Methodology

How this site counts

Every number on OP Box Index can be traced to a rule on this page. We publish it because market data is easy to fake and easy to get subtly wrong, and the only defense a reader has is knowing exactly how a figure was made. If you catch a number that breaks one of these rules, that is a bug — tell us and we will fix the pipeline, not just the page.

Auctions are read after they end, never before

Asking prices are wishes. Even a live auction's current bid is not a price: we once tracked a box that sat at $107 twenty-one minutes before closing and finished at $135 — a 26% move in the final stretch, which is normal behavior when bidders wait until the end. So the auction figures on this site come from a settlement pass that re-reads each listing after it closes and records three things: whether it actually sold, what the winning bid was, and how many bidders took part. Unsold listings are kept — they are the denominator of every sell-through rate we show. A listing we fail to re-read within 30 hours of closing is dropped rather than guessed at.

A sample, not a census — and we say so

eBay runs more One Piece card auctions than any collector site can settle one by one, so we work from a fixed daily sample and publish the coverage estimate right next to the data (as of August 2026, roughly two-fifths of the auctions that end each day). Ratios like sell-through survive sampling well; absolute totals do not, which is why we present totals as "what we recorded", never as "the market". Where a rate is shown with a ± figure, that is a Wilson 95% interval — and because our sample is chosen by eBay's search ranking rather than at random, the true uncertainty is somewhat wider than the ± printed. Below a floor of a few dozen settled auctions we print no rate at all: a percentage built on a handful of listings is noise wearing a suit.

Which listings we remove, and what that does to the numbers

Our asking-price figures are filtered. We remove sellers located in China, Hong Kong and Macau, plus accounts we verified as China-based that ship from US warehouses. In July 2026 we found clusters of accounts undercutting the lowest price at identical amounts; we judged the reseal and counterfeit risk too high and have been blocking them one by one since. The list currently holds 30 accounts and 3 countries.

The filter is not small. As of 23 September 2026 it removed 348 of 709 collected box listings (49%). Recomputing the median with those listings included lowers it by 27% on average across 21 Japanese sets — 48% for OP-01. So the “mid ask” on this site is not the market's middle price; it is the middle price among sellers that passed verification. The home page shows the excluded listings' floor price separately, without a link — we do not take affiliate revenue from listings we flagged as risky.

The completed-sale ledger has no such filter. It judges only title and language, never the seller, so sold medians and their trends reflect the whole market including China and Hong Kong sellers. Keep that asymmetry in mind when reading the two side by side: asks are a screened sample, sales are not.

Three rules, applied in order.
① Country — if eBay reports the item location as China, Hong Kong or Macau, it is removed. This is a blanket rule that ignores the individual seller, so honest sellers are caught too. We use it anyway because case-by-case checking could not keep pace with how fast new accounts appeared.
② Wording — removed if the seller name or location text contains china, chinese, hongkong, shenzhen, guangzhou, shanghai, beijing, cn or hk.
③ Manual review — accounts that evade both need a human. We require both conditions: (a) the account is undercutting the floor price, and (b) its eBay About page or feedback profile shows a China or Hong Kong location. One without the other is not enough. Each block records the date and the evidence as a code comment.
The list is public. All 30 accounts and their reasons live in the repository source. If you disagree with a call, you can check it yourself.
There is no appeal process yet. If you are a seller who was blocked in error, contact us — we will verify, remove the entry, and log it in Corrections.

Every number has a rule — metric definitions

Classification rules that keep categories honest

• Lots are not singles. A "Set of 100" bundle is a different product from one card. Listings whose titles read as lots are counted in their own category; mixing them would drag down the single-card sell-through and poison the median price. When a lot's card count is known we divide to a per-card price; when it is not, the listing stays out of price statistics entirely.
• Graded is split from raw. A slabbed PSA 10 and a raw near-mint copy are different markets. Grade is parsed from the listing title, and the source of that parse is stored with the row so it can be re-checked later.
• Japanese and English are never summed. Different print runs, different pull rates, different buyers. Population counts and prices are recorded per printing, and any figure that would require adding them together is a figure we do not publish.
• One listing, one category, one game. Search terms overlap — a Japanese Pokemon listing matches more than one query — so every listing is claimed once, by the narrowest match, and can never be counted twice.

Ledgers are append-only

Completed auctions are written to one file per day, and a day that has passed is never rewritten. Aggregates — daily, weekly, monthly — are rebuilt from those ledgers on every run, so if a rule improves, history is recomputed under the new rule instead of being hand-edited. Days when collection was interrupted are flagged as partial and gaps are listed explicitly, because a missing day silently drawn as zero would invent a crash that never happened. Grading population numbers work the same way: each grader's own public report is appended weekly and never modified after the fact.

The guard system: more than 40 checks between the data and you

Before anything publishes, an automated suite re-derives the site's claims from the raw ledgers and fails the build on any mismatch. Some examples of what it enforces: category totals must sum exactly to the overall total; a sell-through rate may not appear when its sample is under the floor; a median may not appear when price observations are too few; an unsold listing may never carry a winning price; the same auction may not appear in two daily files or two games; and weekly listing counts must be averages of daily snapshots, not sums — adding up snapshots would count the same listing once per day it stayed live. When a check fails, the site does not update until a human looks at why. We add a new check every time we find a new way to be wrong, which is often.

What we refuse to do

• No estimates in place of data. A blank cell means "we could not verify this", and we prefer it to a plausible guess.
• No currency blending. Sold amounts are aggregated in USD only; a sale in another currency is counted as a sale but kept out of dollar totals rather than converted at some arbitrary moment's rate.
• No survivor-only ratios. Every rate we print keeps its failures — unsold auctions, ungraded submissions — in the denominator.
• No repackaging of other sites' analysis. Our inputs are primary: marketplace records read through eBay's official API, grading population (PSA via GemRate's set-level aggregation; CGC and TAG from each grader's own report), Japanese NM retail from Yuyu-tei, and release facts from Bandai's official product pages.

Corrections

When we find an error we fix the pipeline that produced it, rebuild history from the ledgers, and record it in the corrections & change history. The version of every dataset is stamped into the files themselves, so a reader comparing yesterday's download to today's can see exactly what moved. Three of our datasets are also downloadable as CSV under CC BY 4.0 — re-run our numbers yourself; that is the point of publishing them.