How Many Sales Does a Pokémon Card Price Need?

Every graded price moonstone publishes carries a sample size beside it, and the sample size is the more informative half of the pair. We measured what it is worth. Across the 29,676 card-and-grade rows our daily graded tape carried on September 22, 2026, the median sold median stands on eight verified sales, and 56.3% of them stand on fewer than ten. Then we watched 25,022 of those series every day for eight days. When a median built on three or four sales moved at all, it moved 19.5% in a single day. When a median built on 100 or more moved, it moved 5.4%. That ratio is the whole argument for printing an n.

Start with the shape of the thing. Our graded tape publishes a sold median for a card and a grade only when at least three verified sales support it, and the floor is doing heavy lifting: on September 22, 2026 the tape carried no row at all below three comps, 27.9% of rows at exactly three or four, and another 28.4% between five and nine. The mean is 27.7 sales and the median is eight, which is the usual signature of a distribution with a long right tail and a crowded floor. Read the mean and you will imagine a market. Read the median and you will imagine a rumour. Both are in the same table.

Price does almost nothing to fix this. Rows under $25 sit on a median of six sales and rows at $2,000 and up sit on a median of ten, and the share standing on fewer than ten sales only falls from 64.1% to 49.8% across that entire range. A four-figure slab price is very slightly better evidenced than a twenty-dollar one, and not by enough to change how you should read it. The grade matters more than the price: PSA 10 rows carry a median of eleven sales and CGC 9.5 rows carry four.

A census of sample sizes is a description. The useful question is what a thin sample does to the number standing on it, and that needs a panel rather than a snapshot. We took every card-and-grade series present on all eight days from September 15 to September 22, 2026, which is 25,022 series across 10,333 cards, and looked at all 175,154 consecutive-day pairs inside it. Most pairs are identical, so the interesting statistic is conditional: given that the median moved, how far did it move?

Two columns move in opposite directions and both of them matter. A deep series reprices constantly and gently: a third of its days carry a change, and the typical change is 5.4%. A thin series reprices rarely and violently: 1.3% of its days carry a change, and when one arrives it is 19.5%, with almost one move in five exceeding 50%. The same pattern shows up over the whole window rather than day to day. Among series that moved at least once in the eight days, the high-over-low range was 1.254x at five to nine comps and 1.134x at 100 or more, and 14.7% of the thin movers ranged by a factor of two or more against 2.8% of the deep ones.

The mistake this data exposes most often is reading stillness as agreement. Of the 25,022 series, 15,104 did not change by a single cent across the whole eight days. Their median comp count is six, and 72.5% of them stand on fewer than ten sales. Those are not prices the market has settled on. They are prices nothing has tested since the last time somebody bought one. The quiet is the symptom, and the gap will close as a jump rather than a drift.

One card carries the whole lesson. The PSA 10 median for Charizard from Dragon (the 2003 EX set) read $5,087.50 on 59 verified sales on September 7, 2026. Seven days later it read $23,786.00 on 30. It has read $23,786.00 on 30 on every one of the eight recorded days since, through September 22. A price that has not moved in eight days looks like a settled number, and this one is a number whose pool was cut in half on September 14 and has not taken a new sale since.

Under ten comps, treat the number as an order of magnitude. One move in five on that tier is a 50% move, so quoting two decimal places is false precision and a percentage change over a week is mostly arrival timing. Between ten and fifty, expect the number itself to step 13% to 17% whenever it steps, which makes the second decimal place noise. Above 100, treat it as a market price: it reprices on a third of all days and the typical step is 5.4%, which is what price discovery looks like when it is actually happening. These thresholds sit close to the publication floors our earlier census of how many sales is a price described from 94 series, and this study is the whole-population version of that piece.

The rule generalises past our own data. Any site that prints a graded price without a sample size is asking you to assume the sample is adequate, and on this tape the median sample is eight. The comp count is not a footnote to the price. It is the error bar.

One observation kills the finding: a longer panel in which the typical one-day move is flat across comp depth, or in which thin series move no further than deep ones once you condition on moving. If a ninety-day panel shows the 3-to-4 band moving 6% and the 100-plus band moving 5%, the relationship here was a short-window artifact and the thresholds above are worthless. A second and weaker check: if the share of thin-pool moves exceeding 50% falls below the deep-pool share in any future window, the mechanism we are claiming is backwards.

It cannot tell you the true price of anything. It measures how far our published estimate travels, not how far the underlying market moved, and a thin pool that jumps 40% may be correcting toward the truth rather than away from it. It cannot speak to any window longer than eight days, because that is how much dense daily graded tape exists. It says nothing about raw prices, which are asks and are measured in our companion studies on the ask against the sold price and what stands behind a quote at all. And the panel is the set of cards we track, which leans toward cards people grade; the untracked tail of the catalog is thinner still, not deeper.

Source: the `graded_snapshots` table, one row per card key, grade and UTC day, written by our daily graded cron. A series is a (card key, grade) pair. The census is every row with a median above zero on day 20718 (September 22, 2026 UTC), 29,676 rows across 12,449 card keys and seven grade labels (PSA 9, PSA 10, CGC 9, CGC 9.5, CGC 10, BGS 9, BGS 9.5). The panel is every series present on all eight days 20711 to 20718 (September 15 to September 22, 2026), 25,022 series across 10,333 cards, which is a balanced panel by construction: series appearing on some days and not others are excluded, and that exclusion removes cards the cron reached intermittently rather than a price stratum. Day 20719 exists in the table and is excluded because it was still being written. Medians and percentiles use the upper-middle nearest-rank convention: the value at index floor(p x n) of the ascending sample, which puts the median at the upper middle on even samples. COMP-DEPTH BANDS ARE CUT ON THE SERIES' SAMPLE SIZE ON DAY 20718, the last day of the window, so every day-pair and every eight-day range belonging to one series falls in one band and the bands partition all 175,154 pairs and all 25,022 series exactly once. Cutting instead on the earlier day of each pair moves the 3-to-4 row to 45,877 pairs, a 2.3% change rate and a 24.3% typical move, and widens rather than narrows the gap to the 100-plus row, so the direction of this result does not depend on the choice. A move is a change of any size in the median between two consecutive recorded days; the conditional statistics are computed only over pairs where the median changed. `sample_size` is the provider's count of kept verified sold comps for that card and grade; it is not capped. All prices here are SOLD medians from verified graded sales, never asks. A row absent from the table is not recorded, not zero. The 221,817-pair monotonicity check runs over the full recorded history of the table, day 20655 to 20719. Nothing here is investment advice.