — and the range widens at every step to tell you we did.

1. Where the number comes from

Accountants have a standard for exactly this problem — valuing something when there isn't an active market for it. It's called ASC 820, and it sorts evidence into three levels by how directly observable it is. We use the same three levels, honestly, and publish which one every card falls into.

Level 1 — this exact card actually sold, recently

4,709 cards

Real sales of this exact card, in this exact grade, within the last 180 days. No adjustment beyond recency weighting. This is the strongest evidence there is, and the range is at its narrowest here.

Level 2 — observable, but it needs adjusting

15,711 cards

Still real sales — just not a clean, current quote for this exact card. Two kinds:

Stale own sales (8,314 cards). This card sold, but not for over 180 days. ASC 820 calls this "a quoted price in a market that is not active" — observable, but you cannot use a two-year-old price as today's price. We time-adjust it with a market index (section 3).
Comparable specs (7,397 cards). This exact card in this grade never sold, but the same card sold in a different grade, or in a different variety. We project across using grade and variety ratios measured from cards that have sold in both.

Level 3 — a model estimate, no direct comp

38,762 cards

Nothing about this specific card has ever sold at any grade. We fall back to what similar cards in the same set and grade go for, adjusted for population scarcity and how that character generally performs. This is a genuine estimate. It is the majority of our catalogue, and its range is the widest we publish — deliberately.

Tier	Cards	Share	Typical range width
L1	4,709	8.0%	4.23× low to high
L2a	8,314	14.0%	6.56× low to high
L2b	7,397	12.5%	8.9× low to high
L3	38,762	65.5%	12.17× low to high

Roughly 22% of priced cards have sales history of their own. The rest are estimates and are labelled as such.

2. Recent sales count more

When a card has several sales we do not simply average them. Each sale is weighted by age on a 30-day half-life: weight = 0.5 ^ (days_old / 30).

In plain terms: a sale from today counts fully, a sale from a month ago counts half as much, two months ago a quarter, three months an eighth. A card that sold for $40 last week and $10 last year reads much closer to $40 than to $25. This is the single biggest reason our numbers move faster than guides that average a whole year flat.

We also trim extremes and require several distinct sale days before we let a sudden move reprice a card, so one odd auction — or a batch of listings that all ended in the same second — cannot yank a value on its own.

3. Old sales get adjusted to today

If a card last sold two years ago, that price is not today's price. Something has to bridge the gap. We use a repeat-sales index — the same technique behind the Case-Shiller house price index.

It works by only ever comparing a card to itself: take every card that sold in two consecutive quarters, measure how much each one moved, and take the median. Because it compares like with like, it cannot be fooled by a quarter that simply happened to contain more expensive cards.

Market segment	Index today (first quarter = 100)
mid (1990–2012)	130
modern (2013+)	215
vintage (pre-1990)	309

We used to do this badly. The old model applied a fixed ladder — roughly "add 5% if the comp is 3-6 months old, 47% if it's over two years" — and only ever adjusted upward. When we tested it properly against sales it had never seen, that ladder was worse than making no adjustment at all:

Time-adjustment method	Median error	Average error
no time adjustment use the stale comp as-is	41.7%	66.3%
flat drift ladder the previous shipped model	42.0%	69.6%
global Landfill index whole-market repeat-sales index	40.2%	56.3%
segment (era) index per-era repeat-sales index	40.0%	59.2%
BLENDED (shipped) ← now live vintage uses its era index; others use the global index	40.0%	56.0%

Tested on 7,646 real sales the model had not seen. The index wins, so the index shipped. The most interesting finding: the typical adjustment it applies is slightly downward, not upward — several segments of the GPK market have cooled from where those old comps were struck, and a ladder that only ever adds was systematically overpricing stale cards.

4. The range, and what it means

The range is not a guess and it is not a fixed percentage. We took 40,028 real sales, asked the model to predict each one using only what was knowable the day before that sale, and recorded how wrong it was. The published low and high are simply the 10th and 90th percentile of that measured error, for cards with that kind of evidence.

So the range is an 80% band: about 8 sales in 10 should land inside it. That is the claim, and section 5 checks whether it holds.

