Index methodology

Confirmed sold transactions only: witnessed sales in the comp ledger and provider rows the sold-data gate has confirmed (USD, price-confirmed, provider-final basis, not a currency mirror). Asking prices never qualify. A listing that merely disappeared is not a sale.

Each constituent is one exact card identity — year, product, card number, parallel, image variation, print run, signature status and physical form — priced at one grade. A raw card is never priced off a slab and a Topps Now #5 never prices a #23.

7 series are defined in code (engine/benchmark.ts), each a named identity and treatment. A series is ACTIVE only with at least 5 comparable confirmed sales inside 365 days; an inactive series is shown with its reason and is never priced. Every series definition carries a version; changing a definition opens a new series and nothing is ever differenced across the change.

• 2018 Topps Chrome RC #150 (raw) — the canonical US rookie card in its rawest, most-traded form — the reference series everything else is read against
• 2018 Topps Chrome RC #150 (PSA 10) — the same rookie card as a PSA 10 slab — a separate market at a separate price level, never blended with raw
• 2018 Topps flagship RC #700 (raw) — the mass-market flagship rookie — the highest-volume, lowest-friction Ohtani market there is
• Factory-certified autographs (raw base) — the pack-issued signature market, kept apart from every non-auto series and from the aftermarket-signature flood
• Japanese era 2013–2017 (BBM / Calbee, raw) — the pre-MLB Japanese market — different sellers, different region, priced apart from every US series
• Topps Now milestones (raw) — the milestone-event market — print-to-demand cards whose supply is public and whose demand tracks the news cycle
• Dodgers-era flagship 2024– (raw) — the current-team flagship market — what the newest collectors buy first

Membership is documented in code and on the market page, which prints every series with its state.

On a given day a series is an eligible constituent only when it is active, its latest price point is confirmed, that point rests on at least 3 sales, and it was observed within 45 days. Fewer than 3 eligible series and the day is not published — a one-card index is that card's price chart. A stale constituent drops out rather than carrying yesterday's number as today's.

The index is equal-weighted on returns, not on dollars: each constituent contributes the ratio of its price to its own price on the previous published day, and the index moves by the mean of those ratios. A dollar-weighted mean is exactly how one expensive autograph would drown a rookie card, and the brief names that failure.

The first published day is the base, 100. When the basket's membership changes, the level is chain-linked: the day's return is measured over the constituents present on both days, and a day with no overlap is not published. Constituents and weights behind every published value are stored with it (market_index_points.constituents), so any level can be audited.

Methodology version 1 is part of the stored key (shohei-card-market). A change to any rule above bumps it; the previous series stops and a new one begins. Historical values are never rewritten.

The per-series index is refreshed by the daily scan pass; the composite is recomputed after it and again in the evening (.github/workflows/market-journal.yml). Both are idempotent: the whole history is recomputed from stored points on every run, so a retry cannot inflate anything.

“Shohei vs markets” rebases both series to 100 on the first day inside the chosen window (1M, 3M, 6M, YTD, 1Y, 3Y, 5Y). A side needs at least 2 observations inside the window to be drawn, and the comparison is drawn only when both sides have them; otherwise each side states what it lacks and no line is drawn. The S&P 500 series comes from a provider adapter that ships off by default — displaying index data on a public site is a licensing decision, recorded in docs/MARKET-JOURNAL.md.

A mover is one exact identity whose median delivered confirmed price over the last 30 days differs from its median over the 90 days before by at least 5%. Both windows need at least 3 confirmed sales; medians, never means, so one odd sale cannot carry a headline. A card that traded on a smaller sample is counted and named as insufficient rather than silently omitted.

• Unvalidated: no history, no backtest. Every constant above is an unreviewed default (2026-09-01) and is marked so in code.
• Small basket: 7 defined series, most of them expected to be inactive until real sold data flows. The index will be sparse before it is anything else.
• Survivorship by construction: a series that stops trading leaves the basket. The chained level is continuous, but the basket the level describes is not the same basket over time; the stored constituents say which.
• An index level is not a price and not a forecast. Confirmed sales are witnessed prices, not a valuation of any card you hold.
• Evidence classes on every figure: confirmed, modeled, insufficient evidence, unavailable. A blank with a reason is always preferred to a number the evidence cannot support.