Evidence	Age of newest comp	Low	High	Fitted on
L1	0-30	0.61×	1.58×	9,341 sales
L1	0-30	0.84×	1.14×	87 sales
L1	0-30	0.55×	1.71×	408 sales
L1	0-30	0.53×	1.83×	353 sales
L1	0-30	0.58×	1.66×	954 sales
L1	0-30	0.61×	1.56×	7,539 sales
L1	31-90	0.53×	1.85×	5,484 sales
L1	31-90	0.50×	2.54×	363 sales
L1	31-90	0.43×	2.43×	331 sales
L1	31-90	0.46×	2.08×	809 sales
L1	31-90	0.55×	1.73×	3,945 sales
L1	91-180	0.46×	2.08×	3,245 sales
L1	91-180	0.40×	2.61×	321 sales
L1	91-180	0.50×	2.10×	302 sales
L1	91-180	0.41×	2.21×	779 sales
L1	91-180	0.51×	1.90×	1,826 sales
L2a	181-365	0.48×	2.60×	2,196 sales
L2a	181-365	0.54×	3.01×	356 sales
L2a	181-365	0.44×	3.09×	333 sales
L2a	181-365	0.46×	2.76×	599 sales
L2a	181-365	0.50×	2.18×	902 sales
L2a	366-730	0.53×	2.96×	1,287 sales
L2a	366-730	0.56×	3.82×	327 sales
L2a	366-730	0.53×	2.62×	258 sales
L2a	366-730	0.54×	2.85×	396 sales
L2a	366-730	0.52×	2.22×	299 sales
L2a	730+	0.49×	3.61×	1,480 sales
L2a	730+	0.47×	3.76×	747 sales
L2a	730+	0.47×	3.51×	314 sales
L2a	730+	0.49×	3.44×	278 sales
L2a	730+	0.58×	2.63×	110 sales
L2b	all	0.36×	3.19×	2,193 sales
L2b	all	0.36×	3.20×	2,161 sales
L3	all	0.34×	4.06×	2,194 sales
L3	all	0.34×	4.11×	2,178 sales

Read a row as: a card with this evidence typically sells for between low × value and high × value. Note the ranges are not symmetric — the upside tail is much longer than the downside. A card can sell for three times the guide; it cannot sell for less than nothing. Anyone quoting you a tidy ±20% on a thin market is not measuring.

5. How accurate is any of this?

Measured, not asserted. Holding out real sales and predicting them blind:

Evidence tier	Sales tested	Median error	Inside the range
L1 — recent observable sales of this exact spec	26,125	23.9%	77.2%
L2a — this exact spec, but the comps are stale (>180d) and index-adjusted	7,646	40.0%	79.0%
L2b — same card at another grade or variety, projected	3,106	47.7%	79.7%
L3 — set-peer model estimate — no direct comp for this card	3,151	57.6%	78.4%
All cards	40,028	30.1%	77.8%

"Median error 30.1%" means: for half of all test sales we were within 30.1% of the real price, and for half we were further off. On a Level 1 card that number is 23.9%. On a Level 3 estimate it is 57.6%. That gap is the whole reason we publish tiers at all.

Error also grows with the age of the evidence — from about 18.8% when the card sold within the last month, to 46.8% when the newest comp is over two years old. See the live accuracy ledger →

6. What we do not know

Most of our catalogue has never sold. 66% of priced cards are Level 3 — a model estimate with no direct comp. Treat those as a starting point for a conversation, not a price.
A thin market has no single "true" price. When a card trades twice a year, the two prices are often 50% apart. No method fixes that; it is a property of the market, not of the model. That is what the range is for.
We do not see condition within a grade. A high-end PSA 8 and a low-end PSA 8 are the same card to us. Eye appeal, centering and print defects move real money and we cannot measure them.
We do not model hype. A character going viral, a new set announcement, an influencer video — the index catches these only after they show up in completed sales, never before.
Our sales data is not the whole market. We see public completed sales. Private sales, card shows, Facebook groups and Discord trades are invisible to us.
Level 3 error is high and we are not hiding it. Median 57.6% on estimates means half the time we are off by more than that. We would rather show you the number than a false precision.
Backtests flatter models. Ours is measured on sales the model never saw, which is the honest way to do it, but the live ledger on the accuracy page is the real scoreboard — it is frozen before the sale happens and cannot be adjusted afterwards.

Valuation Standard v1.0. Values are estimates for information only, not an offer to buy or sell, and not an appraisal. Every figure on this page is generated from live data and updates as the data does.

Accuracy ledger → · Coverage board →